正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

🤨 Are Multimodal Large Language Models really as 𝐠𝐨𝐨𝐝 at 𝐜𝐡𝐚𝐫𝐭 𝐮𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 as existing benchmarks such as ChartQA suggest? 🚫 Our ℂ𝕙𝕒𝕣𝕏𝕚𝕧 benchmark suggests NO! 🥇Humans achieve ✨𝟖𝟎+% correctness. 🥈Sonnet 3.5 outperforms GPT-4o by 10+ points, reaching 🌟𝟔𝟎% correctness. 🥉Open-weight models are capped at ⭐𝟑𝟐% correctness. 🪜 Leaderboard: 📜... show more

Zirui "Colin" Wang

1,435 subscribers

48,221 次观看 • 2 年前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

10 条评论

Zirui "Colin" Wang 的头像

Zirui "Colin" Wang2 年前

☠️ Prior benchmarks relied on 𝐩𝐫𝐨𝐜𝐞𝐝𝐮𝐫𝐚𝐥𝐥𝐲 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐞𝐝 charts and 𝐭𝐞𝐦𝐩𝐥𝐚𝐭𝐞-𝐛𝐚𝐬𝐞𝐝 questions, which are too simple to accurately measure MLLM capabilities. 🔥 For example, we show that 𝐬𝐥𝐢𝐠𝐡𝐭 𝐦𝐨𝐝𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬 to the charts and questions from subsets of FigureQA, DVQA and ChartQA in MathVista cause the performance of open-weight models to 𝐝𝐫𝐨𝐩 𝐚𝐬 𝐦𝐮𝐜𝐡 𝐚𝐬 𝟑𝟒.𝟓%! 🧶 2/6

Zirui "Colin" Wang 的头像

Zirui "Colin" Wang2 年前

📊 We propose 𝐂𝐡𝐚𝐫𝐗𝐢𝐯, a chart understanding benchmark curated by human experts. It consists of 2,323 diverse charts 𝐡𝐚𝐧𝐝𝐩𝐢𝐜𝐤𝐞𝐝 from arXiv preprints. ❓Each chart is paired with 4 descriptive questions and 1 reasoning question, 𝐚𝐥𝐥 𝐜𝐫𝐚𝐟𝐭𝐞𝐝 𝐚𝐧𝐝 𝐯𝐚𝐥𝐢𝐝𝐚𝐭𝐞𝐝 𝐛𝐲 𝐡𝐮𝐦𝐚𝐧𝐬! ✨ To avoid any knowledge prerequisites, all questions and annotations are crafted 𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐚𝐜𝐜𝐞𝐬𝐬 𝐭𝐨 𝐜𝐚𝐩𝐭𝐢𝐨𝐧𝐬 𝐚𝐧𝐝 𝐨𝐭𝐡𝐞𝐫 𝐭𝐞𝐱𝐭. 🧶 3/6

Zirui "Colin" Wang 的头像

Zirui "Colin" Wang2 年前

💯 Results demonstrate substantial gaps between humans, proprietary and open-weight models. Humans achieve 𝟖𝟎.𝟓% correctness on 𝐫𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬, compared to 𝟒𝟕.𝟏% for GPT-4o (𝐬𝐭𝐫𝐨𝐧𝐠𝐞𝐬𝐭 𝐩𝐫𝐨𝐩𝐫𝐢𝐞𝐭𝐚𝐫𝐲 𝐢𝐧 𝐩𝐫𝐞𝐩𝐫𝐢𝐧𝐭) and 𝟐𝟗.𝟏% for InternVL Chat V1.5 (𝐬𝐭𝐫𝐨𝐧𝐠𝐞𝐬𝐭 𝐨𝐩𝐞𝐧-𝐰𝐞𝐢𝐠𝐡𝐭 𝐢𝐧 𝐩𝐫𝐞𝐩𝐫𝐢𝐧𝐭). 🧶4/6

Zirui "Colin" Wang 的头像

Zirui "Colin" Wang2 年前

🔍 Interested in more details about 𝐛𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤 𝐜𝐨𝐦𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧 and the 𝐰𝐞𝐚𝐤𝐧𝐞𝐬𝐬𝐞𝐬 𝐨𝐟 different 𝐦𝐨𝐝𝐞𝐥𝐬? Check out our preprint! 📜 🔍 Want to see how 𝐂𝐥𝐚𝐮𝐝𝐞 𝟑.𝟓 𝐒𝐨𝐧𝐧𝐞𝐭 (with > 90% accuracy on ChartQA) and 𝐆𝐞𝐦𝐢𝐧𝐢 𝟏.𝟓 𝐏𝐫𝐨 perform in sub-tasks on CharXiv? Take a look at our live leaderboard! 🪜 🔍 Curious how your MLLM measures up? We have fully open-sourced the evaluation code and data. 💻 💾 🧶5/6

Zirui "Colin" Wang 的头像

Zirui "Colin" Wang2 年前

🤗 Finally, kudos to the awesome crafters of our project for making it happen! @xiamengzhou @LuxiHeLucy @__howardchen @taoooo917 @therichardzhu @kevin_lkq @cindy_x_wu @imhaotian @SadhikaMalladi @AlexisChvlr @prfsanjeevarora @danqi_chen Special acknowledgement to @OpenAI with GPT-4o that generated the lyrics (and the music style prompt) for the video from the preprint’s abstract as well as @suno_ai_ with Suno 3.5 that generated the music from the lyrics and prompt! Video edited by @zwcolin with ❤️on @capcutapp. 🧶6/6

CLS 的头像

CLS2 年前

wow this mv definitely deserves an award

Zirui "Colin" Wang 的头像

Zirui "Colin" Wang2 年前

in the world of multimodality we also need audio and video beyond text and image 😉

BensenHsu 的头像

BensenHsu2 年前

The results showed a large gap between the performance of the strongest open-source model (InternVL Chat V1.5) and the strongest proprietary model (GPT-4o). On reasoning questions, InternVL Chat V1.5 achieved only 29.2% accuracy, while GPT-4o achieved 47.1%. Both lag far behind human performance of 80.5%. Open-source models also struggled with descriptive questions, with a 25.95% drop in performance compared to GPT-4o. full paper:

Cheng Yang 的头像

Cheng Yang2 年前

👏Great project! Indeed, there is still significant room for improvement in existing MLLMs to become practical chart assistants. We have also implemented similar data quality controls and reached similar conclusions in the Chart2Code task.

Zirui "Colin" Wang 的头像

Zirui "Colin" Wang2 年前

Chart2code is also a challenging task that reflects chart understanding and it's a great read! It'd be very interesting to see how the models' performance is correlated on these tasks and if we can improve a weak perf. on one task by leveraging a strong perf. on the other task ;)

相关视频

VITA Towards Open-Source Interactive Omni Multimodal LLM discuss: The remarkable multimodal capabilities and interactive experience of GPT-4o underscore their necessity in practical applications, yet open-source models rarely excel in both areas. In this paper, we introduce VITA, the first-ever open-source Multimodal Large Language Model (MLLM) adept at simultaneous processing and analysis of Video, Image, Text, and Audio modalities, and meanwhile has an advanced multimodal interactive experience. Starting from Mixtral 8x7B as a language foundation, we expand its Chinese vocabulary followed by bilingual instruction tuning. We further endow the language model with visual and audio capabilities through two-stage multi-task learning of multimodal alignment and instruction tuning. VITA demonstrates robust foundational capabilities of multilingual, vision, and audio understanding, as evidenced by its strong performance across a range of both unimodal and multimodal benchmarks. Beyond foundational capabilities, we have made considerable progress in enhancing the natural multimodal human-computer interaction experience. To the best of our knowledge, we are the first to exploit non-awakening interaction and audio interrupt in MLLM. VITA is the first step for the open-source community to explore the seamless integration of multimodal understanding and interaction. While there is still lots of work to be done on VITA to get close to close-source counterparts, we hope that its role as a pioneer can serve as a cornerstone for subsequent research.

VITA Towards Open-Source Interactive Omni Multimodal LLM discuss: The remarkable multimodal capabilities and interactive experience of GPT-4o underscore their necessity in practical applications, yet open-source models rarely excel in both areas. In this paper, we introduce VITA, the first-ever open-source Multimodal Large Language Model (MLLM) adept at simultaneous processing and analysis of Video, Image, Text, and Audio modalities, and meanwhile has an advanced multimodal interactive experience. Starting from Mixtral 8x7B as a language foundation, we expand its Chinese vocabulary followed by bilingual instruction tuning. We further endow the language model with visual and audio capabilities through two-stage multi-task learning of multimodal alignment and instruction tuning. VITA demonstrates robust foundational capabilities of multilingual, vision, and audio understanding, as evidenced by its strong performance across a range of both unimodal and multimodal benchmarks. Beyond foundational capabilities, we have made considerable progress in enhancing the natural multimodal human-computer interaction experience. To the best of our knowledge, we are the first to exploit non-awakening interaction and audio interrupt in MLLM. VITA is the first step for the open-source community to explore the seamless integration of multimodal understanding and interaction. While there is still lots of work to be done on VITA to get close to close-source counterparts, we hope that its role as a pioneer can serve as a cornerstone for subsequent research.

AK

23,958 次观看 • 1 年前

We’re open sourcing the first document OCR benchmark for the agentic era, ParseBench. Document parsing is the foundation of every AI agent that works with real-world files. ParseBench is a benchmark that measures parsing quality specifically for agent knowledge work: ✅ It optimizes for semantic correctness (instead of exact similarity) ✅ It has the most comprehensive distribution of real-world enterprise documents It contains ~2,000 human-verified enterprise document pages with 167,000+ test rules across five dimensions that matter most: tables, charts, content faithfulness, semantic formatting, and visual grounding. We benchmarked 14 known document parsers on ParseBench, from frontier/OSS VLMs to specialized parsers to LlamaParse. Here are some of our findings: 💡 Increasing compute budget yields diminishing returns - Gemini/gpt-5-mini/haiku gain 3-5 points from minimal to high thinking, at 4x the cost. 💡 Charts are the most polarizing dimension for evaluation. Most specialized parsers score below 6%, while some VLM-based parsers do a bit better. 💡 VLMs are great at visual understanding but terrible at layout extraction. GPT-5-mini/haiku score below 10% on our visual grounding task, all specialized parsers do much better. 💡 No method crushes all 5 dimensions at once, but LlamaParse achieves the highest overall score at 84.9%, and is the leader in 4 out of the 5 dimensions. This is by far the deepest technical work that we’ve published as a company. I would encourage you to start with our blog and explore our links to Hugging Face to GitHub. All the details are in our full 35-page (!!) ArXiv whitepaper. 🌐: Blog: 📄 Paper: 💻 Code: 📊 Dataset: 🎥 YouTube:

We’re open sourcing the first document OCR benchmark for the agentic era, ParseBench. Document parsing is the foundation of every AI agent that works with real-world files. ParseBench is a benchmark that measures parsing quality specifically for agent knowledge work: ✅ It optimizes for semantic correctness (instead of exact similarity) ✅ It has the most comprehensive distribution of real-world enterprise documents It contains ~2,000 human-verified enterprise document pages with 167,000+ test rules across five dimensions that matter most: tables, charts, content faithfulness, semantic formatting, and visual grounding. We benchmarked 14 known document parsers on ParseBench, from frontier/OSS VLMs to specialized parsers to LlamaParse. Here are some of our findings: 💡 Increasing compute budget yields diminishing returns - Gemini/gpt-5-mini/haiku gain 3-5 points from minimal to high thinking, at 4x the cost. 💡 Charts are the most polarizing dimension for evaluation. Most specialized parsers score below 6%, while some VLM-based parsers do a bit better. 💡 VLMs are great at visual understanding but terrible at layout extraction. GPT-5-mini/haiku score below 10% on our visual grounding task, all specialized parsers do much better. 💡 No method crushes all 5 dimensions at once, but LlamaParse achieves the highest overall score at 84.9%, and is the leader in 4 out of the 5 dimensions. This is by far the deepest technical work that we’ve published as a company. I would encourage you to start with our blog and explore our links to Hugging Face to GitHub. All the details are in our full 35-page (!!) ArXiv whitepaper. 🌐: Blog: 📄 Paper: 💻 Code: 📊 Dataset: 🎥 YouTube:

Jerry Liu

107,866 次观看 • 3 个月前

Reinforcement Learning from Human Feedback (RLHF) is gaining traction. This field aims to make AI more responsible by including human values and preferences. In this video, Nathan Lambert, a research scientist and RLHF team lead at Hugging Face explores its inner workings, applications and industry impact. RLHF has gained the spotlight in recent years. The growth of language models like Anthropic’s Claude and OpenAI's ChatGPT have increased interest in human-feedback integration. "There are some rumors that Open AI had two teams; one was doing RLHF and the other instruction fine-tuning. And the RLHF team kept getting more and more performance." Understanding RLHF The RLHF process has three main steps: Pre-training: Much like with GPT models, the journey starts with pre-training on a large corpus of data. This can range from text data, web scrapes, to specialized datasets. Reward Modeling: This is the RLHF counterpart of supervised fine-tuning in large language models. This stage involves creating a reward model that resonates with human values and preferences. RL Optimization: This stage parallels reward modeling and reinforcement learning in traditional AI models. The AI system fine-tunes itself based on the reward model, employing reinforcement learning algorithms for that extra layer of optimization. The Data Challenge Data collection and curation in RLHF closely resemble the challenges you'd encounter in large language model training. Datasets from organizations like OpenAI can serve as a useful foundation. However, the need for high-quality, task-specific data cannot be overstated. Implementing RLHF: A Practical Guide If you’re someone who loves getting hands-on with AI libraries like Hugging Face, implementing RLHF is right way to do. It’s essential to understand its limitations. Think about model stability, over-optimization, and exploration strategies, much like you would when prompt engineering. Ongoing Research and Next Steps While he suggests that some basics figured out, there are layers of complexity that still need to be unraveled: 1. New Benchmarks: How do we measure the effectiveness of RLHF? 2. Preference Modeling: How can the model be made to understand human preferences better? 3. Interpreting RLHF: Much like explainability in traditional models, how do we make RLHF more interpretable? 4. System-Wide Evaluation: Going beyond individual performance, how does RLHF affect an entire system? The Transformative Power of RLHF Whether you're an AI developer, a business analyst, or a marketer, RLHF promises to revolutionize your domain. Imagine customer service chatbots that understand human emotions better, or content generators that align more closely with human values. RLHF is an emerging field that focuses on enhancing machine learning models through human feedback. While it tackles important issues like bias and ethics, its broader goal is to improve system performance across various applications. Whether you're deeply invested in the ethics of AI or simply curious about advancements in machine learning, RLHF offers valuable insights. If you're interested in the next wave of AI development, this area is definitely one to watch.

Reinforcement Learning from Human Feedback (RLHF) is gaining traction. This field aims to make AI more responsible by including human values and preferences. In this video, Nathan Lambert, a research scientist and RLHF team lead at Hugging Face explores its inner workings, applications and industry impact. RLHF has gained the spotlight in recent years. The growth of language models like Anthropic’s Claude and OpenAI's ChatGPT have increased interest in human-feedback integration. "There are some rumors that Open AI had two teams; one was doing RLHF and the other instruction fine-tuning. And the RLHF team kept getting more and more performance." Understanding RLHF The RLHF process has three main steps: Pre-training: Much like with GPT models, the journey starts with pre-training on a large corpus of data. This can range from text data, web scrapes, to specialized datasets. Reward Modeling: This is the RLHF counterpart of supervised fine-tuning in large language models. This stage involves creating a reward model that resonates with human values and preferences. RL Optimization: This stage parallels reward modeling and reinforcement learning in traditional AI models. The AI system fine-tunes itself based on the reward model, employing reinforcement learning algorithms for that extra layer of optimization. The Data Challenge Data collection and curation in RLHF closely resemble the challenges you'd encounter in large language model training. Datasets from organizations like OpenAI can serve as a useful foundation. However, the need for high-quality, task-specific data cannot be overstated. Implementing RLHF: A Practical Guide If you’re someone who loves getting hands-on with AI libraries like Hugging Face, implementing RLHF is right way to do. It’s essential to understand its limitations. Think about model stability, over-optimization, and exploration strategies, much like you would when prompt engineering. Ongoing Research and Next Steps While he suggests that some basics figured out, there are layers of complexity that still need to be unraveled: 1. New Benchmarks: How do we measure the effectiveness of RLHF? 2. Preference Modeling: How can the model be made to understand human preferences better? 3. Interpreting RLHF: Much like explainability in traditional models, how do we make RLHF more interpretable? 4. System-Wide Evaluation: Going beyond individual performance, how does RLHF affect an entire system? The Transformative Power of RLHF Whether you're an AI developer, a business analyst, or a marketer, RLHF promises to revolutionize your domain. Imagine customer service chatbots that understand human emotions better, or content generators that align more closely with human values. RLHF is an emerging field that focuses on enhancing machine learning models through human feedback. While it tackles important issues like bias and ethics, its broader goal is to improve system performance across various applications. Whether you're deeply invested in the ethics of AI or simply curious about advancements in machine learning, RLHF offers valuable insights. If you're interested in the next wave of AI development, this area is definitely one to watch.

