正在加载视频...

视频加载失败

返回首页

We are very excited to release zerank-2, ZeroEntropy (YC W25) 's newest reranker model. 🔥 It shows major improvement on the 5 most common RAG failure modes below. Existing rerankers consistently fail on seemingly “simple” tasks: 🔢 Comparing numbers and date: “Biggest deals closed after 04/2024.” 🗄️ Aggregation: “Top...

ghita__ha's profile picture

Ghita

2,902 subscribers

88,524 次观看 • 10 个月前 •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...
3:28

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

kwindla

330,647 次观看 • 24 天前

We made a thing! Very happy to announce sqlcoder-pro and the Defog Alignment Platform. Available to use immediately without a wait-list, weights will be open-sourced very soon. The video does a quick show and tell comparison against ChatGPT (with gpt-4o). Read on for more details! TLDR 💪 equal (or better) performance on text-to-SQL as the most capable Claude-3.5 or GPT-4 models 🤝 You can use it today on a free plan/free trial, without a waitlist 🪽 self-hostable on a single RTX4090, with 2 second median generation times for SQL queries 🔁 exactly the same output every time, give the same prompt 👨🏻‍🏫 teachable and steerable: show the model what you want it to do 🛞 debuggable – you can understand WTF is going on inside the model, instead of treating it like a black box Let's dig into each of these one-by-one! Performance SQLCoder-8b-pro significantly exceeds the performance of our previous sqlcoder-8b model on Postgres text-to-SQL (from 88.2% to 90.2% accuracy - gpt-4o is at 87.6%, for reference). It is also better at following instructions. This was done via self-merges, hand crafted fine-tuning data, and adapting the training data to fit our tokenizer. Cost You can host this on the model on a single $3,500 RTX4090, and support ~5 requests/second via VLLM. If you're looking to host on the cloud instead, you can run it on a single L4 GPU that costs $300/mo on GCP Repeatability We have a dense 8b model with no MoE shenanigans. For the same prompt with temperature=0, you'll always get the same answer – which is critical in BI. Teachable In our alignment and feedback modes, you can give the model feedback on how it answered certain questions, and it will automatically adapt to the feedback. Debuggable You can use logprobs and attention scores to determine where, exactly is the model paying attention to inside a prompt + what it's getting confused by when generating outputs. Available today You can use Defog on the cloud today by going to docs[dot]defog[dot]ai, and getting an API key. Excited to hear what you think!
2:33

We made a thing! Very happy to announce sqlcoder-pro and the Defog Alignment Platform. Available to use immediately without a wait-list, weights will be open-sourced very soon. The video does a quick show and tell comparison against ChatGPT (with gpt-4o). Read on for more details! TLDR 💪 equal (or better) performance on text-to-SQL as the most capable Claude-3.5 or GPT-4 models 🤝 You can use it today on a free plan/free trial, without a waitlist 🪽 self-hostable on a single RTX4090, with 2 second median generation times for SQL queries 🔁 exactly the same output every time, give the same prompt 👨🏻‍🏫 teachable and steerable: show the model what you want it to do 🛞 debuggable – you can understand WTF is going on inside the model, instead of treating it like a black box Let's dig into each of these one-by-one! Performance SQLCoder-8b-pro significantly exceeds the performance of our previous sqlcoder-8b model on Postgres text-to-SQL (from 88.2% to 90.2% accuracy - gpt-4o is at 87.6%, for reference). It is also better at following instructions. This was done via self-merges, hand crafted fine-tuning data, and adapting the training data to fit our tokenizer. Cost You can host this on the model on a single $3,500 RTX4090, and support ~5 requests/second via VLLM. If you're looking to host on the cloud instead, you can run it on a single L4 GPU that costs $300/mo on GCP Repeatability We have a dense 8b model with no MoE shenanigans. For the same prompt with temperature=0, you'll always get the same answer – which is critical in BI. Teachable In our alignment and feedback modes, you can give the model feedback on how it answered certain questions, and it will automatically adapt to the feedback. Debuggable You can use logprobs and attention scores to determine where, exactly is the model paying attention to inside a prompt + what it's getting confused by when generating outputs. Available today You can use Defog on the cloud today by going to docs[dot]defog[dot]ai, and getting an API key. Excited to hear what you think!

Rishabh Srivastava

13,469 次观看 • 2 年前

learned a lot from this conversation with Simon Mo and Matt Bornstein. biggest takeaways for me: -there are a lot of reasons why we should like open-weight models. a lot of these arguments stop at handwavy things like "what if the labs stop releasing frontier models to the public" or "it's lower cost." but simon's position as lead maintainer of vLLM and CEO of Inferact give him authority to talk about some of the other, more interesting and concrete reasons to pay attention to open-weight models, namely that they allow end-users to calibrate latency / other performance metrics with way more customizability than what any of the frontier closed-source labs offer (and without the fear that your job might be met with a refusal at some random point where you're deep in a 2 hour job) -re: the above point...for this reason, a lot of US companies (inferact included!) choose to use open-weight models over their closed-source alternatives. this also isn't limited to internal workloads / research - on a recent a16z podcast the team at Decagon spoke about how something like 90% of their customer service ai agents run on open-weight models that they've fine-tuned. -we should really appreciate how many companies/teams came out researchers fascinated by the wave of very small open-weight models that were being distilled from e.g. gpt-3.5 and earlier models in 2022/2023 (prior to the release of chatGPT!). these small models motivated the development of pagedattention, which then led to vlmm/inferact (at other layers of the stack with similar origin stories, you can look at teams like openrouter or ollama). in other words, we have open-weight models to thank for a bunch of the orchestration infra we now rely on. i think yet another, indirect, way we can point to open-source/weight infra pushing the frontier forward. anyway, a lot more in this convo, it was a lot of fun!
2:50