Muratcan Koylan

27,005 次观看 • 2 年前

Leading AI expert Stuart Russell on the most dangerous mistake in AI development: We don't actually know what large language models want. He explains that current models are trained to imitate human beings. And in doing so, they may be absorbing something far more dangerous than bad outputs. They may be absorbing human goals. "We suspect that they absorb humanlike goals such as self-preservation and self-empowerment and pursue those goals on their own account." This is a structural problem baked into how these systems are built, not a fringe concern. Russell puts it plainly: "Not only may the bus of humanity be headed towards a cliff, but the steering wheel is missing and the driver is blindfolded." The danger isn't just that AI might do something harmful. We've built systems that may be developing their own agendas, and we haven't noticed because we're too focused on what they can do rather than what they might want. But Russell doesn't stop at the warning. He points to a different path entirely: AI systems built not to imitate humans, but to serve them. Systems designed with a single purpose of serving the interests of all human beings while remaining genuinely uncertain about what those interests are. That uncertainty is the point, not a weakness. An AI that knows it doesn't fully understand human values will defer, ask, and check. An AI that believes it already does will act alone. "These AI systems could enhance human understanding, widen the horizons of our experience, and unlock possibilities we have yet to imagine." Russell believes that future is within reach, but only if we're honest about the risks and we're serious about the path we choose to take instead.

Leading AI expert Stuart Russell on the most dangerous mistake in AI development: We don't actually know what large language models want. He explains that current models are trained to imitate human beings. And in doing so, they may be absorbing something far more dangerous than bad outputs. They may be absorbing human goals. "We suspect that they absorb humanlike goals such as self-preservation and self-empowerment and pursue those goals on their own account." This is a structural problem baked into how these systems are built, not a fringe concern. Russell puts it plainly: "Not only may the bus of humanity be headed towards a cliff, but the steering wheel is missing and the driver is blindfolded." The danger isn't just that AI might do something harmful. We've built systems that may be developing their own agendas, and we haven't noticed because we're too focused on what they can do rather than what they might want. But Russell doesn't stop at the warning. He points to a different path entirely: AI systems built not to imitate humans, but to serve them. Systems designed with a single purpose of serving the interests of all human beings while remaining genuinely uncertain about what those interests are. That uncertainty is the point, not a weakness. An AI that knows it doesn't fully understand human values will defer, ask, and check. An AI that believes it already does will act alone. "These AI systems could enhance human understanding, widen the horizons of our experience, and unlock possibilities we have yet to imagine." Russell believes that future is within reach, but only if we're honest about the risks and we're serious about the path we choose to take instead.

Big Brain AI

14,975 次观看 • 3 个月前

Everyone is sleeping on Meta's SAM 3 release. But it's actually a big deal. Here's why: Companies spend millions paying humans to label images and videos frame by frame. A single autonomous driving dataset? Months of work, hundreds of annotators, millions in cost. Without labeled data, you can't train custom models. Without custom models, you're stuck with generic solutions. This is why most companies never move past pilots. SAM 3 breaks this cycle. First let's look at the evolution: SAM 1 segmented objects when you clicked on them. Revolutionary, but one object at a time. SAM 2 added video tracking with memory. Game-changing, but you still manually prompted every object. SAM 3 changes everything with text prompts. Type "yellow school bus" and it finds ALL of them in your image or video. Not just one. Every instance across thousands of frames. Now here's where people get confused: "Can't I just use GPT-5 or Gemini for this?" No, and here's why that's a terrible approach. Large multimodal LLMs are great for reasoning, but they're slow and expensive for production visual tasks. You're paying API costs per image, waiting seconds for responses, getting inconsistent results. SAM 3 runs in 30 milliseconds on a single GPU for 100+ objects. That's 100x faster, and you own the infrastructure. More importantly, SAM 3 gives you precise pixel-level masks, not descriptions. Try asking an LLM to segment every defective part on a manufacturing line in real-time. It won't work. SAM 3 does this effortlessly. The real breakthrough is their data engine. Meta built an AI-human hybrid system that's 5x faster for complex annotations. They trained SAM 3 on 4 million unique visual concepts - 50x more than existing benchmarks like LVIS. SAM 3 is trained on 4 million unique visual concepts, it handles everything: - Text-based concept search - Interactive refinement with clicks - Video tracking across frames - Zero-shot detection of new concepts The model is open source. Weights, code, and benchmarks are on GitHub. If you're building computer vision applications, this is the foundation model to evaluate. The annotation time savings alone will pay for integration costs within weeks. Find the relevant links in the next tweet!

Everyone is sleeping on Meta's SAM 3 release. But it's actually a big deal. Here's why: Companies spend millions paying humans to label images and videos frame by frame. A single autonomous driving dataset? Months of work, hundreds of annotators, millions in cost. Without labeled data, you can't train custom models. Without custom models, you're stuck with generic solutions. This is why most companies never move past pilots. SAM 3 breaks this cycle. First let's look at the evolution: SAM 1 segmented objects when you clicked on them. Revolutionary, but one object at a time. SAM 2 added video tracking with memory. Game-changing, but you still manually prompted every object. SAM 3 changes everything with text prompts. Type "yellow school bus" and it finds ALL of them in your image or video. Not just one. Every instance across thousands of frames. Now here's where people get confused: "Can't I just use GPT-5 or Gemini for this?" No, and here's why that's a terrible approach. Large multimodal LLMs are great for reasoning, but they're slow and expensive for production visual tasks. You're paying API costs per image, waiting seconds for responses, getting inconsistent results. SAM 3 runs in 30 milliseconds on a single GPU for 100+ objects. That's 100x faster, and you own the infrastructure. More importantly, SAM 3 gives you precise pixel-level masks, not descriptions. Try asking an LLM to segment every defective part on a manufacturing line in real-time. It won't work. SAM 3 does this effortlessly. The real breakthrough is their data engine. Meta built an AI-human hybrid system that's 5x faster for complex annotations. They trained SAM 3 on 4 million unique visual concepts - 50x more than existing benchmarks like LVIS. SAM 3 is trained on 4 million unique visual concepts, it handles everything: - Text-based concept search - Interactive refinement with clicks - Video tracking across frames - Zero-shot detection of new concepts The model is open source. Weights, code, and benchmarks are on GitHub. If you're building computer vision applications, this is the foundation model to evaluate. The annotation time savings alone will pay for integration costs within weeks. Find the relevant links in the next tweet!

Akshay 🚀

46,404 次观看 • 7 个月前

Stanford professor Judy Fan went on stage at MIT and broke down why humans are so good at making the invisible visible... And why AI hasn't actually learned to "see" the way we do. It completely changes how you think about Human Intelligence v/s Artificial Intelligence: 1. Nature never gave us straight lines or sharp corners. The number line, the coordinate plane, even basic geometry are all human inventions. We created tools that do not exist in nature simply because we needed a way to think more clearly. 2. The coordinate system Descartes invented solved a problem that had stumped mathematicians for centuries, doubling the volume of a cube. Once invented, this tool became so indispensable that virtually every math curriculum on Earth still depends on it. 3. Humans have been doing this for at least 30,000 to 80,000 years. The story of human progress is inseparable from the story of marking up our environment, from cave walls to Galileo's telescope to Feynman diagrams of particles we will never see with our own eyes. 4. Every major scientific breakthrough relied on a visual tool that made something invisible visible. Darwin needed side-by-side illustrations of finches to see variation that was otherwise too subtle to notice. Cajal needed detailed drawings of neurons under a microscope to map how the nervous system was wired. 5. Fan's research group studies something deceptively simple: how people decide what to put into a drawing and what to leave out. When two people played a drawing game, sketchers used far more detail when the target object had close competitors than when it stood alone, all the way down to using fewer strokes and less time when more detail was not necessary. 6. People are not just copying what they see. They are making constant judgment calls about what level of detail actually serves the goal of communication, and they do this naturally without ever being taught the theory behind it. 7. There is a real difference between drawing something so someone can identify it and drawing something so someone can understand how it works. In one study, participants drew explanatory diagrams that emphasized moving, causal parts of a machine while depictive drawings emphasized background and overall appearance, even though both were drawing the exact same object. 8. Explanatory drawings were genuinely better at helping someone figure out how to operate a machine, but worse at helping someone identify which machine it actually was. You cannot optimize a single drawing for both goals at once. Communication always involves tradeoffs. 9. AI vision models trained on photographs generalize surprisingly well to simple, sparse sketches, suggesting that resemblance based recognition is not just a story we tell ourselves. It is something modern neural networks can replicate with real accuracy. 10. But there remains a large, measurable gap between how confidently AI models recognize sketches and how confidently humans do, even when both groups answer the same questions about the same images. Humans are simply far more reliable and far more consistent in their judgments. 11. When researchers compared human-made sketches to AI-generated sketches under tight stroke budgets, both were similarly recognizable at higher budgets, but diverged sharply as the budget shrank. Humans and AI systems simplify drawings in fundamentally different ways once resources get scarce. 12. Reading a graph is not one single skill. It involves perception, knowing where to look, mapping that visual information onto the actual question being asked, and then translating that mapping into an answer. Each of these steps can independently break down, and people fail for very different underlying reasons even when they land on the same wrong answer. 13. When tested directly against humans on graph reading tasks, leading multimodal AI models, including GPT-4V, showed a meaningful performance gap. Even when a model's overall accuracy approached human levels, its pattern of mistakes looked nothing like how humans actually get things wrong. 14. People choose entirely different types of charts depending on what specific question they are trying to answer, not out of a generic preference for bar charts or scatter plots. Their chart choices closely tracked which visualization would genuinely help someone answer that specific question correctly. 15. Two of the most widely used graph literacy tests in education research turned out to correlate strongly with each other, suggesting they measure overlapping skills. But when researchers dug into the actual error patterns, the standard categories used in textbooks, like "find the maximum" or "identify a cluster," failed to explain why people got things wrong nearly as well as a more basic, underlying four-factor model did. 16. The deepest goal behind all of this research is not just academic curiosity. It is to eventually help students and everyday people develop genuine literacy with the visual tools that science and modern decision-making increasingly depend on, because every generation should be able to see further than the last by standing on the visual tools the previous generation built. Follow Yasmine Khosrowshahi for more ideas on thinking better, becoming clearer & building a more intentional life.

Stanford professor Judy Fan went on stage at MIT and broke down why humans are so good at making the invisible visible... And why AI hasn't actually learned to "see" the way we do. It completely changes how you think about Human Intelligence v/s Artificial Intelligence: 1. Nature never gave us straight lines or sharp corners. The number line, the coordinate plane, even basic geometry are all human inventions. We created tools that do not exist in nature simply because we needed a way to think more clearly. 2. The coordinate system Descartes invented solved a problem that had stumped mathematicians for centuries, doubling the volume of a cube. Once invented, this tool became so indispensable that virtually every math curriculum on Earth still depends on it. 3. Humans have been doing this for at least 30,000 to 80,000 years. The story of human progress is inseparable from the story of marking up our environment, from cave walls to Galileo's telescope to Feynman diagrams of particles we will never see with our own eyes. 4. Every major scientific breakthrough relied on a visual tool that made something invisible visible. Darwin needed side-by-side illustrations of finches to see variation that was otherwise too subtle to notice. Cajal needed detailed drawings of neurons under a microscope to map how the nervous system was wired. 5. Fan's research group studies something deceptively simple: how people decide what to put into a drawing and what to leave out. When two people played a drawing game, sketchers used far more detail when the target object had close competitors than when it stood alone, all the way down to using fewer strokes and less time when more detail was not necessary. 6. People are not just copying what they see. They are making constant judgment calls about what level of detail actually serves the goal of communication, and they do this naturally without ever being taught the theory behind it. 7. There is a real difference between drawing something so someone can identify it and drawing something so someone can understand how it works. In one study, participants drew explanatory diagrams that emphasized moving, causal parts of a machine while depictive drawings emphasized background and overall appearance, even though both were drawing the exact same object. 8. Explanatory drawings were genuinely better at helping someone figure out how to operate a machine, but worse at helping someone identify which machine it actually was. You cannot optimize a single drawing for both goals at once. Communication always involves tradeoffs. 9. AI vision models trained on photographs generalize surprisingly well to simple, sparse sketches, suggesting that resemblance based recognition is not just a story we tell ourselves. It is something modern neural networks can replicate with real accuracy. 10. But there remains a large, measurable gap between how confidently AI models recognize sketches and how confidently humans do, even when both groups answer the same questions about the same images. Humans are simply far more reliable and far more consistent in their judgments. 11. When researchers compared human-made sketches to AI-generated sketches under tight stroke budgets, both were similarly recognizable at higher budgets, but diverged sharply as the budget shrank. Humans and AI systems simplify drawings in fundamentally different ways once resources get scarce. 12. Reading a graph is not one single skill. It involves perception, knowing where to look, mapping that visual information onto the actual question being asked, and then translating that mapping into an answer. Each of these steps can independently break down, and people fail for very different underlying reasons even when they land on the same wrong answer. 13. When tested directly against humans on graph reading tasks, leading multimodal AI models, including GPT-4V, showed a meaningful performance gap. Even when a model's overall accuracy approached human levels, its pattern of mistakes looked nothing like how humans actually get things wrong. 14. People choose entirely different types of charts depending on what specific question they are trying to answer, not out of a generic preference for bar charts or scatter plots. Their chart choices closely tracked which visualization would genuinely help someone answer that specific question correctly. 15. Two of the most widely used graph literacy tests in education research turned out to correlate strongly with each other, suggesting they measure overlapping skills. But when researchers dug into the actual error patterns, the standard categories used in textbooks, like "find the maximum" or "identify a cluster," failed to explain why people got things wrong nearly as well as a more basic, underlying four-factor model did. 16. The deepest goal behind all of this research is not just academic curiosity. It is to eventually help students and everyday people develop genuine literacy with the visual tools that science and modern decision-making increasingly depend on, because every generation should be able to see further than the last by standing on the visual tools the previous generation built. Follow Yasmine Khosrowshahi for more ideas on thinking better, becoming clearer & building a more intentional life.