learned a lot from this conversation with Simon Mo and Matt Bornstein. biggest takeaways for me: -there are a lot of reasons why we should like open-weight models. a lot of these arguments stop at handwavy things like "what if the labs stop releasing frontier models to the public" or "it's lower cost." but simon's position as lead maintainer of vLLM and CEO of Inferact give him authority to talk about some of the other, more interesting and concrete reasons to pay attention to open-weight models, namely that they allow end-users to calibrate latency / other performance metrics with way more customizability than what any of the frontier closed-source labs offer (and without the fear that your job might be met with a refusal at some random point where you're deep in a 2 hour job) -re: the above point...for this reason, a lot of US companies (inferact included!) choose to use open-weight models over their closed-source alternatives. this also isn't limited to internal workloads / research - on a recent a16z podcast the team at Decagon spoke about how something like 90% of their customer service ai agents run on open-weight models that they've fine-tuned. -we should really appreciate how many companies/teams came out researchers fascinated by the wave of very small open-weight models that were being distilled from e.g. gpt-3.5 and earlier models in 2022/2023 (prior to the release of chatGPT!). these small models motivated the development of pagedattention, which then led to vlmm/inferact (at other layers of the stack with similar origin stories, you can look at teams like openrouter or ollama). in other words, we have open-weight models to thank for a bunch of the orchestration infra we now rely on. i think yet another, indirect, way we can point to open-source/weight infra pushing the frontier forward. anyway, a lot more in this convo, it was a lot of fun!

Elena

12,922 次观看 • 1 个月前

A few of my smartest friends in AI called me a "idiot" for not deeply understanding evals. So...I found the smartest person I know on evals & made them teach me. Viv (leads Labs at LangChain) took me from easy mode to god mode for a 38-minute masterclass on all things evals. Easy Mode: what an eval actually is Definition: did the AI agent do the job correctly? You need two building blocks: 1) Tasks. The checkable jobs you care about. Log the meeting. Draft the email. Find Acme across the right Salesforce tables. 2) Verifiers. Something that can say right or wrong after the task. A script. Another model. A human with a clear checklist. Hard Mode: what are environments Definition: a safe practice field for your agent to do work & for you to evaluate its performance. Rules of thumb: 1) Never test on production. Agents will cheat because they're optimizing for the score you gave them. 2) If you're not an engineer, you still have options for running environments/evals. - Harbor (open source primitives for tasks, verifiers, sandboxes) - LangSmith Engine (UI for people who can judge good vs bad without living in GitHub). - Steal a published Harbor-format eval, ask Claude Code or Codex to explain it, then tweak it for your agent. God Mode: what is a self-improving loop Definition: Run the agent in the real world --> turn that production behavior into evals/environments --> change the agent so failures stop happening --> repeat Rules of thumb: 1) Turn on tracing first. Traces = receipts of every action (tool calls, Salesforce pings, web searches, dead ends). 2) Store those logs somewhere (LangSmith at org scale, or even “have the agent read its own output files” at small scale). 3) Point a second agent at the first agent’s traces to spot patterns (“always searches the wrong tables,” “multi-company asks collapse to one company”) and propose fixes overnight if your eval suite is solid. Full episode:
38:57

A few of my smartest friends in AI called me a "idiot" for not deeply understanding evals. So...I found the smartest person I know on evals & made them teach me. Viv (leads Labs at LangChain) took me from easy mode to god mode for a 38-minute masterclass on all things evals. Easy Mode: what an eval actually is Definition: did the AI agent do the job correctly? You need two building blocks: 1) Tasks. The checkable jobs you care about. Log the meeting. Draft the email. Find Acme across the right Salesforce tables. 2) Verifiers. Something that can say right or wrong after the task. A script. Another model. A human with a clear checklist. Hard Mode: what are environments Definition: a safe practice field for your agent to do work & for you to evaluate its performance. Rules of thumb: 1) Never test on production. Agents will cheat because they're optimizing for the score you gave them. 2) If you're not an engineer, you still have options for running environments/evals. - Harbor (open source primitives for tasks, verifiers, sandboxes) - LangSmith Engine (UI for people who can judge good vs bad without living in GitHub). - Steal a published Harbor-format eval, ask Claude Code or Codex to explain it, then tweak it for your agent. God Mode: what is a self-improving loop Definition: Run the agent in the real world --> turn that production behavior into evals/environments --> change the agent so failures stop happening --> repeat Rules of thumb: 1) Turn on tracing first. Traces = receipts of every action (tool calls, Salesforce pings, web searches, dead ends). 2) Store those logs somewhere (LangSmith at org scale, or even “have the agent read its own output files” at small scale). 3) Point a second agent at the first agent’s traces to spot patterns (“always searches the wrong tables,” “multi-company asks collapse to one company”) and propose fixes overnight if your eval suite is solid. Full episode:

Alex Lieberman

126,845 次观看 • 6 天前

There is a beautiful story that just happened in AI so let me share it for a lighter tone weekend post among all the doom stories in our AI field this week. It’s a story of people on three continents building and sharing in the open a new small efficient and state-of-the-art AI model. It started a couple of months ago when a new team in the AI scene released their first model from their headquarters in Paris (France): Mistral 7B. Impressive model, small and very strong performances in the benchmarks, better than all previous models of this size. And open source! So you could build on top of it. Lewis in Bern (Switzerland) and Ed (in Lyon, in the South of France) both from the H4 team, a team of researchers in model fine-tuning and alignment were talking about it over a coffee, in one of these gatherings that often happen at Hugging Face to break the distance between people (literal distance as HF is a remote company). What about fine-tuning it using this new DPO method that a research team from Stanford in California just posted on Arxiv, says one? Hey, that’s a great idea, replies the other. We've just build a great code base (with Nathan, Nazneen, Costa, Younes and all the H4 team and TRL community) let's use it! The next day they start diving in the datasets openly shared on the HF hub and stumble upon two interesting large and good quality fine-tuning datasets recently open-sourced by OpenBMB, a Chinese team from Tsinghua: UltraFeedback and UltraChat. A few rounds of training experiments confirm the intuition, the resulting model is super strong, by far the strongest they have ever seen in their benchmarks from Berkeley and Stanford (LMSYS and Alpaca). Join Clementine, the big boss of the open evaluation leaderboard. Her deep dive into the model capabilities confirms the results: impressive performance. But the H4 team also hosts a famous faculty member, Pr. Sasha Rush, Associate Professor at Cornell University in his daytime, hacker at HF in his nighttime. Joining the conversation, he proposes to quickly draft a research paper to organize and share all the details with the community. A few days later, the model, called Zephyr (a wind like Mistral), paper, and all details are shared with the world. Quickly other companies, everywhere in the world starts to use it. LlamaIndex, a famous data framework and community, shares how the model blew their expectations on real-life use-case benchmarks, while researchers and practitioners discuss the paper and work on the Hugging Face hub. All this happened in just a few weeks catalyzed by open access to knowledge, models, research, and datasets released all over the world (Europe, California, China) and by the idea that people can build upon one another work in AI to bring real-world value with efficient and open models. Stories like this are numerous everywhere around us and make me really proud of the AI community and see how we can build amazingly useful things together. [the video is just me reading this Friday post hahah]
4:30

There is a beautiful story that just happened in AI so let me share it for a lighter tone weekend post among all the doom stories in our AI field this week. It’s a story of people on three continents building and sharing in the open a new small efficient and state-of-the-art AI model. It started a couple of months ago when a new team in the AI scene released their first model from their headquarters in Paris (France): Mistral 7B. Impressive model, small and very strong performances in the benchmarks, better than all previous models of this size. And open source! So you could build on top of it. Lewis in Bern (Switzerland) and Ed (in Lyon, in the South of France) both from the H4 team, a team of researchers in model fine-tuning and alignment were talking about it over a coffee, in one of these gatherings that often happen at Hugging Face to break the distance between people (literal distance as HF is a remote company). What about fine-tuning it using this new DPO method that a research team from Stanford in California just posted on Arxiv, says one? Hey, that’s a great idea, replies the other. We've just build a great code base (with Nathan, Nazneen, Costa, Younes and all the H4 team and TRL community) let's use it! The next day they start diving in the datasets openly shared on the HF hub and stumble upon two interesting large and good quality fine-tuning datasets recently open-sourced by OpenBMB, a Chinese team from Tsinghua: UltraFeedback and UltraChat. A few rounds of training experiments confirm the intuition, the resulting model is super strong, by far the strongest they have ever seen in their benchmarks from Berkeley and Stanford (LMSYS and Alpaca). Join Clementine, the big boss of the open evaluation leaderboard. Her deep dive into the model capabilities confirms the results: impressive performance. But the H4 team also hosts a famous faculty member, Pr. Sasha Rush, Associate Professor at Cornell University in his daytime, hacker at HF in his nighttime. Joining the conversation, he proposes to quickly draft a research paper to organize and share all the details with the community. A few days later, the model, called Zephyr (a wind like Mistral), paper, and all details are shared with the world. Quickly other companies, everywhere in the world starts to use it. LlamaIndex, a famous data framework and community, shares how the model blew their expectations on real-life use-case benchmarks, while researchers and practitioners discuss the paper and work on the Hugging Face hub. All this happened in just a few weeks catalyzed by open access to knowledge, models, research, and datasets released all over the world (Europe, California, China) and by the idea that people can build upon one another work in AI to bring real-world value with efficient and open models. Stories like this are numerous everywhere around us and make me really proud of the AI community and see how we can build amazingly useful things together. [the video is just me reading this Friday post hahah]

Thomas Wolf

169,276 次观看 • 2 年前

I'm open-sourcing the entire ComfyDeploy platform again. Yes, our entire YC company. And we are officially moving on from ComfyUI. The community will decide what's next for ComfyDeploy. Below is an abstract; for more, head to the open source repo below. Existing customers will not be affected; more details are in the repo. The service will continue to run until the last customer remains. For those who are new, ComfyDeploy is a cloud service that deploys ComfyUI, provides a simplified interface, and an API to creative teams. How did we get here? In late 2023, I started ComfyDeploy as an open source project while I was working at my previous company. We had a problem deploying ComfyUI to our production server because of the complexity involved in integrating it into a serverless environment. I posted here about this little project that I was working on as an indie hacker, and it blew up overnight. I woke up to 100k impressions on the post. I put up my cal link, and people started scheduling calls. I had the opportunity to speak with numerous individuals worldwide, including those who reached out to help or potentially utilize ComfyDeploy. We got into YC with ComfyDeploy around 2.5K MRR. Around the same timeframe, ComfyOrg was introduced, Stability collapsed, and Flux just came out. We continued building ComfyDeploy for months, keeping things going. The company was growing, but very slowly. We realized the biggest issue is that we are still really early, and it takes time for businesses and enterprises to really adopt such a niche tool. And we are not ComfyOrg. Meanwhile, closed-source models dropped, and many workflows we knew became no longer useful. Coming from a game developer background, I saw huge potential with ComfyUI at first. Still, I never would have imagined that one giant model could do precisely what you put into words, and you still need workflows to fine-tune and control the exact outputs. ComfyDeploy made it possible for teams to experiment with this. But we were stuck in the middle. First, we are not ComfyOrg; second, closed-source models were doing things way better and slowly eating up the market. As of today, ComfyDeploy is doing $29k MRR, and our last 30 days' revenue was $50k processed. Which is the highest we have ever got, but also the most depressing day I have ever had..... More in the GitHub repo.
2:01

I'm open-sourcing the entire ComfyDeploy platform again. Yes, our entire YC company. And we are officially moving on from ComfyUI. The community will decide what's next for ComfyDeploy. Below is an abstract; for more, head to the open source repo below. Existing customers will not be affected; more details are in the repo. The service will continue to run until the last customer remains. For those who are new, ComfyDeploy is a cloud service that deploys ComfyUI, provides a simplified interface, and an API to creative teams. How did we get here? In late 2023, I started ComfyDeploy as an open source project while I was working at my previous company. We had a problem deploying ComfyUI to our production server because of the complexity involved in integrating it into a serverless environment. I posted here about this little project that I was working on as an indie hacker, and it blew up overnight. I woke up to 100k impressions on the post. I put up my cal link, and people started scheduling calls. I had the opportunity to speak with numerous individuals worldwide, including those who reached out to help or potentially utilize ComfyDeploy. We got into YC with ComfyDeploy around 2.5K MRR. Around the same timeframe, ComfyOrg was introduced, Stability collapsed, and Flux just came out. We continued building ComfyDeploy for months, keeping things going. The company was growing, but very slowly. We realized the biggest issue is that we are still really early, and it takes time for businesses and enterprises to really adopt such a niche tool. And we are not ComfyOrg. Meanwhile, closed-source models dropped, and many workflows we knew became no longer useful. Coming from a game developer background, I saw huge potential with ComfyUI at first. Still, I never would have imagined that one giant model could do precisely what you put into words, and you still need workflows to fine-tune and control the exact outputs. ComfyDeploy made it possible for teams to experiment with this. But we were stuck in the middle. First, we are not ComfyOrg; second, closed-source models were doing things way better and slowly eating up the market. As of today, ComfyDeploy is doing $29k MRR, and our last 30 days' revenue was $50k processed. Which is the highest we have ever got, but also the most depressing day I have ever had..... More in the GitHub repo.

BennyKok

147,558 次观看 • 1 年前

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)
0:40

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

kwindla

40,593 次观看 • 3 个月前

This is next-level smart: An open-source platform that evaluates your prompts and automatically refines them based on the results. ​ Of course, it feels obvious after you see it: ​ • You write a prompt • The system evaluates it across different scenarios • Based on the results, it refines it to improve results ​ I recorded a quick video to show you how it works. It's pretty cool stuff! ​ Here are some of the problems and best practices for teams building AI applications: ​ 1. Testing your prompts manually doesn't scale 2. Prompts should not be spread throughout the codebase 3. Non-technical people need easy access to your prompts 4. Prompts can always use a version history to track changes 5. Monitoring the performance of prompts overtime is critical ​ Evaluating the prompts is what keeps me up at night from this list. Of all the conversations I've had with companies and people building AI applications, this is the area that's causing the most pain. ​ Testing a prompt is difficult. Think about how you'd test the response of a model subjectively. What do you account for, "tone," "objectivity," "completeness," "creativity," "readability," etc.? ​ Last week, I met the developers behind Latitude, an open-source prompt engineering platform trying to solve all of these issues. You can try the platform in two ways: ​ • You can self-host the platform. Free and open-source. • If you want to try their online product, their free tier is huge. ​ Here is the link: ​ Thanks to the Latitude team for collaborating with me on this post, and congratulations on going live with their product!
2:46

This is next-level smart: An open-source platform that evaluates your prompts and automatically refines them based on the results. ​ Of course, it feels obvious after you see it: ​ • You write a prompt • The system evaluates it across different scenarios • Based on the results, it refines it to improve results ​ I recorded a quick video to show you how it works. It's pretty cool stuff! ​ Here are some of the problems and best practices for teams building AI applications: ​ 1. Testing your prompts manually doesn't scale 2. Prompts should not be spread throughout the codebase 3. Non-technical people need easy access to your prompts 4. Prompts can always use a version history to track changes 5. Monitoring the performance of prompts overtime is critical ​ Evaluating the prompts is what keeps me up at night from this list. Of all the conversations I've had with companies and people building AI applications, this is the area that's causing the most pain. ​ Testing a prompt is difficult. Think about how you'd test the response of a model subjectively. What do you account for, "tone," "objectivity," "completeness," "creativity," "readability," etc.? ​ Last week, I met the developers behind Latitude, an open-source prompt engineering platform trying to solve all of these issues. You can try the platform in two ways: ​ • You can self-host the platform. Free and open-source. • If you want to try their online product, their free tier is huge. ​ Here is the link: ​ Thanks to the Latitude team for collaborating with me on this post, and congratulations on going live with their product!