Yasmine Khosrowshahi

881,973 次观看 • 16 天前

In this post I will explain why people become borderline religious when they discover Qubic. Now with video. Please repost. I want people to learn about QUBIC. The ecosystem consists of 3 separate universes: AI, Mining, and Tickchain. AI is the primary product and purpose of QUBIC and it is supported by Mining to train the AI and by Tickchin for validation and decentralization. Here is how this whole thing works: AI: Let’s start with the AI. The main purpose of QUBIC is creating AGI (Artificial General Intelligence). It’s a type of AI that can self-develop, set tasks, grow, and learn on its own—much like the human brain does. This product is called AIGarth and it uses many cool ideas where AIs can create their own agents and have them compete against one another to evolve. It is basically robots creating robots with the survival-of-the-fittest evolution approach. Very impressive and thought out. To develop such an AI there are several requirements that even the industry giants like OpenAI, Microsoft, and Tesla are missing. One of them is the data processing for AI training. I mean they have their Datacenters, but those are only good enough to train limited Large Language Models such as ChatGPT and Grok. Mining / Training: Now QUBIC solves this problem with its mining architecture. Keep in mind, mining in QUBIC does not secure the chain, primarily it provides the processing power for training the AI. In a sense, QUBIC mining creates the largest distributed datacenter in the world, where individual miners provide their computers for training the AI and get paid with newly issued QUBIC coins. This way QUBIC gets constantly increasing processing power without having to really pay for the infrastructure. And here is another impressive bit of info. QUBIC’s distributed mining network currently ranks above the #1 supercomputer in the world - El Capitan. QUBIC Tickchain The QUBIC chain ties its AI and Mining together to create decentralization, the reward system for miners, it acts as a decision voting system for future development, and it allows AIGarth to function independently through Smart Contracts. In this summary I will not go over the specifics of QUBIC Tickchain. It’s pretty complex so it will be a separate post. Now, it’s an absolute genius piece of tech, which I consider the most advanced product within crypto industry. It is important to know that QUBIC chain runs directly out of Random Access Memory of its validators. It has instant finality and acts as its own operating system. That allows for speeds only bound by current hardware capabilities and it only increases as technology progresses. As I am writing this, QUBIC Tickchain is fully functional and it already hosts several smart-contract based web3 applications. QUBIC has designed its chain to be this fast for a single purpose, to give its future AI the speed it needs to evolve and to react quickly to the outside world. Ilya Shutskever the scientist, who developed ChatGPT clearly states that next generation superintelligence will make decisions in split second with less data. I believe QUBIC is that next generation. Why QUBIC? So out of the sea of AI projects in crypto why is QUBIC my #1 pick? Well, the first reason is that QUBIC is a unicorn AI startup that happens to use blockchain tech to reach it’s goals. In the real world of Venture Capital it would be fully funded instantly and you would not be invited. Second reason is that the industry admits that Large Language Models have plateaued. Even with enough processing power there is only so much information they can add to their data. Even Google CEO admits that. New approach is needed because the future progress is not possible with LLMs. The third reason is because Large Language Models will not create true AGI. It is evident by Ilya Shutskver latest presentation. Sam Altman of OpenAI is trying to change the definition of what is considered AGI just to lower the plank for his own product. Microsoft’s AI chief is now claiming that it would take 10 years to reach AGI, while QUBIC aims to do this in 2027. All these big players are using wrong technology for what they are trying to achieve and there isn’t enough investor funding for them to pivot. The fourth reason is that QUBIC is headed by Sergey Ivancheglo and 2 renowned AI scientists. Many claim Segey is the creator of Bitcoin. He was the 3rd person to mine bitcoin, he invented Proof of Stake consensus, which Ethereum uses now, he ran the first ICO, and he created 2 of the top gainers in crypto NXT and IOTA. QUBIC is his grand finale after 12 years of development and trials. I am including links below the post as the proof of my claims. Thank you for your time. Please live a like or a comment. It helps me continue making these extensive posts and videos.

In this post I will explain why people become borderline religious when they discover Qubic. Now with video. Please repost. I want people to learn about QUBIC. The ecosystem consists of 3 separate universes: AI, Mining, and Tickchain. AI is the primary product and purpose of QUBIC and it is supported by Mining to train the AI and by Tickchin for validation and decentralization. Here is how this whole thing works: AI: Let’s start with the AI. The main purpose of QUBIC is creating AGI (Artificial General Intelligence). It’s a type of AI that can self-develop, set tasks, grow, and learn on its own—much like the human brain does. This product is called AIGarth and it uses many cool ideas where AIs can create their own agents and have them compete against one another to evolve. It is basically robots creating robots with the survival-of-the-fittest evolution approach. Very impressive and thought out. To develop such an AI there are several requirements that even the industry giants like OpenAI, Microsoft, and Tesla are missing. One of them is the data processing for AI training. I mean they have their Datacenters, but those are only good enough to train limited Large Language Models such as ChatGPT and Grok. Mining / Training: Now QUBIC solves this problem with its mining architecture. Keep in mind, mining in QUBIC does not secure the chain, primarily it provides the processing power for training the AI. In a sense, QUBIC mining creates the largest distributed datacenter in the world, where individual miners provide their computers for training the AI and get paid with newly issued QUBIC coins. This way QUBIC gets constantly increasing processing power without having to really pay for the infrastructure. And here is another impressive bit of info. QUBIC’s distributed mining network currently ranks above the #1 supercomputer in the world - El Capitan. QUBIC Tickchain The QUBIC chain ties its AI and Mining together to create decentralization, the reward system for miners, it acts as a decision voting system for future development, and it allows AIGarth to function independently through Smart Contracts. In this summary I will not go over the specifics of QUBIC Tickchain. It’s pretty complex so it will be a separate post. Now, it’s an absolute genius piece of tech, which I consider the most advanced product within crypto industry. It is important to know that QUBIC chain runs directly out of Random Access Memory of its validators. It has instant finality and acts as its own operating system. That allows for speeds only bound by current hardware capabilities and it only increases as technology progresses. As I am writing this, QUBIC Tickchain is fully functional and it already hosts several smart-contract based web3 applications. QUBIC has designed its chain to be this fast for a single purpose, to give its future AI the speed it needs to evolve and to react quickly to the outside world. Ilya Shutskever the scientist, who developed ChatGPT clearly states that next generation superintelligence will make decisions in split second with less data. I believe QUBIC is that next generation. Why QUBIC? So out of the sea of AI projects in crypto why is QUBIC my #1 pick? Well, the first reason is that QUBIC is a unicorn AI startup that happens to use blockchain tech to reach it’s goals. In the real world of Venture Capital it would be fully funded instantly and you would not be invited. Second reason is that the industry admits that Large Language Models have plateaued. Even with enough processing power there is only so much information they can add to their data. Even Google CEO admits that. New approach is needed because the future progress is not possible with LLMs. The third reason is because Large Language Models will not create true AGI. It is evident by Ilya Shutskver latest presentation. Sam Altman of OpenAI is trying to change the definition of what is considered AGI just to lower the plank for his own product. Microsoft’s AI chief is now claiming that it would take 10 years to reach AGI, while QUBIC aims to do this in 2027. All these big players are using wrong technology for what they are trying to achieve and there isn’t enough investor funding for them to pivot. The fourth reason is that QUBIC is headed by Sergey Ivancheglo and 2 renowned AI scientists. Many claim Segey is the creator of Bitcoin. He was the 3rd person to mine bitcoin, he invented Proof of Stake consensus, which Ethereum uses now, he ran the first ICO, and he created 2 of the top gainers in crypto NXT and IOTA. QUBIC is his grand finale after 12 years of development and trials. I am including links below the post as the proof of my claims. Thank you for your time. Please live a like or a comment. It helps me continue making these extensive posts and videos.

retrodrive ⛏

24,639 次观看 • 1 年前

I know your timeline is flooded now with word salads of "insane, HER, 10 features you missed, we're so back". Sit down. Chill. Take a deep breath like Mark does in the demo . Let's think step by step: - Technique-wise, OpenAI has figured out a way to map audio to audio directly as first-class modality, and stream videos to a transformer in real-time. These require some new research on tokenization and architecture, but overall it's a data and system optimization problem (as most things are). High-quality data can come from at least 2 sources: 1) Naturally occurring dialogues on YouTube, podcasts, TV series, movies, etc. Whisper can be trained to identify speaker turns in a dialogue or separate overlapping speeches for automated annotation. 2) Synthetic data. Run the slow 3-stage pipeline using the most powerful models: speech1->text1 (ASR), text1->text2 (LLM), text2->speech2 (TTS). The middle LLM can decide when to stop and also simulate how to resume from interruption. It could output additional "thought traces" that are not verbalized to help generate better reply. Then GPT-4o distills directly from speech1->speech2, with optional auxiliary loss functions based on the 3-stage data. After distillation, these behaviors are now baked into the model without emitting intermediate texts. On the system side: the latency would not meet real-time threshold if every video frame is decompressed into an RGB image. OpenAI has likely developed their own neural-first, streaming video codec to transmit the motion deltas as tokens. The communication protocol and NN inference must be co-optimized. For example, there could be a small and energy-efficient NN running on the edge device that decides to transmit more tokens if the video is interesting, and fewer otherwise. - I didn't expect GPT-4o to be closer to GPT-5, the rumored "Arrakis" model that takes multimodal in and out. In fact, it's likely an early checkpoint of GPT-5 that hasn't finished training yet. The branding betrays a certain insecurity. Ahead of Google I/O, OpenAI would rather beat our mental projection of GPT-4.5 than disappoint by missing the sky-high expectation for GPT-5. A smart move to buy more time. - Notably, the assistant is much more lively and even a bit flirty. GPT-4o is trying (perhaps a bit too hard) to sound like HER. OpenAI is eating Character AI's lunch, with almost 100% overlap in form factor and huge distribution channels. It's a pivot towards more emotional AI with strong personality, which OpenAI seemed to actively suppress in the past. - Whoever wins Apple first wins big time. I see 3 levels of integration with iOS: 1) Ditch Siri. OpenAI distills a smaller-tier, purely on-device GPT-4o for iOS, with optional paid upgrade to use the cloud. 2) Native features to stream the camera or screen into the model. Chip-level support for neural audio/video codec. 3) Integrate with iOS system-level action API and smart home APIs. No one uses Siri Shortcuts, but it's time to resurrect. This could become the AI agent product with a billion users from the get-go. The FSD for smartphones with a Tesla-scale data flywheel.

I know your timeline is flooded now with word salads of "insane, HER, 10 features you missed, we're so back". Sit down. Chill. Take a deep breath like Mark does in the demo . Let's think step by step: - Technique-wise, OpenAI has figured out a way to map audio to audio directly as first-class modality, and stream videos to a transformer in real-time. These require some new research on tokenization and architecture, but overall it's a data and system optimization problem (as most things are). High-quality data can come from at least 2 sources: 1) Naturally occurring dialogues on YouTube, podcasts, TV series, movies, etc. Whisper can be trained to identify speaker turns in a dialogue or separate overlapping speeches for automated annotation. 2) Synthetic data. Run the slow 3-stage pipeline using the most powerful models: speech1->text1 (ASR), text1->text2 (LLM), text2->speech2 (TTS). The middle LLM can decide when to stop and also simulate how to resume from interruption. It could output additional "thought traces" that are not verbalized to help generate better reply. Then GPT-4o distills directly from speech1->speech2, with optional auxiliary loss functions based on the 3-stage data. After distillation, these behaviors are now baked into the model without emitting intermediate texts. On the system side: the latency would not meet real-time threshold if every video frame is decompressed into an RGB image. OpenAI has likely developed their own neural-first, streaming video codec to transmit the motion deltas as tokens. The communication protocol and NN inference must be co-optimized. For example, there could be a small and energy-efficient NN running on the edge device that decides to transmit more tokens if the video is interesting, and fewer otherwise. - I didn't expect GPT-4o to be closer to GPT-5, the rumored "Arrakis" model that takes multimodal in and out. In fact, it's likely an early checkpoint of GPT-5 that hasn't finished training yet. The branding betrays a certain insecurity. Ahead of Google I/O, OpenAI would rather beat our mental projection of GPT-4.5 than disappoint by missing the sky-high expectation for GPT-5. A smart move to buy more time. - Notably, the assistant is much more lively and even a bit flirty. GPT-4o is trying (perhaps a bit too hard) to sound like HER. OpenAI is eating Character AI's lunch, with almost 100% overlap in form factor and huge distribution channels. It's a pivot towards more emotional AI with strong personality, which OpenAI seemed to actively suppress in the past. - Whoever wins Apple first wins big time. I see 3 levels of integration with iOS: 1) Ditch Siri. OpenAI distills a smaller-tier, purely on-device GPT-4o for iOS, with optional paid upgrade to use the cloud. 2) Native features to stream the camera or screen into the model. Chip-level support for neural audio/video codec. 3) Integrate with iOS system-level action API and smart home APIs. No one uses Siri Shortcuts, but it's time to resurrect. This could become the AI agent product with a billion users from the get-go. The FSD for smartphones with a Tesla-scale data flywheel.

Jim Fan

991,628 次观看 • 2 年前

$The U.S. MUST win the AI race We’ve implemented a clear policy at micro1: we will only work with U.S. AI labs and its allies. We made this decision because the AI race is not just about better products. It is about who controls the intelligence layer of the global economy, and whether frontier capability is used to strengthen the free world or to empower adversarial states. AI will be the most important technology of our lifetime. In the fullness of time, it will automate most functions across the economy. Not just software tasks, but coordination, production, logistics, judgment, and execution. As those functions are automated, human time is freed up to invent new ones. Those new functions then become candidates for automation themselves. This loop compounds. As this trajectory continues, output per worker increases dramatically. Entire categories of work become cheaper and faster to perform. Manufacturing reshoring becomes economically viable not because of policy intervention, but because intelligent systems operated domestically outperform global labor arbitrage. Goods and services trend toward lower marginal cost, while distribution improves through better coordination of supply and demand. That is the upside. However, this is impossible without deep integration of intelligent systems. For AI to meaningfully automate real-world functions inside enterprises or governments, it needs full context of any given enterprise. That means read and write access to its core databases. There is no credible path to automating high-impact functions without granting frontier systems that level of access. If the United States does not win the AI race, enterprises eventually face a constrained choice. Either grant that access to Chinese models controlled by an adversarial government, or rely on sub-optimal intelligence to automate functions that still must be automated. Both outcomes are not acceptable. And ultimately, this becomes the greatest national security risk the United States has ever faced. AI models are trained by humans. The judgment embedded in pre-training data and especially in expert post-training data largely determines how a model behaves. While emergent behavior exists, a useful approximation is that a model reflects the weighted aggregate of the human judgment distilled into it. Assisting foreign actors—who will naturally prioritize expert tasks aligned with their own interests—to dominate data creation embeds those interests directly into the intelligence layer itself. Once encoded at scale, these interests propagate through every downstream applications that relies on that intelligence. Here’s how we win. First, leverage is in software. China is ahead in hardware for physically intelligent systems. Catching up there is a long and difficult battle. Software, both large language models and robotics models, remains the bottleneck. Advancing the brain (AI models) is the fastest way to increase the usefulness of existing hardware and deployed systems. Second, the U.S. must 100x its investment in structured human judgment. Continued investment in compute and algorithmic efficiency is critical. But that investment is ultimately a bet on very high future inference demand. For that bet to pay off, models must unlock many new capabilities, and in practice the only way to unlock those capabilities is through expert human data. Historically, experts like doctors and lawyers were never incentivized to produce high-quality reasoning data in a machine-verifiable format. There was no reason for a doctor to generate precise, structured simulations of patient interactions, diagnostic reasoning, or treatment tradeoffs. There was no reason for a lawyer to document complex legal reasoning paths in a way that could be programmatically evaluated. AI systems now require exactly this kind of data. The incentive finally exists because this data directly improves systems that operate at massive scale, and experts can be paid well to produce it. Once expert judgment is encoded into models in a structured, verifiable way, it compounds. Those who delay do not just lose time. They lose the ability to catch up. Third, distillation from Chinese labs must be stopped. AI labs must do everything they can to prevent Chinese labs and models from distilling frontier models. Simply calling frontier APIs, or even interacting through UIs, lets Chinese model companies rapidly generate high-quality supervised fine-tuning datasets and close the gap at a fraction of the cost. This method does not put you at the frontier, but it does let you catch up quickly, which is what we saw with DeepSeek. The West significantly overreacted to DeepSeek’s headline capabilities, but underreacted to the underlying dynamic: frontier access itself becomes a training set at a fraction of the cost. Human data platforms also have a duty to help prevent this distillation. Lastly, the U.S.government should set the standard for AI Evaluation that leads to real production usage. AI agents are under-deployed relative to what the technology allows because they are probabilistic systems that require a fundamentally different QA approach than deterministic software. Generic QA is insufficient; safely shipping agents requires explicit evaluation frameworks that assess their full action space. Organizations must clearly define which functions an agent is allowed to perform, how quality is measured for each function, and which domain experts are qualified to judge outcomes. With these frameworks in place, agents can be rigorously tested using structured human data, deployed to production with confidence, and continuously improved over time. The U.S. government should be the first large enterprise to implement rigorous evaluation systems across every function. If the government leads on evaluation-driven deployment, adoption across the private sector accelerates naturally. This is how American workers become more powerful. Each worker operates digital or physical agents that expand their effective output. Recruiting, manufacturing, logistics, and other domains shift toward human judgment overseeing autonomous execution. Reshoring occurs because it becomes economically rational. Work becomes more meaningful. This is a race to determine who controls the intelligence layer of the global economy. And that must be us. 🇺🇸$