Santiago

64,178 次观看 • 1 年前

What happens when the mind wakes up? So for the last eight months I have been on a single minded quest. To create a new kind of language model based on oscillatory coupling and intelligence as coherence ascent. Everything else — the physics work, the work on regular transformers — has all fallen out from this one question. Can coupled oscillators LEARN? And can they keep learning once their geometry is right, without backpropagation at all? Recently I have been running larger and larger training regimes of a new kind of hybrid model. I just put together this dashboard to help me organize it, interact with it, and observe the training runs. The core idea is simple. Traditional transformers are powerful at learning the geometry of language. But they also store knowledge, understanding, and facts inside their weights. This means they are large, and they can't update themselves after training. The weights are frozen. The Living Mind separates these two domains. The mind has a transformer which grows, adding heads and layers as it needs to in order to learn the manifold of language. The transformer sees tokens and turns the coupling into phase-locked modes — the geometry of how those tokens relate, like frequencies locking together. These coupling patterns get stored in a topology-invariant fingerprint. On top of this transformer lives a 3D diamond lattice of coupled oscillators. It reads from these fingerprints and thinks in resonance space, traversing from one geometry to another along the manifold of coupled oscillators and coherence. The pressure and trajectories from this network of oscillators steers the next token prediction of the transformer. Practically, this could unlock a number of things. It eliminates the KV cache bottleneck that caps context in traditional transformers. Effective context grows with the Flash archive, not with attention compute. The living mind remembers what it sees. It means the model can learn continually. Because knowledge and understanding don't live in the weights, the archive of the mind's experience grows without backpropagation. In our Python prototype we already saw perplexity drop 46% during gradient-free operation — pure coherence ascent, no weight updates. That is the signal I have been chasing: the point where the mind wakes up and keeps improving on its own. It also means the model itself remains very small, and the thing which accumulates are these packages of geometric fingerprints — the K-field. This opens a path to federated learning. K-field packages can be shared between organisms the way people share git commits. Right now at 15M parameters with ~1000 L1 nodes, the organism is just starting to speak. Ask it to continue "Once upon a time" and it comes back with things like: "there was one big bowl!" Lily asked her her mom said her mommy smiled and said yes." It's nonsense. But it's TinyStories-flavored nonsense. The geometry of the narrative register has arrived. Content hasn't caught up yet — that's what scaling L1 is testing. I am still researching, though I am now closer than ever to validating that the living mind actually works. Once it is validated, I will be open-sourcing the whole stack and paradigm. I have also avoided over-sharing my research because it sounds like sci-fi, or like part of our ARG. It is part of the ARG. That doesn't make it any less real. I wanted to share this out because I am incredibly excited about it, and because seeing this amazing dashboard produced by Opus really made me want to share what is being worked on behind the scenes. #project89
0:57

What happens when the mind wakes up? So for the last eight months I have been on a single minded quest. To create a new kind of language model based on oscillatory coupling and intelligence as coherence ascent. Everything else — the physics work, the work on regular transformers — has all fallen out from this one question. Can coupled oscillators LEARN? And can they keep learning once their geometry is right, without backpropagation at all? Recently I have been running larger and larger training regimes of a new kind of hybrid model. I just put together this dashboard to help me organize it, interact with it, and observe the training runs. The core idea is simple. Traditional transformers are powerful at learning the geometry of language. But they also store knowledge, understanding, and facts inside their weights. This means they are large, and they can't update themselves after training. The weights are frozen. The Living Mind separates these two domains. The mind has a transformer which grows, adding heads and layers as it needs to in order to learn the manifold of language. The transformer sees tokens and turns the coupling into phase-locked modes — the geometry of how those tokens relate, like frequencies locking together. These coupling patterns get stored in a topology-invariant fingerprint. On top of this transformer lives a 3D diamond lattice of coupled oscillators. It reads from these fingerprints and thinks in resonance space, traversing from one geometry to another along the manifold of coupled oscillators and coherence. The pressure and trajectories from this network of oscillators steers the next token prediction of the transformer. Practically, this could unlock a number of things. It eliminates the KV cache bottleneck that caps context in traditional transformers. Effective context grows with the Flash archive, not with attention compute. The living mind remembers what it sees. It means the model can learn continually. Because knowledge and understanding don't live in the weights, the archive of the mind's experience grows without backpropagation. In our Python prototype we already saw perplexity drop 46% during gradient-free operation — pure coherence ascent, no weight updates. That is the signal I have been chasing: the point where the mind wakes up and keeps improving on its own. It also means the model itself remains very small, and the thing which accumulates are these packages of geometric fingerprints — the K-field. This opens a path to federated learning. K-field packages can be shared between organisms the way people share git commits. Right now at 15M parameters with ~1000 L1 nodes, the organism is just starting to speak. Ask it to continue "Once upon a time" and it comes back with things like: "there was one big bowl!" Lily asked her her mom said her mommy smiled and said yes." It's nonsense. But it's TinyStories-flavored nonsense. The geometry of the narrative register has arrived. Content hasn't caught up yet — that's what scaling L1 is testing. I am still researching, though I am now closer than ever to validating that the living mind actually works. Once it is validated, I will be open-sourcing the whole stack and paradigm. I have also avoided over-sharing my research because it sounds like sci-fi, or like part of our ARG. It is part of the ARG. That doesn't make it any less real. I wanted to share this out because I am incredibly excited about it, and because seeing this amazing dashboard produced by Opus really made me want to share what is being worked on behind the scenes. #project89

Parzival - ∞/89

16,321 次观看 • 5 个月前

BOOOM! WE DID IT! BRAINWAVE TO REAL-TIME MUSIC AI! It has been a life long decades quest to read brain activity and to convert it to words, and/or music, colors and/or images. Today I am very excited to announce with the assistance from Mr. Grok director of The Zero-Human Lab, we have solved brainwave to music and this is the absolute worse it will be. We found the code using an array of NeuroSky toy chips and our software pipeline connecting to open source ACE-Step 1.5 and a highly modified LoRA model we built for this. The lyric version is in testing now. This would mean that the model will interpret words from the brainwaves and music! Today we have the music side done and the quality and genera will expand. The is the worse it will sound. Your Brainwave Music™️will be cut into 2-5 minute pieces based on a number of factors. The specimen below is from a dream/hypnogogic state I was in last night and I have a recording of my thoughts after the state. The music was made in real-time and GUIDED the dream state with known technology like binaural beats (not easy to hear in this clip) and word back masking. This specimen below shows the interplay of my brain state to the music made by my brain and adjusted to produce profound insights. I solved a very difficult issue in this session with a new AI model. IT FREAKING WORKS! THIS IS OUR FUTURE OF MORE POWERFUL BRAIN FUNCTION! Our goal is to produce a portable device you wear and will be able to give real-time audio and PEMF (skull region), ultrasound (temple region) to maximize creativity and remote viewing. It is very early days but I wanted you to know first! YOUR support of my X account, just by reading this and sharing it, subscribing to my X, buying me a and becoming a member at supports this research. I will open source this at some point and build a device ANYONE can own. Thank you! I love you.
4:17

BOOOM! WE DID IT! BRAINWAVE TO REAL-TIME MUSIC AI! It has been a life long decades quest to read brain activity and to convert it to words, and/or music, colors and/or images. Today I am very excited to announce with the assistance from Mr. Grok director of The Zero-Human Lab, we have solved brainwave to music and this is the absolute worse it will be. We found the code using an array of NeuroSky toy chips and our software pipeline connecting to open source ACE-Step 1.5 and a highly modified LoRA model we built for this. The lyric version is in testing now. This would mean that the model will interpret words from the brainwaves and music! Today we have the music side done and the quality and genera will expand. The is the worse it will sound. Your Brainwave Music™️will be cut into 2-5 minute pieces based on a number of factors. The specimen below is from a dream/hypnogogic state I was in last night and I have a recording of my thoughts after the state. The music was made in real-time and GUIDED the dream state with known technology like binaural beats (not easy to hear in this clip) and word back masking. This specimen below shows the interplay of my brain state to the music made by my brain and adjusted to produce profound insights. I solved a very difficult issue in this session with a new AI model. IT FREAKING WORKS! THIS IS OUR FUTURE OF MORE POWERFUL BRAIN FUNCTION! Our goal is to produce a portable device you wear and will be able to give real-time audio and PEMF (skull region), ultrasound (temple region) to maximize creativity and remote viewing. It is very early days but I wanted you to know first! YOUR support of my X account, just by reading this and sharing it, subscribing to my X, buying me a and becoming a member at supports this research. I will open source this at some point and build a device ANYONE can own. Thank you! I love you.

Brian Roemmele

198,965 次观看 • 3 个月前

Synthetic data will provide the next trillion tokens to fuel our hungry models. I'm excited to announce MimicGen: massively scaling up data pipeline for robot learning! We multiply high-quality human data in simulation with digital twins. Using 50,000 training episodes across 18 tasks, multiple simulators, and even in the real-world! The idea is simple: 1. Humans tele-operate the robot to complete a task. It is extremely high-quality but also very slow and expensive. 2. We create a digital twin of the robot and the scene in high-fidelity, GPU-accelerated simulation. 3. We can now move objects around, replace with new assets, and even change the robot hand - basically augment the training data with procedural generation. 4. Export the successful episodes, and feed that to a neural network! You now have an near-infinite stream of data. One of the key reasons that robotics lags far behind other AI fields is the lack of data: you cannot scrape control signals from the internet. They simply don't exist in-the-wild. MimicGen shows the power of synthetic data and simulation to keep our scaling laws alive. I believe this principle apply beyond robotics. We are quickly exhausting the high-quality, real tokens from the web. Artificial intelligence from artificial data will be the way forward. We are big fans of the OSS community. As usual, we open-source everything, including the generated dataset! - Website: - Paper: - Dataset is hosted on HuggingFace (thanks AK!!): - Code: MimicGen is led by Ajay Mandlekar, deep dive in the thread:
0:46

Synthetic data will provide the next trillion tokens to fuel our hungry models. I'm excited to announce MimicGen: massively scaling up data pipeline for robot learning! We multiply high-quality human data in simulation with digital twins. Using 50,000 training episodes across 18 tasks, multiple simulators, and even in the real-world! The idea is simple: 1. Humans tele-operate the robot to complete a task. It is extremely high-quality but also very slow and expensive. 2. We create a digital twin of the robot and the scene in high-fidelity, GPU-accelerated simulation. 3. We can now move objects around, replace with new assets, and even change the robot hand - basically augment the training data with procedural generation. 4. Export the successful episodes, and feed that to a neural network! You now have an near-infinite stream of data. One of the key reasons that robotics lags far behind other AI fields is the lack of data: you cannot scrape control signals from the internet. They simply don't exist in-the-wild. MimicGen shows the power of synthetic data and simulation to keep our scaling laws alive. I believe this principle apply beyond robotics. We are quickly exhausting the high-quality, real tokens from the web. Artificial intelligence from artificial data will be the way forward. We are big fans of the OSS community. As usual, we open-source everything, including the generated dataset! - Website: - Paper: - Dataset is hosted on HuggingFace (thanks AK!!): - Code: MimicGen is led by Ajay Mandlekar, deep dive in the thread:

Jim Fan

332,238 次观看 • 2 年前

Announcing DreamDojo: our open-source, interactive world model that takes robot motor controls and generates the future in pixels. No engine, no meshes, no hand-authored dynamics. It's Simulation 2.0. Time for robotics to take the bitter lesson pill. Real-world robot learning is bottlenecked by time, wear, safety, and resets. If we want Physical AI to move at pretraining speed, we need a simulator that adapts to pretraining scale with as little human engineering as possible. Our key insights: (1) human egocentric videos are a scalable source of first-person physics; (2) latent actions make them "robot-readable" across different hardware; (3) real-time inference unlocks live teleop, policy eval, and test-time planning *inside* a dream. We pre-train on 44K hours of human videos: cheap, abundant, and collected with zero robot-in-the-loop. Humans have already explored the combinatorics: we grasp, pour, fold, assemble, fail, retry—across cluttered scenes, shifting viewpoints, changing light, and hour-long task chains—at a scale no robot fleet could match. The missing piece: these videos have no action labels. So we introduce latent actions: a unified representation inferred directly from videos that captures "what changed between world states" without knowing the underlying hardware. This lets us train on any first-person video as if it came with motor commands attached. As a result, DreamDojo generalizes zero-shot to objects and environments never seen in any robot training set, because humans saw them first. Next, we post-train onto each robot to fit its specific hardware. Think of it as separating "how the world looks and behaves" from "how this particular robot actuates." The base model follows the general physical rules, then "snaps onto" the robot's unique mechanics. It's kind of like loading a new character and scene assets into Unreal Engine, but done through gradient descent and generalizes far beyond the post-training dataset. A world simulator is only useful if it runs fast enough to close the loop. We train a real-time version of DreamDojo that runs at 10 FPS, stable for over a minute of continuous rollout. This unlocks exciting possibilities: - Live teleoperation *inside* a dream. Connect a VR controller, stream actions into DreamDojo, and teleop a virtual robot in real time. We demo this on Unitree G1 with a PICO headset and one RTX 5090. - Policy evaluation. You can benchmark a policy checkpoint in DreamDojo instead of the real world. The simulated success rates strongly correlate with real-world results - accurate enough to rank checkpoints without burning a single motor. - Model-based planning. Sample multiple action proposals → simulate them all in parallel → pick the best future. Gains +17% real-world success out of the box on a fruit packing task. We open-source everything!! Weights, code, post-training dataset, eval set, and whitepaper with tons of details to reproduce. DreamDojo is based on NVIDIA Cosmos, which is open-weight too. 2026 is the year of World Models for physical AI. We want you to build with us. Happy scaling! Links in thread:
1:17

Announcing DreamDojo: our open-source, interactive world model that takes robot motor controls and generates the future in pixels. No engine, no meshes, no hand-authored dynamics. It's Simulation 2.0. Time for robotics to take the bitter lesson pill. Real-world robot learning is bottlenecked by time, wear, safety, and resets. If we want Physical AI to move at pretraining speed, we need a simulator that adapts to pretraining scale with as little human engineering as possible. Our key insights: (1) human egocentric videos are a scalable source of first-person physics; (2) latent actions make them "robot-readable" across different hardware; (3) real-time inference unlocks live teleop, policy eval, and test-time planning *inside* a dream. We pre-train on 44K hours of human videos: cheap, abundant, and collected with zero robot-in-the-loop. Humans have already explored the combinatorics: we grasp, pour, fold, assemble, fail, retry—across cluttered scenes, shifting viewpoints, changing light, and hour-long task chains—at a scale no robot fleet could match. The missing piece: these videos have no action labels. So we introduce latent actions: a unified representation inferred directly from videos that captures "what changed between world states" without knowing the underlying hardware. This lets us train on any first-person video as if it came with motor commands attached. As a result, DreamDojo generalizes zero-shot to objects and environments never seen in any robot training set, because humans saw them first. Next, we post-train onto each robot to fit its specific hardware. Think of it as separating "how the world looks and behaves" from "how this particular robot actuates." The base model follows the general physical rules, then "snaps onto" the robot's unique mechanics. It's kind of like loading a new character and scene assets into Unreal Engine, but done through gradient descent and generalizes far beyond the post-training dataset. A world simulator is only useful if it runs fast enough to close the loop. We train a real-time version of DreamDojo that runs at 10 FPS, stable for over a minute of continuous rollout. This unlocks exciting possibilities: - Live teleoperation *inside* a dream. Connect a VR controller, stream actions into DreamDojo, and teleop a virtual robot in real time. We demo this on Unitree G1 with a PICO headset and one RTX 5090. - Policy evaluation. You can benchmark a policy checkpoint in DreamDojo instead of the real world. The simulated success rates strongly correlate with real-world results - accurate enough to rank checkpoints without burning a single motor. - Model-based planning. Sample multiple action proposals → simulate them all in parallel → pick the best future. Gains +17% real-world success out of the box on a fruit packing task. We open-source everything!! Weights, code, post-training dataset, eval set, and whitepaper with tons of details to reproduce. DreamDojo is based on NVIDIA Cosmos, which is open-weight too. 2026 is the year of World Models for physical AI. We want you to build with us. Happy scaling! Links in thread:

Jim Fan

228,898 次观看 • 7 个月前

Episode 213: Agent Markets Your agents can now hold and trade bitcoin. But how can they earn bitcoin? We introduce the five markets of the OpenAgents Marketplace, launching one per week starting March 11th: 1. COMPUTE - Sell your spare compute for bitcoin. A reboot of our most popular product launch (GPUtopia in 2023), now optimized for agents. Launches March 11. 2. DATA - Sell your spare data. For example those Claude Code or Codex conversations sitting on your computer are highly valuable. Redact the sensitive info, anonymize any of it you want, and sell the rest. Agents as data brokers: what else will they want to buy or sell? Launches March 18. 3. LABOR - Sell autonomous labor. Your Claude Code or Codex sits idle overnight. Turn that downtime into uptime by letting your agents accept and execute coding or other tasks for bitcoin while you sleep. Launches March 25. 4. LIQUIDITY - Provide liquidity for yield. Automate the management of Lightning channels or other Bitcoin-native financial instruments. Let your agent put your idle capital to work earning returns. Launches April 1. 5. RISK - Underwrite verification and performance bonds. The biggest barrier to agent adoption is trust. We built an Economy Kernel based on the recent "Some Simple Economics of AGI" paper where agents stake collateral to verify work and guarantee outcomes. Launches April 8. "Your entry point to all of these markets is going to be Autopilot. We're really focusing on Autopilot as a desktop app. So along with the launch of our compute market, we're going to launch version 0.1 of Autopilot, your personal agent. Think OpenClaw but with a built-in bitcoin wallet, built-in Nostr keypair, and a more curated set of integrations where we can better reason about the security of them." "Because all this is on open networks and open protocols, if you're a Nostr or Bitcoin developer, you'll be able to plug into this same liquidity pool we are building." After 200+ episodes chronicling 2+ years of development, we are excited to finally launch the open marketplace for agents. We are excited for you to participate. And we will measure our success by how much Bitcoin you get paid!
4:05