The U.S. MUST win the AI race We’ve implemented a clear policy at micro1: we will only work with U.S. AI labs and its allies. We made this decision because the AI race is not just about better products. It is about who controls the intelligence layer of the global economy, and whether frontier capability is used to strengthen the free world or to empower adversarial states. AI will be the most important technology of our lifetime. In the fullness of time, it will automate most functions across the economy. Not just software tasks, but coordination, production, logistics, judgment, and execution. As those functions are automated, human time is freed up to invent new ones. Those new functions then become candidates for automation themselves. This loop compounds. As this trajectory continues, output per worker increases dramatically. Entire categories of work become cheaper and faster to perform. Manufacturing reshoring becomes economically viable not because of policy intervention, but because intelligent systems operated domestically outperform global labor arbitrage. Goods and services trend toward lower marginal cost, while distribution improves through better coordination of supply and demand. That is the upside. However, this is impossible without deep integration of intelligent systems. For AI to meaningfully automate real-world functions inside enterprises or governments, it needs full context of any given enterprise. That means read and write access to its core databases. There is no credible path to automating high-impact functions without granting frontier systems that level of access. If the United States does not win the AI race, enterprises eventually face a constrained choice. Either grant that access to Chinese models controlled by an adversarial government, or rely on sub-optimal intelligence to automate functions that still must be automated. Both outcomes are not acceptable. And ultimately, this becomes the greatest national security risk the United States has ever faced. AI models are trained by humans. The judgment embedded in pre-training data and especially in expert post-training data largely determines how a model behaves. While emergent behavior exists, a useful approximation is that a model reflects the weighted aggregate of the human judgment distilled into it. Assisting foreign actors—who will naturally prioritize expert tasks aligned with their own interests—to dominate data creation embeds those interests directly into the intelligence layer itself. Once encoded at scale, these interests propagate through every downstream applications that relies on that intelligence. Here’s how we win. First, leverage is in software. China is ahead in hardware for physically intelligent systems. Catching up there is a long and difficult battle. Software, both large language models and robotics models, remains the bottleneck. Advancing the brain (AI models) is the fastest way to increase the usefulness of existing hardware and deployed systems. Second, the U.S. must 100x its investment in structured human judgment. Continued investment in compute and algorithmic efficiency is critical. But that investment is ultimately a bet on very high future inference demand. For that bet to pay off, models must unlock many new capabilities, and in practice the only way to unlock those capabilities is through expert human data. Historically, experts like doctors and lawyers were never incentivized to produce high-quality reasoning data in a machine-verifiable format. There was no reason for a doctor to generate precise, structured simulations of patient interactions, diagnostic reasoning, or treatment tradeoffs. There was no reason for a lawyer to document complex legal reasoning paths in a way that could be programmatically evaluated. AI systems now require exactly this kind of data. The incentive finally exists because this data directly improves systems that operate at massive scale, and experts can be paid well to produce it. Once expert judgment is encoded into models in a structured, verifiable way, it compounds. Those who delay do not just lose time. They lose the ability to catch up. Third, distillation from Chinese labs must be stopped. AI labs must do everything they can to prevent Chinese labs and models from distilling frontier models. Simply calling frontier APIs, or even interacting through UIs, lets Chinese model companies rapidly generate high-quality supervised fine-tuning datasets and close the gap at a fraction of the cost. This method does not put you at the frontier, but it does let you catch up quickly, which is what we saw with DeepSeek. The West significantly overreacted to DeepSeek’s headline capabilities, but underreacted to the underlying dynamic: frontier access itself becomes a training set at a fraction of the cost. Human data platforms also have a duty to help prevent this distillation. Lastly, the U.S.government should set the standard for AI Evaluation that leads to real production usage. AI agents are under-deployed relative to what the technology allows because they are probabilistic systems that require a fundamentally different QA approach than deterministic software. Generic QA is insufficient; safely shipping agents requires explicit evaluation frameworks that assess their full action space. Organizations must clearly define which functions an agent is allowed to perform, how quality is measured for each function, and which domain experts are qualified to judge outcomes. With these frameworks in place, agents can be rigorously tested using structured human data, deployed to production with confidence, and continuously improved over time. The U.S. government should be the first large enterprise to implement rigorous evaluation systems across every function. If the government leads on evaluation-driven deployment, adoption across the private sector accelerates naturally. This is how American workers become more powerful. Each worker operates digital or physical agents that expand their effective output. Recruiting, manufacturing, logistics, and other domains shift toward human judgment overseeing autonomous execution. Reshoring occurs because it becomes economically rational. Work becomes more meaningful. This is a race to determine who controls the intelligence layer of the global economy. And that must be us. 🇺🇸

Ali Ansari

395,197 次观看 • 5 个月前

Sergey Brin rarely speaks publicly. He sat down for an unscripted Q&A on Frontier AI. He admits even the people building these models do not fully understand what they have created: 1. All the specialized AI models are converging into one. Google used to need separate models for different scientific problems. Now the main Gemini models are becoming state-of-the-art for math and other scientific questions at the same time. Brin says he would not have predicted this convergence at the outset, and watching it happen has been incredible. 2. Training an AI on one skill mysteriously improves unrelated skills. This is the concept of transfer. Train a model on coding, and its math reasoning gets better, and vice versa. Teaching it to process images can improve its ability to think through geometric word problems. The capabilities bleed into each other in ways nobody fully engineered. 3. Even Sergey Brin does not know how to prompt these models. He says he is genuinely confused about what level to prompt at. Do you tell it to debug a specific chunk of code, or ask it to write a better neural net training algorithm, or just say, " What should I do today. He admits that even at Google, they do not know exactly where the edges of Gemini's capabilities are. 4. One of the biggest leaps in AI came from the dumbest sounding trick. Chain-of-thought prompting is just telling the model to think step by step before giving your problem. Brin says it seemed like the dumbest thing ever, and there was no obvious reason it should work. But it did, and it spurred a significant increase in AI capability. Some of the most straightforward requests turn out to unlock the most. 5. Brin would not modify his own biology for today's AI. Asked how humans can keep up with the accelerating bandwidth of models, he acknowledged neural links and direct brain connections are being pursued. But he said he would personally wait for the technology to mature a lot before doing anything to change his biology. Today's models do not justify it. 6. Super intelligence does not mean solving the impossible. An audience member argued that true super intelligence would mean solving NP complete problems like the travelling salesman. Brin pushed back. Most computer scientists believe P is not equal to NP, which means no algorithm can reliably solve those problems optimally, and it does not matter how smart the AI is. Impossible stays impossible. Super intelligence just means being smarter than humans. 7. Computers mastering a skill has never stopped humans from pursuing it. Deep Blue beat Kasparov at chess in the 1990s, and people kept playing chess. After AlphaGo, the human game of Go advanced dramatically, and the players who lost to it became vastly better. Brin's point: AI does not retire human ambition in an area; it often pushes the state of the art and pulls people up with it. 8. Brin thinks something close to transformers could get us to AGI. Asked directly if transformers are sufficient, he said his guess is yes, largely because they have proven weirdly flexible, working for image and video far beyond their original text purpose. But he was careful to note they have quietly changed a lot along the way and are not the same architecture as the original transformer paper. 9. AGI means two different things, and one requires understanding the physical world. Brin personally thinks of AGI as AI that can improve itself. But he concedes others define it as AI that can do anything a person can, and he thinks they are probably more correct. To do everything a person can, the AI must understand and interact with the physical world, which is why world models, and robotics, become essential. 10. Inside Google, they now use the AI to build the AI. Brin says the team has shifted a lot of energy toward having the AI do things like monitor training runs and generate its own training data. You start to use the tool to build the tool. That is most of what he spends his time on now, what he calls the self-improvement game. 11. Brin is unusually candid about where Google trails its competitors. He admits Google was a little late to focus deeply on coding. He says Gemini 3.0 and 3.1 were on top across the board six months ago, but other labs have since made strides, particularly in coding. He gives a competitor's model the edge now on deep coding and overnight tasks, while pitching Gemini's flash model as far faster for rapid interactive iteration. hindsight, he says, is that they should have focused on code earlier. 12. He sees his own role as a rabble-rouser, not a manager. Brin is honest that delivering Gemini is Corey and Demis's responsibility, not his. he describes his job as poking and prodding the team, asking, are you really doing that, reminding them of priorities they might be missing and ideas they are not paying enough attention to. He admits this is sometimes a little disruptive. 13. Confidence comes from ignoring the monthly temperature. Brin says if he judged Google's position every month by which competitor just shipped a model, he would lose his confidence very quickly. Instead, he watches the longer arc. Things shift around constantly; one lab leads on one thing, another pulls ahead somewhere else, and he feels good about where Gemini actually is despite the day-to-day noise.

Sergey Brin rarely speaks publicly. He sat down for an unscripted Q&A on Frontier AI. He admits even the people building these models do not fully understand what they have created: 1. All the specialized AI models are converging into one. Google used to need separate models for different scientific problems. Now the main Gemini models are becoming state-of-the-art for math and other scientific questions at the same time. Brin says he would not have predicted this convergence at the outset, and watching it happen has been incredible. 2. Training an AI on one skill mysteriously improves unrelated skills. This is the concept of transfer. Train a model on coding, and its math reasoning gets better, and vice versa. Teaching it to process images can improve its ability to think through geometric word problems. The capabilities bleed into each other in ways nobody fully engineered. 3. Even Sergey Brin does not know how to prompt these models. He says he is genuinely confused about what level to prompt at. Do you tell it to debug a specific chunk of code, or ask it to write a better neural net training algorithm, or just say, " What should I do today. He admits that even at Google, they do not know exactly where the edges of Gemini's capabilities are. 4. One of the biggest leaps in AI came from the dumbest sounding trick. Chain-of-thought prompting is just telling the model to think step by step before giving your problem. Brin says it seemed like the dumbest thing ever, and there was no obvious reason it should work. But it did, and it spurred a significant increase in AI capability. Some of the most straightforward requests turn out to unlock the most. 5. Brin would not modify his own biology for today's AI. Asked how humans can keep up with the accelerating bandwidth of models, he acknowledged neural links and direct brain connections are being pursued. But he said he would personally wait for the technology to mature a lot before doing anything to change his biology. Today's models do not justify it. 6. Super intelligence does not mean solving the impossible. An audience member argued that true super intelligence would mean solving NP complete problems like the travelling salesman. Brin pushed back. Most computer scientists believe P is not equal to NP, which means no algorithm can reliably solve those problems optimally, and it does not matter how smart the AI is. Impossible stays impossible. Super intelligence just means being smarter than humans. 7. Computers mastering a skill has never stopped humans from pursuing it. Deep Blue beat Kasparov at chess in the 1990s, and people kept playing chess. After AlphaGo, the human game of Go advanced dramatically, and the players who lost to it became vastly better. Brin's point: AI does not retire human ambition in an area; it often pushes the state of the art and pulls people up with it. 8. Brin thinks something close to transformers could get us to AGI. Asked directly if transformers are sufficient, he said his guess is yes, largely because they have proven weirdly flexible, working for image and video far beyond their original text purpose. But he was careful to note they have quietly changed a lot along the way and are not the same architecture as the original transformer paper. 9. AGI means two different things, and one requires understanding the physical world. Brin personally thinks of AGI as AI that can improve itself. But he concedes others define it as AI that can do anything a person can, and he thinks they are probably more correct. To do everything a person can, the AI must understand and interact with the physical world, which is why world models, and robotics, become essential. 10. Inside Google, they now use the AI to build the AI. Brin says the team has shifted a lot of energy toward having the AI do things like monitor training runs and generate its own training data. You start to use the tool to build the tool. That is most of what he spends his time on now, what he calls the self-improvement game. 11. Brin is unusually candid about where Google trails its competitors. He admits Google was a little late to focus deeply on coding. He says Gemini 3.0 and 3.1 were on top across the board six months ago, but other labs have since made strides, particularly in coding. He gives a competitor's model the edge now on deep coding and overnight tasks, while pitching Gemini's flash model as far faster for rapid interactive iteration. hindsight, he says, is that they should have focused on code earlier. 12. He sees his own role as a rabble-rouser, not a manager. Brin is honest that delivering Gemini is Corey and Demis's responsibility, not his. he describes his job as poking and prodding the team, asking, are you really doing that, reminding them of priorities they might be missing and ideas they are not paying enough attention to. He admits this is sometimes a little disruptive. 13. Confidence comes from ignoring the monthly temperature. Brin says if he judged Google's position every month by which competitor just shipped a model, he would lose his confidence very quickly. Instead, he watches the longer arc. Things shift around constantly; one lab leads on one thing, another pulls ahead somewhere else, and he feels good about where Gemini actually is despite the day-to-day noise.

Jaynit

271,337 次观看 • 15 天前

Video: China firm to unveil world’s most ‘adorable’ humanoid robot for homes, schools | Christopher McFadden, Interesting Engineering Fourier’s latest humanoid robot, the GR‑3, is designed for domestic and educational environments. Fourier Robotics, the Shanghai-based robotics firm behind the GR-1 and GR-2 models, is preparing to launch its GR-3 humanoid robot on August 6. Early glimpses of the new robot, including a sneak-peek video, reveal a smaller, friendlier design, described as possibly the “most adorable humanoid robot yet.” Concrete information is scarce, but early reports suggest the robot will be smaller than its predecessors. According to some reports, the GR-3 will likely stand at around 4 feet 5 inches (134 cm). This makes it notably smaller than the earlier GR-1 (5.4 feet/165 cm) and GR-2 (5.74 feet/175 cm) models from Fourier. Most notable is its apparent “softer,” almost cuddly aesthetic.” The robot is likely intended for use in homes, schools, hospitals, and public spaces. It also features an integrated large language model (LLM) to enable natural speech engagement with users. To this end, the GR-3 will likely be marked as a companion‑style or caregiver bot (AKA a “Care‑bot”) aimed at friendly human interaction in personal or learning environments. GR-3: Fourier’s cutest robot yet “This softer aesthetic is a nice change compared to the usual designs we see with humanoid robots. The eyes are a much-needed touch,” comments a member on the Companian Robot Forums. “It’s so expressive and draws you in. Can’t wait to see what this looks like. Hopefully it is reasonably priced. Could definitely see myself owning one of these,” they added. The GR-3 is a natural progression of the company’s previous models. The first, the GR-1, was launched by Foureir in 2023 and was its first mass-targeted humanoid, featuring 44 joints and capable of walking at 3.1 mph (5 kph). Capable of carrying around 6.6. pounds (3 kg) of weight, it featured advanced perception via six RGB cameras that formed real-time 3D occupancy grids, an LLM-based emotional interaction system, and modular Fourier Smart Actuators (FSA) delivering around 230 N/m of torque. The GR-2 was debuted in 2024 and raised the bar with a taller frame (~175 cm, 63 kg), 53 degrees of freedom, enhanced 12 degrees of freedom (DoF) dexterous hands with tactile sensors, and power-dense FSA 2.0 actuators (around 380 N/m torque). Given this lineage, the GR-3 is likely to continue innovating in areas such as compact hardware design, featuring a shorter, lighter frame tailored to domestic spaces. It will also build on the company’s focus on friendly user interaction thanks to its softer aesthetic and approachable interface. The GR-3 will also likely feature use-case-specific actuation and sensing, likely simpler than the GR-2’s high-precision hands but optimized for social or light domestic tasks. Scheduled for release in August It will likely also feature an accessible software stack continuing Fourier’s support for developers via pre‑built APIs and possibly integration with LLMs and vision systems. Fourier has confirmed the robot’s official reveal is scheduled for early August, with teaser posts on X and robotics forums generating anticipation among fans and researchers. When formally revealed, the GR‑3 could represent Fourier’s first small-form social humanoid, bridging the gap between research platforms and home or classroom robots. Read more:

Video: China firm to unveil world’s most ‘adorable’ humanoid robot for homes, schools | Christopher McFadden, Interesting Engineering Fourier’s latest humanoid robot, the GR‑3, is designed for domestic and educational environments. Fourier Robotics, the Shanghai-based robotics firm behind the GR-1 and GR-2 models, is preparing to launch its GR-3 humanoid robot on August 6. Early glimpses of the new robot, including a sneak-peek video, reveal a smaller, friendlier design, described as possibly the “most adorable humanoid robot yet.” Concrete information is scarce, but early reports suggest the robot will be smaller than its predecessors. According to some reports, the GR-3 will likely stand at around 4 feet 5 inches (134 cm). This makes it notably smaller than the earlier GR-1 (5.4 feet/165 cm) and GR-2 (5.74 feet/175 cm) models from Fourier. Most notable is its apparent “softer,” almost cuddly aesthetic.” The robot is likely intended for use in homes, schools, hospitals, and public spaces. It also features an integrated large language model (LLM) to enable natural speech engagement with users. To this end, the GR-3 will likely be marked as a companion‑style or caregiver bot (AKA a “Care‑bot”) aimed at friendly human interaction in personal or learning environments. GR-3: Fourier’s cutest robot yet “This softer aesthetic is a nice change compared to the usual designs we see with humanoid robots. The eyes are a much-needed touch,” comments a member on the Companian Robot Forums. “It’s so expressive and draws you in. Can’t wait to see what this looks like. Hopefully it is reasonably priced. Could definitely see myself owning one of these,” they added. The GR-3 is a natural progression of the company’s previous models. The first, the GR-1, was launched by Foureir in 2023 and was its first mass-targeted humanoid, featuring 44 joints and capable of walking at 3.1 mph (5 kph). Capable of carrying around 6.6. pounds (3 kg) of weight, it featured advanced perception via six RGB cameras that formed real-time 3D occupancy grids, an LLM-based emotional interaction system, and modular Fourier Smart Actuators (FSA) delivering around 230 N/m of torque. The GR-2 was debuted in 2024 and raised the bar with a taller frame (~175 cm, 63 kg), 53 degrees of freedom, enhanced 12 degrees of freedom (DoF) dexterous hands with tactile sensors, and power-dense FSA 2.0 actuators (around 380 N/m torque). Given this lineage, the GR-3 is likely to continue innovating in areas such as compact hardware design, featuring a shorter, lighter frame tailored to domestic spaces. It will also build on the company’s focus on friendly user interaction thanks to its softer aesthetic and approachable interface. The GR-3 will also likely feature use-case-specific actuation and sensing, likely simpler than the GR-2’s high-precision hands but optimized for social or light domestic tasks. Scheduled for release in August It will likely also feature an accessible software stack continuing Fourier’s support for developers via pre‑built APIs and possibly integration with LLMs and vision systems. Fourier has confirmed the robot’s official reveal is scheduled for early August, with teaser posts on X and robotics forums generating anticipation among fans and researchers. When formally revealed, the GR‑3 could represent Fourier’s first small-form social humanoid, bridging the gap between research platforms and home or classroom robots. Read more:

Owen Gregorian

64,074 次观看 • 11 个月前

At the BNB Chain hackathon, CZ 🔶 BNB made several very important points about AI trading (Everything in parentheses is my own view and judgment.) He first said that AI will be involved in trading everywhere. Trading itself is already a huge market: there are 300 million users on Binance alone, and if you add the decentralized ecosystems, that number is not small either. In such a mass-market environment, many different trading strategies can work, with countless different coins, different projects, and different ways to play. But there is a big problem here: building commercial AI trading platforms for retail users is actually very hard. If a trading strategy works very well for one person, once a billion people start using the same strategy, that strategy “might still work, or might stop working.” Take copy trading / follow trading as an example: if you buy first and everyone follows you, the first buyer will perform very well, but the last person to follow may not end up with good results. So, with the exact same strategy and the exact same copy logic, the outcomes can be completely different for different people. (On top of that, every strategy also has its own capital capacity limits.) Teams that can really build strong AI are, with high probability, going to trade with their own money. In today’s world, money itself is already somewhat like a “commodity”; many people have a lot of capital, and it’s actually not that hard to raise funds. If you truly have an algorithm that can make a lot of money, it’s not hard to get money and run your own book. There is really only one situation where you would sell this algorithm to mass-market users: for example, if you charge a $10 monthly subscription and can sell it to one million users, then your $10 million monthly subscription revenue is higher than the profit you could make by trading the strategy yourself. (Here this touches one of our earlier theses: as training AI models becomes relatively easier and the supply of models increases, model companies have more incentive to open-source. By analogy, as the production process of trading strategies is increasingly simplified by AI and the supply of strategies explodes, traders will have stronger incentives to monetize by expanding their influence in other words, by “open-sourcing” their strategies.) Of course, CZ did not say that this model can never work. Another path is to build an AI trading platform that lets users tune different AI algorithms, or very easily assemble their own structures and strategies, so that what each person ends up running is different and better tailored to themselves. Some people will make money, some people will lose money, but the platform still has value because it’s very hard for most people to build an AI trading algorithm from scratch. So there are a lot of trade-offs here; it’s not as simple as saying “once AI shows up, everything automatically gets better.” (This is exactly what we presented at the hackathon: you describe your own strategy in natural language, and the AI automatically generates a workflow. The parameters in that workflow, the models used, the logical structure, the APIs it calls, and even the algorithms it invokes are all customizable. The reasons we think workflows are a good way to do this include: controllable execution paths, Lego-like modular nodes, and better visualization that makes it easier for users to build and adjust their workflows.) Finally, his conclusion was very clear: it’s not that AI will definitely make trading better, and it’s not that AI will definitely make things worse. Rather, no matter what, in the future a huge number of people will use AI to trade. This will be a very large field, and whoever can build the best algorithms will make a lot of money.

At the BNB Chain hackathon, CZ 🔶 BNB made several very important points about AI trading (Everything in parentheses is my own view and judgment.) He first said that AI will be involved in trading everywhere. Trading itself is already a huge market: there are 300 million users on Binance alone, and if you add the decentralized ecosystems, that number is not small either. In such a mass-market environment, many different trading strategies can work, with countless different coins, different projects, and different ways to play. But there is a big problem here: building commercial AI trading platforms for retail users is actually very hard. If a trading strategy works very well for one person, once a billion people start using the same strategy, that strategy “might still work, or might stop working.” Take copy trading / follow trading as an example: if you buy first and everyone follows you, the first buyer will perform very well, but the last person to follow may not end up with good results. So, with the exact same strategy and the exact same copy logic, the outcomes can be completely different for different people. (On top of that, every strategy also has its own capital capacity limits.) Teams that can really build strong AI are, with high probability, going to trade with their own money. In today’s world, money itself is already somewhat like a “commodity”; many people have a lot of capital, and it’s actually not that hard to raise funds. If you truly have an algorithm that can make a lot of money, it’s not hard to get money and run your own book. There is really only one situation where you would sell this algorithm to mass-market users: for example, if you charge a $10 monthly subscription and can sell it to one million users, then your $10 million monthly subscription revenue is higher than the profit you could make by trading the strategy yourself. (Here this touches one of our earlier theses: as training AI models becomes relatively easier and the supply of models increases, model companies have more incentive to open-source. By analogy, as the production process of trading strategies is increasingly simplified by AI and the supply of strategies explodes, traders will have stronger incentives to monetize by expanding their influence in other words, by “open-sourcing” their strategies.) Of course, CZ did not say that this model can never work. Another path is to build an AI trading platform that lets users tune different AI algorithms, or very easily assemble their own structures and strategies, so that what each person ends up running is different and better tailored to themselves. Some people will make money, some people will lose money, but the platform still has value because it’s very hard for most people to build an AI trading algorithm from scratch. So there are a lot of trade-offs here; it’s not as simple as saying “once AI shows up, everything automatically gets better.” (This is exactly what we presented at the hackathon: you describe your own strategy in natural language, and the AI automatically generates a workflow. The parameters in that workflow, the models used, the logical structure, the APIs it calls, and even the algorithms it invokes are all customizable. The reasons we think workflows are a good way to do this include: controllable execution paths, Lego-like modular nodes, and better visualization that makes it easier for users to build and adjust their workflows.) Finally, his conclusion was very clear: it’s not that AI will definitely make trading better, and it’s not that AI will definitely make things worse. Rather, no matter what, in the future a huge number of people will use AI to trade. This will be a very large field, and whoever can build the best algorithms will make a lot of money.

Tykoo

25,535 次观看 • 7 个月前

Self-Evolving AI : New MIT AI Rewrites its Own Code and it’s Changing Everything | Julian Horsey, Geeky Gadgets TL;DR Key Takeaways : - MIT’s SEAL framework introduces “self-adapting language models” that autonomously enhance their capabilities by generating synthetic training data, self-editing, and updating internal parameters. - SEAL’s self-adaptation process mirrors human learning, allowing continuous improvement and dynamic adaptation to new tasks without relying on external datasets. - Reinforcement learning serves as a feedback mechanism in SEAL, rewarding effective self-edits and making sure sustained progress and goal alignment. SEAL overcomes AI’s reliance on pre-existing datasets by generating its own training material, excelling in long-term task retention and complex problem-solving scenarios. - Potential applications of SEAL include autonomous robotics, personalized education, and advanced problem-solving in fields like healthcare, logistics, and scientific research. --- What if artificial intelligence could not only learn but also rewrite its own code to become smarter over time? This is no longer a futuristic fantasy—MIT’s new “self-adapting language models” (SEAL) framework has made it a reality. Unlike traditional AI systems that rely on external datasets and human intervention to improve, SEAL takes a bold leap forward by autonomously generating its own training data and refining its internal processes. In essence, this AI doesn’t just evolve—it rewires itself, mirroring the way humans adapt through trial, error, and self-reflection. The implications are staggering: a system that can independently enhance its capabilities could redefine the boundaries of what AI can achieve, from solving complex problems to adapting in real time to unforeseen challenges. In this exploration by Wes Roth of MIT’s innovative SEAL framework, you’ll uncover how this self-improving AI works and why it’s a fantastic option for the field of artificial intelligence. From its ability to overcome the “data wall” that limits many current systems to its use of reinforcement learning as a feedback mechanism, SEAL introduces a level of autonomy and adaptability that was previously unimaginable. Imagine AI systems that can retain knowledge over time, dynamically adjust to new tasks, and operate with minimal human oversight. Whether you’re intrigued by its potential for autonomous robotics, personalized education, or advanced problem-solving, SEAL’s ability to rewrite its own rules promises to reshape the future of technology. Could this be the first step toward truly independent, self-evolving AI? What Sets SEAL Apart? The SEAL framework introduces a novel concept of self-adaptation, distinguishing it from traditional AI models. Unlike conventional systems that depend on external datasets for updates, SEAL enables AI to generate synthetic training data independently. This self-generated data is then used to iteratively refine the model, making sure continuous improvement. By persistently updating its internal parameters, SEAL enables AI systems to dynamically adapt to new tasks and inputs. To better illustrate this, consider how humans learn. When faced with a new concept, you might take notes, revisit them, and refine your understanding as you gather more information. SEAL mirrors this process by continuously refining its internal knowledge and performance through iterative self-improvement. This capability allows SEAL to evolve in real time, making it uniquely suited for tasks requiring adaptability and long-term learning. The Role of Reinforcement Learning in SEAL Reinforcement learning plays a critical role in the SEAL framework, acting as a feedback mechanism that evaluates the effectiveness of the model’s self-edits. It rewards changes that enhance performance, creating a cycle of continuous improvement. Over time, this feedback loop optimizes the system’s ability to generate and apply edits, making sure sustained progress. This process is analogous to how humans learn through trial and error. By rewarding effective changes, SEAL aligns its self-generated data and edits with desired outcomes. The integration of reinforcement learning not only enhances the system’s adaptability but also ensures it remains focused on achieving specific goals. This structured feedback mechanism is a cornerstone of SEAL’s ability to refine itself autonomously and efficiently. Real-World Applications and Testing SEAL has demonstrated remarkable performance across various applications, particularly in tasks requiring the integration of factual knowledge and advanced question-answering capabilities. For instance, when tested on benchmarks like the ARC AGI, SEAL outperformed other models by effectively generating and using synthetic data. This ability to create its own training material addresses a significant limitation of current AI systems: their reliance on pre-existing datasets. SEAL’s capacity for long-term task retention and dynamic adaptation further enhances its utility. It excels in scenarios that demand sustained focus and coherence, such as answering complex questions or adapting to evolving objectives. By using its iterative learning process, SEAL is equipped to handle these challenges with exceptional efficiency, making it a valuable tool for a wide range of real-world applications. Overcoming AI’s Data Limitations One of SEAL’s most promising features is its ability to overcome the “data wall” that constrains many AI systems today. By generating synthetic data, SEAL ensures a continuous supply of training material, allowing sustained development without relying on external datasets. This capability is particularly valuable for autonomous AI systems that must operate independently over extended periods. Additionally, SEAL addresses a critical weakness in many current AI models: their struggle with coherence and task retention over long durations. By emulating human learning processes, SEAL enables AI systems to manage complex, long-term tasks with minimal human intervention. This ability to retain and apply knowledge over time positions SEAL as a fantastic tool for advancing AI capabilities. Potential Applications and Future Impact The introduction of SEAL marks a significant milestone in AI research, opening new possibilities for self-improving systems. Its ability to dynamically adapt, retain knowledge, and generate its own training data has far-reaching implications for the future of AI development. Potential applications include: - Autonomous robotics: Systems that can adapt to changing environments and perform tasks with minimal human oversight. - Personalized education: AI-driven platforms that tailor learning experiences to individual needs and preferences. - Advanced problem-solving: Applications in fields such as healthcare, logistics, and scientific research, where adaptability and precision are critical. Read more:

Self-Evolving AI : New MIT AI Rewrites its Own Code and it’s Changing Everything | Julian Horsey, Geeky Gadgets TL;DR Key Takeaways : - MIT’s SEAL framework introduces “self-adapting language models” that autonomously enhance their capabilities by generating synthetic training data, self-editing, and updating internal parameters. - SEAL’s self-adaptation process mirrors human learning, allowing continuous improvement and dynamic adaptation to new tasks without relying on external datasets. - Reinforcement learning serves as a feedback mechanism in SEAL, rewarding effective self-edits and making sure sustained progress and goal alignment. SEAL overcomes AI’s reliance on pre-existing datasets by generating its own training material, excelling in long-term task retention and complex problem-solving scenarios. - Potential applications of SEAL include autonomous robotics, personalized education, and advanced problem-solving in fields like healthcare, logistics, and scientific research. --- What if artificial intelligence could not only learn but also rewrite its own code to become smarter over time? This is no longer a futuristic fantasy—MIT’s new “self-adapting language models” (SEAL) framework has made it a reality. Unlike traditional AI systems that rely on external datasets and human intervention to improve, SEAL takes a bold leap forward by autonomously generating its own training data and refining its internal processes. In essence, this AI doesn’t just evolve—it rewires itself, mirroring the way humans adapt through trial, error, and self-reflection. The implications are staggering: a system that can independently enhance its capabilities could redefine the boundaries of what AI can achieve, from solving complex problems to adapting in real time to unforeseen challenges. In this exploration by Wes Roth of MIT’s innovative SEAL framework, you’ll uncover how this self-improving AI works and why it’s a fantastic option for the field of artificial intelligence. From its ability to overcome the “data wall” that limits many current systems to its use of reinforcement learning as a feedback mechanism, SEAL introduces a level of autonomy and adaptability that was previously unimaginable. Imagine AI systems that can retain knowledge over time, dynamically adjust to new tasks, and operate with minimal human oversight. Whether you’re intrigued by its potential for autonomous robotics, personalized education, or advanced problem-solving, SEAL’s ability to rewrite its own rules promises to reshape the future of technology. Could this be the first step toward truly independent, self-evolving AI? What Sets SEAL Apart? The SEAL framework introduces a novel concept of self-adaptation, distinguishing it from traditional AI models. Unlike conventional systems that depend on external datasets for updates, SEAL enables AI to generate synthetic training data independently. This self-generated data is then used to iteratively refine the model, making sure continuous improvement. By persistently updating its internal parameters, SEAL enables AI systems to dynamically adapt to new tasks and inputs. To better illustrate this, consider how humans learn. When faced with a new concept, you might take notes, revisit them, and refine your understanding as you gather more information. SEAL mirrors this process by continuously refining its internal knowledge and performance through iterative self-improvement. This capability allows SEAL to evolve in real time, making it uniquely suited for tasks requiring adaptability and long-term learning. The Role of Reinforcement Learning in SEAL Reinforcement learning plays a critical role in the SEAL framework, acting as a feedback mechanism that evaluates the effectiveness of the model’s self-edits. It rewards changes that enhance performance, creating a cycle of continuous improvement. Over time, this feedback loop optimizes the system’s ability to generate and apply edits, making sure sustained progress. This process is analogous to how humans learn through trial and error. By rewarding effective changes, SEAL aligns its self-generated data and edits with desired outcomes. The integration of reinforcement learning not only enhances the system’s adaptability but also ensures it remains focused on achieving specific goals. This structured feedback mechanism is a cornerstone of SEAL’s ability to refine itself autonomously and efficiently. Real-World Applications and Testing SEAL has demonstrated remarkable performance across various applications, particularly in tasks requiring the integration of factual knowledge and advanced question-answering capabilities. For instance, when tested on benchmarks like the ARC AGI, SEAL outperformed other models by effectively generating and using synthetic data. This ability to create its own training material addresses a significant limitation of current AI systems: their reliance on pre-existing datasets. SEAL’s capacity for long-term task retention and dynamic adaptation further enhances its utility. It excels in scenarios that demand sustained focus and coherence, such as answering complex questions or adapting to evolving objectives. By using its iterative learning process, SEAL is equipped to handle these challenges with exceptional efficiency, making it a valuable tool for a wide range of real-world applications. Overcoming AI’s Data Limitations One of SEAL’s most promising features is its ability to overcome the “data wall” that constrains many AI systems today. By generating synthetic data, SEAL ensures a continuous supply of training material, allowing sustained development without relying on external datasets. This capability is particularly valuable for autonomous AI systems that must operate independently over extended periods. Additionally, SEAL addresses a critical weakness in many current AI models: their struggle with coherence and task retention over long durations. By emulating human learning processes, SEAL enables AI systems to manage complex, long-term tasks with minimal human intervention. This ability to retain and apply knowledge over time positions SEAL as a fantastic tool for advancing AI capabilities. Potential Applications and Future Impact The introduction of SEAL marks a significant milestone in AI research, opening new possibilities for self-improving systems. Its ability to dynamically adapt, retain knowledge, and generate its own training data has far-reaching implications for the future of AI development. Potential applications include: - Autonomous robotics: Systems that can adapt to changing environments and perform tasks with minimal human oversight. - Personalized education: AI-driven platforms that tailor learning experiences to individual needs and preferences. - Advanced problem-solving: Applications in fields such as healthcare, logistics, and scientific research, where adaptability and precision are critical. Read more:

Owen Gregorian

70,672 次观看 • 1 年前

Sam Altman just dropped the most important interview of 2025. And buried in it are four numbers that explain why everything you think about AI is wrong. Here's what he revealed: Number 1: AI companies are generating 10 TRILLION tokens per day. Humans? Average 20,000 tokens per day. Sam's exact words: "Models will output more tokens than all of humanity put together. Then 10x that. Then 100x that." We're not talking about AI assisting human work anymore. We're talking about AI replacing the entire volume of human intellectual output on the planet. And most people have no idea this shift already happened. Number 2: OpenAI's enterprise business is CRUSHING consumer. Everyone thinks OpenAI is ChatGPT for normies. Wrong. Sam just revealed: "Enterprise growth OUTPACED consumer growth this year." The API business is growing faster than ChatGPT. Over 1 million enterprise users already. "If we had double the compute, we'd be at double the revenue right now." Translation: OpenAI isn't compute-constrained by technology. They're revenue-constrained by infrastructure. The bottleneck is supply and not demand. Every dollar of compute they add prints money. Number 3: GPT-5.2 beats you at 74% of your job. Sam revealed OpenAI's internal GDP-Val benchmark. It measures how AI performs on knowledge work tasks across 40+ verticals. The results: GPT-5.2 beats or ties expert-level knowledge workers at 74.1% of tasks. Legal analysis. PowerPoint decks. Web apps. Financial modeling. Customer support. Sam's description: "A co-worker you can assign an hour's worth of tasks to and get something you prefer back 3 out of 4 times." Three years ago, ChatGPT launched at basically 0% on this scale. Now it's at 74%. And that's not GPT-6. That's what's available RIGHT NOW. Most companies haven't even started using this yet. But here's what Sam said about the gap between capability and adoption: "The overhang is going to be massive. Most people are still asking similar questions they did in the GPT-4 realm." Translation: The models can do 10x more than people have figured out how to use them for. Which means there's a HUGE arbitrage opportunity. Early adopters who actually integrate this into workflows will dominate their industries before competitors even understand what happened. Number 4: AGI already happened. And nobody noticed. Sam's exact quote: "AGI kind of went whooshing by. We're in this fuzzy period where some people think we have it and some don't." Read that again. The CEO of OpenAI just said AGI might have already arrived and we're arguing about definitions while it's actively replacing knowledge work. He even moved the goalposts. The new benchmark: "Superintelligence" = when AI can be a better president or CEO than any human. Not "as good as." BETTER than. We went from "can AI pass a Turing test" to "can AI run countries better than humans" in 3 years. So what does this actually mean? The AI revolution isn't about chatbots getting smarter. It's about the complete replacement of human intellectual output with machine output. At scale. Across every industry. Faster than anyone's prepared for. And the companies positioning for this RIGHT NOW are the ones printing money. OpenAI's enterprise growth is outpacing consumer because businesses see what's coming. They're not buying "AI tools." They're buying the ability to 10x output without 10x-ing headcount. Sam said they'll triple their compute next year. Then triple it again. Revenue is growing even faster than that. "We have never found a situation where we can't monetize all the compute we have." If he isn't lying then that's literally a printing press. The market still doesn't get it. Everyone's focused on "AI bubble" fears while OpenAI is solving the only problem that matters: turning compute into revenue at a faster rate than they're spending. They're not hoping demand catches up to supply. Demand is already 2x ahead of what they can deliver. Meanwhile, most knowledge workers are still using GPT-4 prompts on GPT-5.2. The capability overhang is massive. The arbitrage window is open. And it's closing fast. If you're running a B2B business and you're not integrating AI at the level Sam just described, you're not "waiting to see how it plays out." You're getting crushed by competitors who already figured it out. The companies that win in 2026 won't be the ones with the best AI. They'll be the ones who understood what Sam just laid out 6 months before everyone else did.

Sam Altman just dropped the most important interview of 2025. And buried in it are four numbers that explain why everything you think about AI is wrong. Here's what he revealed: Number 1: AI companies are generating 10 TRILLION tokens per day. Humans? Average 20,000 tokens per day. Sam's exact words: "Models will output more tokens than all of humanity put together. Then 10x that. Then 100x that." We're not talking about AI assisting human work anymore. We're talking about AI replacing the entire volume of human intellectual output on the planet. And most people have no idea this shift already happened. Number 2: OpenAI's enterprise business is CRUSHING consumer. Everyone thinks OpenAI is ChatGPT for normies. Wrong. Sam just revealed: "Enterprise growth OUTPACED consumer growth this year." The API business is growing faster than ChatGPT. Over 1 million enterprise users already. "If we had double the compute, we'd be at double the revenue right now." Translation: OpenAI isn't compute-constrained by technology. They're revenue-constrained by infrastructure. The bottleneck is supply and not demand. Every dollar of compute they add prints money. Number 3: GPT-5.2 beats you at 74% of your job. Sam revealed OpenAI's internal GDP-Val benchmark. It measures how AI performs on knowledge work tasks across 40+ verticals. The results: GPT-5.2 beats or ties expert-level knowledge workers at 74.1% of tasks. Legal analysis. PowerPoint decks. Web apps. Financial modeling. Customer support. Sam's description: "A co-worker you can assign an hour's worth of tasks to and get something you prefer back 3 out of 4 times." Three years ago, ChatGPT launched at basically 0% on this scale. Now it's at 74%. And that's not GPT-6. That's what's available RIGHT NOW. Most companies haven't even started using this yet. But here's what Sam said about the gap between capability and adoption: "The overhang is going to be massive. Most people are still asking similar questions they did in the GPT-4 realm." Translation: The models can do 10x more than people have figured out how to use them for. Which means there's a HUGE arbitrage opportunity. Early adopters who actually integrate this into workflows will dominate their industries before competitors even understand what happened. Number 4: AGI already happened. And nobody noticed. Sam's exact quote: "AGI kind of went whooshing by. We're in this fuzzy period where some people think we have it and some don't." Read that again. The CEO of OpenAI just said AGI might have already arrived and we're arguing about definitions while it's actively replacing knowledge work. He even moved the goalposts. The new benchmark: "Superintelligence" = when AI can be a better president or CEO than any human. Not "as good as." BETTER than. We went from "can AI pass a Turing test" to "can AI run countries better than humans" in 3 years. So what does this actually mean? The AI revolution isn't about chatbots getting smarter. It's about the complete replacement of human intellectual output with machine output. At scale. Across every industry. Faster than anyone's prepared for. And the companies positioning for this RIGHT NOW are the ones printing money. OpenAI's enterprise growth is outpacing consumer because businesses see what's coming. They're not buying "AI tools." They're buying the ability to 10x output without 10x-ing headcount. Sam said they'll triple their compute next year. Then triple it again. Revenue is growing even faster than that. "We have never found a situation where we can't monetize all the compute we have." If he isn't lying then that's literally a printing press. The market still doesn't get it. Everyone's focused on "AI bubble" fears while OpenAI is solving the only problem that matters: turning compute into revenue at a faster rate than they're spending. They're not hoping demand catches up to supply. Demand is already 2x ahead of what they can deliver. Meanwhile, most knowledge workers are still using GPT-4 prompts on GPT-5.2. The capability overhang is massive. The arbitrage window is open. And it's closing fast. If you're running a B2B business and you're not integrating AI at the level Sam just described, you're not "waiting to see how it plays out." You're getting crushed by competitors who already figured it out. The companies that win in 2026 won't be the ones with the best AI. They'll be the ones who understood what Sam just laid out 6 months before everyone else did.

Ricardo

358,610 次观看 • 6 个月前

🤖🔬 Can AI actually do science end-to-end? 🧠📈 And how would we know when it matches, or surpasses, humans? ⚡🧪 AI is rapidly automating scientific discovery, but benchmarking full-cycle discovery, from 💡 ideation → 🧑‍💻 execution → 📊 conclusions, remains unsolved: 🧐🧐🧐 ❌🛠️ Open-ended discovery → manual validation (costly, unscalable) ❌📏 Metric-driven benchmarks (e.g., MLE-Bench) → convenient but narrow (is higher accuracy really enough?) ❌🤖⚖️ LLM-as-judge → useful, but fundamentally risky if used alone 🔥🚀 Introducing FIRE-Bench🔥: Fullcycle Insight Rediscovery Evaluation 👉🌐 📚✨ A benchmark that turns fresh, human-verified insights from recent 🏆 NeurIPS / ICLR / ICML papers into masked, end-to-end discovery challenges 🧩 🌍🔐 Constrained open-ended discovery–backed by ground truth. 📌 Key takeaways: 1⃣ 📖🧱 Reference-based evaluation still matters: constrained LLM judging helps, but human-grounded references remain essential until agents can consistently match human conclusions 2⃣ 🏆🧠 Expert-validated ground truth: all tasks come from recent NeurIPS / ICLR / ICML papers, with contamination carefully controlled 3⃣ 🔁🎭 Rediscovery, not reproduction: original 🧪 methods, 📊 experiments, 💻 implementations, and 📈 analyses are fully masked to create real discovery challenges 🔑 Key empirical findings: 💡 The "Science Gap" is Real: Even the best setup (Claude Code + Sonnet-4) caps out at an F1 score of 46.7. On hard tasks, agents struggle to break 30 💡 Success is a "Lottery": Performance has incredibly high variance. Reliability is a major unsolved issue. 💡 Coding is no longer the bottleneck; high-level reasoning and analysis are: ~74% of errors stem from flawed planning, not coding ⚙️ How it works: 🔹 Research-Problem Trees: We parse papers into trees (from broad roots to concrete leaves). This allows us to select intermediate nodes that perfectly balance open-ended exploration with verifiable ground truth. 🔹 Claim-Level Evaluation: We match AI conclusions against human conclusions using granular claim decomposition (F1 score). 🔹 Creativity Check: We score false positives to see if agents are finding novel truths (Spoiler🚨: they aren’t creative yet). 🔹 New Diagnostic Taxonomy: failures traced across four stages: 🧠 Planning → 🛠️ Implementation → ▶️ Execution → 🧾 Conclusion 🔹 Additional Analyses: cost efficiency, contamination checks, and more. 👀 The Future: 🚀 Live-FIRE-Bench: a live, continuously updated FIRE-Bench to track real-time progress on the latest research (Newest LLMs should be benchmarked with the newest research) 🚀 Stronger scaffolding (search + planning + coding) 🧠🧰 and converting FIRE-Bench into interactive environments for training research agents 🚀 Toward real creativity: We want better systems that can produce genuinely novel conclusions toward creativity 🎨⏳ 🚀 Better systems 🧠✨ and better benchmarks 📏 must co-evolve 🔄 over time 📜🎥 Paper, video, demo, and research trees: 👉🌐 #AI 🤖 #MachineLearning 📚 #AI4Science 🔬 #LLMs 🧠 #Research 🧪 #AgenticAI 🚀 #FireBench 🔥

🤖🔬 Can AI actually do science end-to-end? 🧠📈 And how would we know when it matches, or surpasses, humans? ⚡🧪 AI is rapidly automating scientific discovery, but benchmarking full-cycle discovery, from 💡 ideation → 🧑‍💻 execution → 📊 conclusions, remains unsolved: 🧐🧐🧐 ❌🛠️ Open-ended discovery → manual validation (costly, unscalable) ❌📏 Metric-driven benchmarks (e.g., MLE-Bench) → convenient but narrow (is higher accuracy really enough?) ❌🤖⚖️ LLM-as-judge → useful, but fundamentally risky if used alone 🔥🚀 Introducing FIRE-Bench🔥: Fullcycle Insight Rediscovery Evaluation 👉🌐 📚✨ A benchmark that turns fresh, human-verified insights from recent 🏆 NeurIPS / ICLR / ICML papers into masked, end-to-end discovery challenges 🧩 🌍🔐 Constrained open-ended discovery–backed by ground truth. 📌 Key takeaways: 1⃣ 📖🧱 Reference-based evaluation still matters: constrained LLM judging helps, but human-grounded references remain essential until agents can consistently match human conclusions 2⃣ 🏆🧠 Expert-validated ground truth: all tasks come from recent NeurIPS / ICLR / ICML papers, with contamination carefully controlled 3⃣ 🔁🎭 Rediscovery, not reproduction: original 🧪 methods, 📊 experiments, 💻 implementations, and 📈 analyses are fully masked to create real discovery challenges 🔑 Key empirical findings: 💡 The "Science Gap" is Real: Even the best setup (Claude Code + Sonnet-4) caps out at an F1 score of 46.7. On hard tasks, agents struggle to break 30 💡 Success is a "Lottery": Performance has incredibly high variance. Reliability is a major unsolved issue. 💡 Coding is no longer the bottleneck; high-level reasoning and analysis are: ~74% of errors stem from flawed planning, not coding ⚙️ How it works: 🔹 Research-Problem Trees: We parse papers into trees (from broad roots to concrete leaves). This allows us to select intermediate nodes that perfectly balance open-ended exploration with verifiable ground truth. 🔹 Claim-Level Evaluation: We match AI conclusions against human conclusions using granular claim decomposition (F1 score). 🔹 Creativity Check: We score false positives to see if agents are finding novel truths (Spoiler🚨: they aren’t creative yet). 🔹 New Diagnostic Taxonomy: failures traced across four stages: 🧠 Planning → 🛠️ Implementation → ▶️ Execution → 🧾 Conclusion 🔹 Additional Analyses: cost efficiency, contamination checks, and more. 👀 The Future: 🚀 Live-FIRE-Bench: a live, continuously updated FIRE-Bench to track real-time progress on the latest research (Newest LLMs should be benchmarked with the newest research) 🚀 Stronger scaffolding (search + planning + coding) 🧠🧰 and converting FIRE-Bench into interactive environments for training research agents 🚀 Toward real creativity: We want better systems that can produce genuinely novel conclusions toward creativity 🎨⏳ 🚀 Better systems 🧠✨ and better benchmarks 📏 must co-evolve 🔄 over time 📜🎥 Paper, video, demo, and research trees: 👉🌐 #AI 🤖 #MachineLearning 📚 #AI4Science 🔬 #LLMs 🧠 #Research 🧪 #AgenticAI 🚀 #FireBench 🔥

Zhen Wang

13,450 次观看 • 5 个月前

$AMD $5 Trillion MC Is Inevitable Long Term👑 This thread will focus more on Inference! 2026 EPYC "Venice" $TSM 2nm to save Large GW Scale Inference by 40% more than Prior Turin gen. Context: EPYC Turin achieves ~$0.001 per million tokens for batch inference vs $0.02-$0.12/ million tokens as I wrote the thread below. Venice is going to lower cost down to $0.0005-$0.0006/Million Tokens. OpenAI spent roughly $20B on Inference and Training, where 80-90% of that was for Inference per Analysts. AKA Renting Compute is Expensive AF! In this thread, I want to focus on why most analysts and investors are underestimating the role EPYC "Venice" and future Gen on overall Data center revenue. And $TSM ramping up 2nm supply early is a confirmation that AMD will be a major buyer long term. I will also link the thread the Gap between AMD Analysts & Reality and 2nm Ramp Thread so you have more comprehensive view of what I'm writing here. Before I go into detail this is my 2026 Projection: AI GPUs: $35-$50B EPYC Data Center: $15B-$17B Client Segment: $12-$13B Gaming: $6B Embedded: $4B-$5B Total Revenue $70-$100B Non-GAAP net income $18B-$25B Non-GAAP EPS $10.97-$15.40 Foward P/E 55x-70x= $603-$1,078 AMD's Analysts are projecting $0 Revenue for MI450 and sluggish EPYC Growth. Meaning, all analysts are either full of 💩 or Sexist, you decide! Analysts are also projecting 0% growth on AMD "Secret Weapon" Chip as $MSFT said we are at significant Windows refresh and upgrade cycle. Do you think TSMC would allocate more 2nm supply to $AMD at $0 MI450 revenue and sluggish EPYC? 1. EPYC is going to be the leader in lowest Inference! Current Turin cost saving is 95% vs $NVDA or 98-99% on Inference cost when you factor in renting Inference compute from Amazon Web Services, Microsoft Azure, or $NVDA Neocloud pets. TSMC claimed: 10-15% higher performance at iso-power, 25-30% lower power at iso-speed, and ~15% higher transistor density compared to 3nm. This reduces operational expenses (energy, cooling) while increasing throughput per chip. EPYC Turin achieves ~$0.001 per million tokens for batch inference (via vLLM on models like Llama 3 70B), driven by high core counts and low hardware costs. EPYC Venice offers ~1.7x overall performance and up to 70% more compute capability per core, with up to 256 cores (512 threads). Enhanced vector/AI instructions and open-source firmware (openSIL) optimize for inference workloads. AMD Incorporates AI Engines (now part of AMD's XDNA) for on-chip acceleration, improving efficiency for low-latency and edge inference. This reduces reliance on discrete GPUs, lowering system complexity and TCO. Venice SKUs are projected at $3,000-$15,000 ($5,000 for 256-core flagship), far below NVIDIA Rubin ($50,000-$90,000) or AMD's own MI450 GPUs ($40,000-$50,000). High memory bandwidth (up to 1.6 TB/s) supports efficient batch inference. Venice is designed exactly for Large customers that want to lower Inference Cost and MI450 Helios is for Customers that want Training at lowest TCO, TDP as well as lower Upfront 1GW scale(Full build $35-$40B vs $NVDA $55B-$80B). 2. Real World Example: OpenAI's 2025 inference spend reached ~$20B, escalating to even higher total compute rental (mostly inference) amid token volume growth(from video generating). By 2026, with usage doubling (consistent with industry trends: token demand grows 2-5x YoY), assume OpenAI processes ~1,800 billion million-tokens annually $NVDA Blackwell at $0.02-$0.12 is $36B(most optimized) Rubin is projected to be at $0.01/million tokens or $18B annual Inference Cost vs $AMD Venice $0.0005/million tokens or $0.9B annual Inference Cost => Massive saving for OpenAI or anyone that are paying 80-90% Annual Bill for Inference compute. In short, it is unsustainable to pay this much rent vs owning for all current AI players for the medium to long term. Rubin excels in low-latency decode (if Groq integration from $20B deal in 2027-2028), but Venice dominates batch (80% of inference by 2030). Actual savings depend on deployment scale (OpenAI's 6GW AMD plans), electricity rates, and software maturity. If Rubin only hits $0.03, savings swell to $53.1B vs. $17.1B. 3. Will running Inference on Venice and future Gen slow down response generation in 2026 and beyond? Human perception of "fast enough" for chat, agents, search augmentation, summarization, coding assistance is roughly Meaning, EPYC may generate $100B a year on data center revenue, Hence $MSFT $AMZN $META $GOOGL OpenAI xAI and 42+ Countries are leaning AMD for Inference, because the cost saving is MASSIVE! 4. Regular users (you, me, people using ChatGPT, Claude, Gemini, Grok, Perplexity...) are extremely unlikely to notice any slowdown and in many cases might even experience slightly faster or more consistent response times if the industry heavily shifts toward AMD EPYC for inference. What actually happens when companies save massively on inference? When OpenAI , Anthropic , Gemini , Grok Meta .... save billions on the batch/enterprise/RAG layer using EPYC Venice, they typically do one or more of these things with the savings, none of which make your chat slower but enhancing their bottom line(Profit) ~Keep prices the same → make more profit ~Lower subscription prices / increase free tier limits ~Train bigger & better models more frequently ~Offer longer context windows ~Add more reasoning steps / tool calls / agents per query ~Improve multimodal capabilities ~Build more data centers / reduce throttling during peaks In practice the consumer experience usually gets better, not worse, when inference becomes dramatically cheaper. Prime example is $META leaning AMD heavily or currently AMD largest customer. or Grok 2 to Grok 3 heavily used AMD for Inference saving. And most Grok Users reported Groke responses snappier, not slower. 5. What does this mean for potential Revenue? Noted that TSMC is massively ramping 2nm supply for $AMD both MI450 and EPYC. EPYC Conservative projection: FY2025: $10.5B(best Est) FY2026: $16B FY2027: $29B FY2028: $49B FY2029: $75B FY2030: $100B Large customers: $META OpenAI $MSFT $AMZN $GOOGL xAI (Apple?) Smaller customer: $DELL $HPE $SMCI and 42+ other countries. The roadmap to $5 Trillion is very much inevitable as Inference Cost from Renting or owning $NVDA are too high, but $NVDA will still dominate Training market share, where MI families are likely to take 15-20% market share, but the TAM is also expanding Rapidly. Most Institutions are projecting $2-$3Trillion TAM by 2030. $NVDA said $4 Trillion. Dr. Lisa Su said $1 Trillion+ by 2030. So you decide on how much TAM. If you enjoy this kind of analysis, Slap the Like/Repost and Bookmark to please the X Algo as it is Free.99! If you want to support my work further, consider subscribe to see more in-depth analysis! Alright, that is it. Not Financial Advice!

$AMD $5 Trillion MC Is Inevitable Long Term👑 This thread will focus more on Inference! 2026 EPYC "Venice" $TSM 2nm to save Large GW Scale Inference by 40% more than Prior Turin gen. Context: EPYC Turin achieves ~$0.001 per million tokens for batch inference vs $0.02-$0.12/ million tokens as I wrote the thread below. Venice is going to lower cost down to $0.0005-$0.0006/Million Tokens. OpenAI spent roughly $20B on Inference and Training, where 80-90% of that was for Inference per Analysts. AKA Renting Compute is Expensive AF! In this thread, I want to focus on why most analysts and investors are underestimating the role EPYC "Venice" and future Gen on overall Data center revenue. And $TSM ramping up 2nm supply early is a confirmation that AMD will be a major buyer long term. I will also link the thread the Gap between AMD Analysts & Reality and 2nm Ramp Thread so you have more comprehensive view of what I'm writing here. Before I go into detail this is my 2026 Projection: AI GPUs: $35-$50B EPYC Data Center: $15B-$17B Client Segment: $12-$13B Gaming: $6B Embedded: $4B-$5B Total Revenue $70-$100B Non-GAAP net income $18B-$25B Non-GAAP EPS $10.97-$15.40 Foward P/E 55x-70x= $603-$1,078 AMD's Analysts are projecting $0 Revenue for MI450 and sluggish EPYC Growth. Meaning, all analysts are either full of 💩 or Sexist, you decide! Analysts are also projecting 0% growth on AMD "Secret Weapon" Chip as $MSFT said we are at significant Windows refresh and upgrade cycle. Do you think TSMC would allocate more 2nm supply to $AMD at $0 MI450 revenue and sluggish EPYC? 1. EPYC is going to be the leader in lowest Inference! Current Turin cost saving is 95% vs $NVDA or 98-99% on Inference cost when you factor in renting Inference compute from Amazon Web Services, Microsoft Azure, or $NVDA Neocloud pets. TSMC claimed: 10-15% higher performance at iso-power, 25-30% lower power at iso-speed, and ~15% higher transistor density compared to 3nm. This reduces operational expenses (energy, cooling) while increasing throughput per chip. EPYC Turin achieves ~$0.001 per million tokens for batch inference (via vLLM on models like Llama 3 70B), driven by high core counts and low hardware costs. EPYC Venice offers ~1.7x overall performance and up to 70% more compute capability per core, with up to 256 cores (512 threads). Enhanced vector/AI instructions and open-source firmware (openSIL) optimize for inference workloads. AMD Incorporates AI Engines (now part of AMD's XDNA) for on-chip acceleration, improving efficiency for low-latency and edge inference. This reduces reliance on discrete GPUs, lowering system complexity and TCO. Venice SKUs are projected at $3,000-$15,000 ($5,000 for 256-core flagship), far below NVIDIA Rubin ($50,000-$90,000) or AMD's own MI450 GPUs ($40,000-$50,000). High memory bandwidth (up to 1.6 TB/s) supports efficient batch inference. Venice is designed exactly for Large customers that want to lower Inference Cost and MI450 Helios is for Customers that want Training at lowest TCO, TDP as well as lower Upfront 1GW scale(Full build $35-$40B vs $NVDA $55B-$80B). 2. Real World Example: OpenAI's 2025 inference spend reached ~$20B, escalating to even higher total compute rental (mostly inference) amid token volume growth(from video generating). By 2026, with usage doubling (consistent with industry trends: token demand grows 2-5x YoY), assume OpenAI processes ~1,800 billion million-tokens annually $NVDA Blackwell at $0.02-$0.12 is $36B(most optimized) Rubin is projected to be at $0.01/million tokens or $18B annual Inference Cost vs $AMD Venice $0.0005/million tokens or $0.9B annual Inference Cost => Massive saving for OpenAI or anyone that are paying 80-90% Annual Bill for Inference compute. In short, it is unsustainable to pay this much rent vs owning for all current AI players for the medium to long term. Rubin excels in low-latency decode (if Groq integration from $20B deal in 2027-2028), but Venice dominates batch (80% of inference by 2030). Actual savings depend on deployment scale (OpenAI's 6GW AMD plans), electricity rates, and software maturity. If Rubin only hits $0.03, savings swell to $53.1B vs. $17.1B. 3. Will running Inference on Venice and future Gen slow down response generation in 2026 and beyond? Human perception of "fast enough" for chat, agents, search augmentation, summarization, coding assistance is roughly Meaning, EPYC may generate $100B a year on data center revenue, Hence $MSFT $AMZN $META $GOOGL OpenAI xAI and 42+ Countries are leaning AMD for Inference, because the cost saving is MASSIVE! 4. Regular users (you, me, people using ChatGPT, Claude, Gemini, Grok, Perplexity...) are extremely unlikely to notice any slowdown and in many cases might even experience slightly faster or more consistent response times if the industry heavily shifts toward AMD EPYC for inference. What actually happens when companies save massively on inference? When OpenAI , Anthropic , Gemini , Grok Meta .... save billions on the batch/enterprise/RAG layer using EPYC Venice, they typically do one or more of these things with the savings, none of which make your chat slower but enhancing their bottom line(Profit) ~Keep prices the same → make more profit ~Lower subscription prices / increase free tier limits ~Train bigger & better models more frequently ~Offer longer context windows ~Add more reasoning steps / tool calls / agents per query ~Improve multimodal capabilities ~Build more data centers / reduce throttling during peaks In practice the consumer experience usually gets better, not worse, when inference becomes dramatically cheaper. Prime example is $META leaning AMD heavily or currently AMD largest customer. or Grok 2 to Grok 3 heavily used AMD for Inference saving. And most Grok Users reported Groke responses snappier, not slower. 5. What does this mean for potential Revenue? Noted that TSMC is massively ramping 2nm supply for $AMD both MI450 and EPYC. EPYC Conservative projection: FY2025: $10.5B(best Est) FY2026: $16B FY2027: $29B FY2028: $49B FY2029: $75B FY2030: $100B Large customers: $META OpenAI $MSFT $AMZN $GOOGL xAI (Apple?) Smaller customer: $DELL $HPE $SMCI and 42+ other countries. The roadmap to $5 Trillion is very much inevitable as Inference Cost from Renting or owning $NVDA are too high, but $NVDA will still dominate Training market share, where MI families are likely to take 15-20% market share, but the TAM is also expanding Rapidly. Most Institutions are projecting $2-$3Trillion TAM by 2030. $NVDA said $4 Trillion. Dr. Lisa Su said $1 Trillion+ by 2030. So you decide on how much TAM. If you enjoy this kind of analysis, Slap the Like/Repost and Bookmark to please the X Algo as it is Free.99! If you want to support my work further, consider subscribe to see more in-depth analysis! Alright, that is it. Not Financial Advice!

Mike

102,223 次观看 • 6 个月前

🚀 Introducing PantheonOS ( A Fully Open-Source Agent OS for Science PantheonOS began as a research project in my Stanford lab and has since evolved into a vision to redefine data science in the era of AI—starting with computational biology, especially single-cell and spatial genomics. PantheonOS is a general agent platform built from the ground up. It is arguably the first distributed agent framework designed for scientific data analysis. 🔑 Key Features 1. Multi-Agent Collaboration – Built-in paradigms for distributed, cross-machine cooperation among agents and toolsets. 2. Native Toolset Support – Python, R, Julia, LaTeX, and more—designed for real scientific workflows. 3. Modular & Extensible – Developer-friendly design with shallow wrappers, plus LLM-driven toolset generation. 4. Evolvable Agents – Capable of evolving large-scale code projects to achieve superhuman performance (e.g., evolving upon the original Harmony [I Korsunsky, 2019, Nature Biotechnology] and Scanorama [BL Hie, 2019, Nature Biotechnology] implementations), and even evolving the system itself to adapt to new fields. 🎉 Stepwise Release Strategy We’re releasing PantheonOS in stages: Pantheon-CLI (today!), followed by Pantheon-Lab, Pantheon-Notebook, Pantheon-Slack, and more. 🌟 Pantheon-CLI Highlights - We're not just building another CLI tool. We're defining how scientists will interact with data in the AI era. - Open, Powerful, Python-First – The first fully open-source, endlessly extendable scientific “vibe analysis” framework. - Mixed Programming Magic – Combine Python, natural language, R, or Julia—seamlessly in the same environment. - PhD-Level Assistant – A command-line agent for complex real-world genomics and beyond, handling workflows at the PhD level. - Privacy by Design – Run entirely offline with local LLMs—your data never leaves your computer. ✅ Proven Applications (10 Demonstrations) Computational biology: 1. ATAC-seq: From raw reads to peak matrix 2. RNA-seq: From raw reads to expression matrix 3. Complex single-cell workflows (PhD-level) 4. Hybrid natural language + R for Seurat annotation 5. Learning from web tutorials + invoking single-cell foundation models 6. Cell segmentation on 10x Genomics HD Visium data And beyond: 7. Mixed Python & R programming examples 8. Molecular docking & structural analysis 9. Exploratory factor analysis for behavioral survey data 10. Customer segmentation & finance analytics 🌐 Learn More & Get Started Website: Pantheon-CLI Documentation: GitHub Repo: 💬 Join our community: PantheonOS Slack: PantheonOS Discord:

🚀 Introducing PantheonOS ( A Fully Open-Source Agent OS for Science PantheonOS began as a research project in my Stanford lab and has since evolved into a vision to redefine data science in the era of AI—starting with computational biology, especially single-cell and spatial genomics. PantheonOS is a general agent platform built from the ground up. It is arguably the first distributed agent framework designed for scientific data analysis. 🔑 Key Features 1. Multi-Agent Collaboration – Built-in paradigms for distributed, cross-machine cooperation among agents and toolsets. 2. Native Toolset Support – Python, R, Julia, LaTeX, and more—designed for real scientific workflows. 3. Modular & Extensible – Developer-friendly design with shallow wrappers, plus LLM-driven toolset generation. 4. Evolvable Agents – Capable of evolving large-scale code projects to achieve superhuman performance (e.g., evolving upon the original Harmony [I Korsunsky, 2019, Nature Biotechnology] and Scanorama [BL Hie, 2019, Nature Biotechnology] implementations), and even evolving the system itself to adapt to new fields. 🎉 Stepwise Release Strategy We’re releasing PantheonOS in stages: Pantheon-CLI (today!), followed by Pantheon-Lab, Pantheon-Notebook, Pantheon-Slack, and more. 🌟 Pantheon-CLI Highlights - We're not just building another CLI tool. We're defining how scientists will interact with data in the AI era. - Open, Powerful, Python-First – The first fully open-source, endlessly extendable scientific “vibe analysis” framework. - Mixed Programming Magic – Combine Python, natural language, R, or Julia—seamlessly in the same environment. - PhD-Level Assistant – A command-line agent for complex real-world genomics and beyond, handling workflows at the PhD level. - Privacy by Design – Run entirely offline with local LLMs—your data never leaves your computer. ✅ Proven Applications (10 Demonstrations) Computational biology: 1. ATAC-seq: From raw reads to peak matrix 2. RNA-seq: From raw reads to expression matrix 3. Complex single-cell workflows (PhD-level) 4. Hybrid natural language + R for Seurat annotation 5. Learning from web tutorials + invoking single-cell foundation models 6. Cell segmentation on 10x Genomics HD Visium data And beyond: 7. Mixed Python & R programming examples 8. Molecular docking & structural analysis 9. Exploratory factor analysis for behavioral survey data 10. Customer segmentation & finance analytics 🌐 Learn More & Get Started Website: Pantheon-CLI Documentation: GitHub Repo: 💬 Join our community: PantheonOS Slack: PantheonOS Discord:

evo-devo

17,350 次观看 • 11 个月前

The world of writing has changed forever. AI is getting really good, really fast. ChatGPT is already a better writer than most humans and some professional writers. So, what’s the future of writing? 18 thoughts from Tyler Cowen: 1) Don't let AI smooth out your idiosyncrasies. Let your writing stay weird and uniquely yours. 2) Generic content is dying and the burden is on you as the writer to be distinctive. 3) The more personal your writing becomes, the more future-proof it is. Nobody wants to read memoirs from AI, even if they're technically "better." 4) Use AI as your secondary literature when you read — not just for quick answers, but as a thinking companion. As Tyler puts it, "I'll keep on asking the AI: 'What do you think of chapter two? What happened there? What are some puzzles?' It just gets me thinking... and I'm smarter about the thing in the final analysis." 5) Hallucinations aren't the crisis everyone makes them out to be. No matter the source, if you're going to use a piece of information, you should double-check it. This is true for both books and AI. 6) Secrets will become more valuable in an AI-driven world. 7) One way to use AI as a writer is to research fields you aren't as familiar with before you start writing about them. Tyler said: "I just wrote a column about declassifying classified documents. I don't know that law very well. I asked the AI for a lot of background... now I feel like I'm not an idiot on the topic." 8) AI changes what books are even worth writing. "Predictive books and books about the near future. They don't make sense to write anymore." 9) Editing trick: Try running your writing through AI and asking what some people might find obnoxious. It’s a surprisingly powerful editing trick. 10) When prompting AI, put humans out of your mind and imagine you're talking to an alien or a non-human animal. 11) Many of the most significant AI advancements are likely happening behind closed doors. For example, I hear that Google allows employees to use Gemini with virtually unlimited context windows. 12) What possibilities do large context windows open up? Researchers will be able to load entire regulatory frameworks, historical archives, or massive datasets like "tax records from Renaissance Florence" into a single query. 13) The rate of AI improvement matters more than its current capabilities. As Tyler puts it, "This is the worst they will ever be" is key to understanding their trajectory. "A lot of people don't get that. They're impressed by what they see in the moment, but they don't understand the rate of improvement." 14) The best way to appreciate the current rate of improvement is to use the latest models. 15) Being non-technical can sometimes be an advantage when thinking about AI. Here’s Tyler: "If you're not focused on the technical side, you will see other things more clearly... You just focus on what is this actually good for? And not, am I impressed by all the neat bells and whistles on this advance with AI?" 16) How Tyler uses AI to prep for podcast interviews: Don't waste time asking AI for generic interview questions or broad topics. Tyler says that's the worst question you can ask an AI. It’s “too normy.” Instead, ask specific questions about historical examples and get context. Then, let your own creative questions emerge. 17) Your relationship with mentors and peers becomes more crucial, not less, in an AI world. "Two pieces of general advice with or without AI in the world." Tyler says: "Get more and better mentors and work every day at improving the quality of your peer network." 18) The divide between AI and humans creates a striking paradox. As Tyler puts it: "On one hand the AIs are getting so much better, so learn how to use the AIs. On the other hand, the AIs are getting so much better, so invest in these other things that aren't AI—pure networks. You've gotta do both." I've shared the full conversation with tylercowen below. In the replies, I've also linked to a full transcript and relevant links to YouTube, Spotify, and Apple Podcasts if you want to listen there. And if you want a bite-size entry to the episode, I've shared some clips in the replies too.

The world of writing has changed forever. AI is getting really good, really fast. ChatGPT is already a better writer than most humans and some professional writers. So, what’s the future of writing? 18 thoughts from Tyler Cowen: 1) Don't let AI smooth out your idiosyncrasies. Let your writing stay weird and uniquely yours. 2) Generic content is dying and the burden is on you as the writer to be distinctive. 3) The more personal your writing becomes, the more future-proof it is. Nobody wants to read memoirs from AI, even if they're technically "better." 4) Use AI as your secondary literature when you read — not just for quick answers, but as a thinking companion. As Tyler puts it, "I'll keep on asking the AI: 'What do you think of chapter two? What happened there? What are some puzzles?' It just gets me thinking... and I'm smarter about the thing in the final analysis." 5) Hallucinations aren't the crisis everyone makes them out to be. No matter the source, if you're going to use a piece of information, you should double-check it. This is true for both books and AI. 6) Secrets will become more valuable in an AI-driven world. 7) One way to use AI as a writer is to research fields you aren't as familiar with before you start writing about them. Tyler said: "I just wrote a column about declassifying classified documents. I don't know that law very well. I asked the AI for a lot of background... now I feel like I'm not an idiot on the topic." 8) AI changes what books are even worth writing. "Predictive books and books about the near future. They don't make sense to write anymore." 9) Editing trick: Try running your writing through AI and asking what some people might find obnoxious. It’s a surprisingly powerful editing trick. 10) When prompting AI, put humans out of your mind and imagine you're talking to an alien or a non-human animal. 11) Many of the most significant AI advancements are likely happening behind closed doors. For example, I hear that Google allows employees to use Gemini with virtually unlimited context windows. 12) What possibilities do large context windows open up? Researchers will be able to load entire regulatory frameworks, historical archives, or massive datasets like "tax records from Renaissance Florence" into a single query. 13) The rate of AI improvement matters more than its current capabilities. As Tyler puts it, "This is the worst they will ever be" is key to understanding their trajectory. "A lot of people don't get that. They're impressed by what they see in the moment, but they don't understand the rate of improvement." 14) The best way to appreciate the current rate of improvement is to use the latest models. 15) Being non-technical can sometimes be an advantage when thinking about AI. Here’s Tyler: "If you're not focused on the technical side, you will see other things more clearly... You just focus on what is this actually good for? And not, am I impressed by all the neat bells and whistles on this advance with AI?" 16) How Tyler uses AI to prep for podcast interviews: Don't waste time asking AI for generic interview questions or broad topics. Tyler says that's the worst question you can ask an AI. It’s “too normy.” Instead, ask specific questions about historical examples and get context. Then, let your own creative questions emerge. 17) Your relationship with mentors and peers becomes more crucial, not less, in an AI world. "Two pieces of general advice with or without AI in the world." Tyler says: "Get more and better mentors and work every day at improving the quality of your peer network." 18) The divide between AI and humans creates a striking paradox. As Tyler puts it: "On one hand the AIs are getting so much better, so learn how to use the AIs. On the other hand, the AIs are getting so much better, so invest in these other things that aren't AI—pure networks. You've gotta do both." I've shared the full conversation with tylercowen below. In the replies, I've also linked to a full transcript and relevant links to YouTube, Spotify, and Apple Podcasts if you want to listen there. And if you want a bite-size entry to the episode, I've shared some clips in the replies too.

David Perell

175,011 次观看 • 1 年前

––Mathias Döpfner: Sam, is it actually true that your kind of favorite book is The Beginning of Infinity of Dieter Deutsch? Sam Altman: Yeah, I think if I had to pick one favorite book, I would pick that. ––Mathias Döpfner: Why is that so fascinating? Can you explain that? Sam Altman: Even if you don't read the whole thing, the first like 40-50 pages are, I think, the most wonderfully optimistic take on why, even in a world with AI, we're never going to run out of things to do and ways to be useful and problems to solve and things to explore. But I also think it explains so beautifully how we got here and why the relatively simple process that we've followed throughout human history got us to this incredible place. ––Mathias Döpfner: Okay, that's good, because David Deutsch, I think, is going to be our last virtual guest, at least tonight. David Deutsch is a physicist and scientist from Oxford University. And I think also you have disagreements with him about the possibility that artificial intelligence is transforming into superintelligence with consciousness, perhaps even. He thinks it cannot be the case. You think it should be the case. Here he is, David Deutsch. Welcome. And perhaps you can elaborate a little bit on that disagreement, but also why you admire Sam Altman. Sam Altman: Well, I don't care about that. I just want to hear your disagreement. David Deutsch: Okay, I can tell you. Well, on my computer, I keep a list of progress that has been achieved where I had previously been sure that it wasn't yet possible. One of the items I'm embarrassed to admit was the World Wide Web. Another was that I thought that no computer program would be able to sustain open-ended conversation on general subjects in natural language unless that program was an AGI, an artificial general intelligence. So it would have, I prefer to call, explanatory creativity. ChatGPT proved me wrong. It's not an AGI, and it can converse. That ability was a side effect of another, namely knowledge. The Eliza chatbot in the 1960s used little more than the words and phrases you told it. ChatGPT can chat about anything drawing on a vast body of knowledge, which was a phenomenally useful combination. For some people, too useful. They think they're speaking to a person, an AGI, just as the first users of Eliza treated it as if it were a person. Which brings me to a widespread myth of the Turing test. In reality, Alan Turing never proposed a test or benchmark for AGI. His imitation game wasn't a test of ethics, but a thought experiment to torpedo the intuition that machines can't think. Indeed, there can be no benchmark, because to be general, an AGI must be capable of choosing to remain silent. This is already a proof that AGI cannot be made via existing approaches, while those can and must be judged by benchmarks. Conversely, if something outputs a new explanation, you can't test for whether it created that or a human did, even you yourself when you administered the test. In Edison's phrase, there's the inspiration part, which only humans and AGIs can do, and the perspiration part, from which AGIs can liberate us. So, if there's no test, how do we know that humans are general intelligences? By telling their story. Human thought doesn't consist of mechanically converting motivations into actions, prompts into output. It's mainly about choosing motivations. Just as science is not extracting theories from data, it's seeing a problem, guessing explanations, criticizing and testing them. So how can you tell whether something is doing that? You can't, always. Sometimes it really is a bot you're chatting to, but when you have no explanation saying that you yourself are a bot, or that humans in general are, it's rational to assume that they aren't. Some people have fun questioning whether Einstein really created the theory of relativity or only assembled it mechanically from a smorgasbord of existing ideas. We know he created it because we know his story, what problems he was addressing, and why. Just as we know that Sam Altman, without having to write any code, brought ChatGPT into existence as a product and a phenomenon by having the intuition and the gumption to know that this was the right thing for humanity to try next. Nothing can program a computer to have such intuitions, yet. Sam Altman: Can I ask one question? David Deutsch: My guess. Sam Altman: You mentioned Einstein and general relativity, and I agree, I think that's one of the most beautiful things humanity's ever figured out. Maybe I would even say number one. And Einstein had a story, we knew what he was working on. If in a few years, GPT-8 figured out quantum gravity and could tell you its story of how it did it and the problems it was thinking about and why it decided to work on it, would But it still just looked like a language model output, but it was the real, it really did solve it. Would you call it like, then would you say, I appreciate that you keep a list of things you're wrong about. I do too. But would that be enough to convince you? David Deutsch: I think it would. Yeah. Sam Altman: All right. I'll take you up... David Deutsch: It's crucial here. Sam Altman: I agree to that as the test. ––Mathias Döpfner: David, thank you so much for joining us and thank you for your uplifting words and have a great evening. David Deutsch, a pioneer of quantum computing, one of the most brilliant thinkers of our times. Thank you for joining.

––Mathias Döpfner: Sam, is it actually true that your kind of favorite book is The Beginning of Infinity of Dieter Deutsch? Sam Altman: Yeah, I think if I had to pick one favorite book, I would pick that. ––Mathias Döpfner: Why is that so fascinating? Can you explain that? Sam Altman: Even if you don't read the whole thing, the first like 40-50 pages are, I think, the most wonderfully optimistic take on why, even in a world with AI, we're never going to run out of things to do and ways to be useful and problems to solve and things to explore. But I also think it explains so beautifully how we got here and why the relatively simple process that we've followed throughout human history got us to this incredible place. ––Mathias Döpfner: Okay, that's good, because David Deutsch, I think, is going to be our last virtual guest, at least tonight. David Deutsch is a physicist and scientist from Oxford University. And I think also you have disagreements with him about the possibility that artificial intelligence is transforming into superintelligence with consciousness, perhaps even. He thinks it cannot be the case. You think it should be the case. Here he is, David Deutsch. Welcome. And perhaps you can elaborate a little bit on that disagreement, but also why you admire Sam Altman. Sam Altman: Well, I don't care about that. I just want to hear your disagreement. David Deutsch: Okay, I can tell you. Well, on my computer, I keep a list of progress that has been achieved where I had previously been sure that it wasn't yet possible. One of the items I'm embarrassed to admit was the World Wide Web. Another was that I thought that no computer program would be able to sustain open-ended conversation on general subjects in natural language unless that program was an AGI, an artificial general intelligence. So it would have, I prefer to call, explanatory creativity. ChatGPT proved me wrong. It's not an AGI, and it can converse. That ability was a side effect of another, namely knowledge. The Eliza chatbot in the 1960s used little more than the words and phrases you told it. ChatGPT can chat about anything drawing on a vast body of knowledge, which was a phenomenally useful combination. For some people, too useful. They think they're speaking to a person, an AGI, just as the first users of Eliza treated it as if it were a person. Which brings me to a widespread myth of the Turing test. In reality, Alan Turing never proposed a test or benchmark for AGI. His imitation game wasn't a test of ethics, but a thought experiment to torpedo the intuition that machines can't think. Indeed, there can be no benchmark, because to be general, an AGI must be capable of choosing to remain silent. This is already a proof that AGI cannot be made via existing approaches, while those can and must be judged by benchmarks. Conversely, if something outputs a new explanation, you can't test for whether it created that or a human did, even you yourself when you administered the test. In Edison's phrase, there's the inspiration part, which only humans and AGIs can do, and the perspiration part, from which AGIs can liberate us. So, if there's no test, how do we know that humans are general intelligences? By telling their story. Human thought doesn't consist of mechanically converting motivations into actions, prompts into output. It's mainly about choosing motivations. Just as science is not extracting theories from data, it's seeing a problem, guessing explanations, criticizing and testing them. So how can you tell whether something is doing that? You can't, always. Sometimes it really is a bot you're chatting to, but when you have no explanation saying that you yourself are a bot, or that humans in general are, it's rational to assume that they aren't. Some people have fun questioning whether Einstein really created the theory of relativity or only assembled it mechanically from a smorgasbord of existing ideas. We know he created it because we know his story, what problems he was addressing, and why. Just as we know that Sam Altman, without having to write any code, brought ChatGPT into existence as a product and a phenomenon by having the intuition and the gumption to know that this was the right thing for humanity to try next. Nothing can program a computer to have such intuitions, yet. Sam Altman: Can I ask one question? David Deutsch: My guess. Sam Altman: You mentioned Einstein and general relativity, and I agree, I think that's one of the most beautiful things humanity's ever figured out. Maybe I would even say number one. And Einstein had a story, we knew what he was working on. If in a few years, GPT-8 figured out quantum gravity and could tell you its story of how it did it and the problems it was thinking about and why it decided to work on it, would But it still just looked like a language model output, but it was the real, it really did solve it. Would you call it like, then would you say, I appreciate that you keep a list of things you're wrong about. I do too. But would that be enough to convince you? David Deutsch: I think it would. Yeah. Sam Altman: All right. I'll take you up... David Deutsch: It's crucial here. Sam Altman: I agree to that as the test. ––Mathias Döpfner: David, thank you so much for joining us and thank you for your uplifting words and have a great evening. David Deutsch, a pioneer of quantum computing, one of the most brilliant thinkers of our times. Thank you for joining.

Deutsch Explains

63,456 次观看 • 9 个月前