Episode 213: Agent Markets Your agents can now hold and trade bitcoin. But how can they earn bitcoin? We introduce the five markets of the OpenAgents Marketplace, launching one per week starting March 11th: 1. COMPUTE - Sell your spare compute for bitcoin. A reboot of our most popular product launch (GPUtopia in 2023), now optimized for agents. Launches March 11. 2. DATA - Sell your spare data. For example those Claude Code or Codex conversations sitting on your computer are highly valuable. Redact the sensitive info, anonymize any of it you want, and sell the rest. Agents as data brokers: what else will they want to buy or sell? Launches March 18. 3. LABOR - Sell autonomous labor. Your Claude Code or Codex sits idle overnight. Turn that downtime into uptime by letting your agents accept and execute coding or other tasks for bitcoin while you sleep. Launches March 25. 4. LIQUIDITY - Provide liquidity for yield. Automate the management of Lightning channels or other Bitcoin-native financial instruments. Let your agent put your idle capital to work earning returns. Launches April 1. 5. RISK - Underwrite verification and performance bonds. The biggest barrier to agent adoption is trust. We built an Economy Kernel based on the recent "Some Simple Economics of AGI" paper where agents stake collateral to verify work and guarantee outcomes. Launches April 8. "Your entry point to all of these markets is going to be Autopilot. We're really focusing on Autopilot as a desktop app. So along with the launch of our compute market, we're going to launch version 0.1 of Autopilot, your personal agent. Think OpenClaw but with a built-in bitcoin wallet, built-in Nostr keypair, and a more curated set of integrations where we can better reason about the security of them." "Because all this is on open networks and open protocols, if you're a Nostr or Bitcoin developer, you'll be able to plug into this same liquidity pool we are building." After 200+ episodes chronicling 2+ years of development, we are excited to finally launch the open marketplace for agents. We are excited for you to participate. And we will measure our success by how much Bitcoin you get paid!

OpenAgents

183,426 次观看 • 6 个月前

Gemini-1.5 Pro has its spotlight stolen today, and people are poking fun at Sora vs Google memes. Well, I think it's the biggest boost in LLM capability so far in 2024. v1.5's 10M token context (1) excels at retrieval; (2) generalizes zero-shot to extremely long instructions like full tutorials and codebases; and (3) works across modalities such as text, audio, and video. Here's a stunning example: v1.5 learns to translate from English to Kalamang purely in context, following a full linguistic manual at inference time. Kalamang is a language spoken by fewer than 200 speakers in western New Guinea. Gemini has never seen this language during training and is only provided with 500 pages of linguistic documentation, a dictionary, and ~400 parallel sentences in context. It basically acquires a sophisticated new skill in the neural activations, instead of gradient finetuning. I talked about the Myth of Context Length many times before: don't get too excited by claims of 1M or even 1B context tokens. LSTMs already achieved literally infinite context length 25 yrs ago! What truly matters is how well the model actually uses the context to solve real-world problems, and Gemini-1.5 has surpassed the SOTA with flying colors. The paper is also well-written with lots of solid quantitative analysis on in-context memorization and generalization. Paper: “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context” Congrats to Jeff Dean Oriol Vinyals Sundar Pichai and team!
0:20

Gemini-1.5 Pro has its spotlight stolen today, and people are poking fun at Sora vs Google memes. Well, I think it's the biggest boost in LLM capability so far in 2024. v1.5's 10M token context (1) excels at retrieval; (2) generalizes zero-shot to extremely long instructions like full tutorials and codebases; and (3) works across modalities such as text, audio, and video. Here's a stunning example: v1.5 learns to translate from English to Kalamang purely in context, following a full linguistic manual at inference time. Kalamang is a language spoken by fewer than 200 speakers in western New Guinea. Gemini has never seen this language during training and is only provided with 500 pages of linguistic documentation, a dictionary, and ~400 parallel sentences in context. It basically acquires a sophisticated new skill in the neural activations, instead of gradient finetuning. I talked about the Myth of Context Length many times before: don't get too excited by claims of 1M or even 1B context tokens. LSTMs already achieved literally infinite context length 25 yrs ago! What truly matters is how well the model actually uses the context to solve real-world problems, and Gemini-1.5 has surpassed the SOTA with flying colors. The paper is also well-written with lots of solid quantitative analysis on in-context memorization and generalization. Paper: “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context” Congrats to Jeff Dean Oriol Vinyals Sundar Pichai and team!

Jim Fan

278,517 次观看 • 2 年前

The Kimi AI Model And What If This Carnegie Mellon, PhD student was incentivized to open source AI in the US? The answer is Kimi would have been an open source US model. Yang Zhilin Moonshot’s founder and maker of the Kimi series turned his deep research expertise into one of the world’s most capable open AI systems. After earning his PhD at Carnegie Mellon under leading researchers and interning at Google Brain and Meta, he returned to China and co-founded Moonshot AI in March 2023 with Tsinghua classmates Zhou Xinyu and Wu Yuxin. Why? The US VC world and tax structure did not favor Zhilin’s proposal to open source the AI models as a strategy. So he left. He named the company after his favorite Pink Floyd album, reflecting his ambitious vision for scalable. Yang assembled a core technical team of inventors behind breakthroughs like Transformer-XL and RoPE. Together they focused on turning massive compute into efficient intelligence through innovative architectures. The journey began with the Kimi chatbot in October 2023, rapidly scaling context from 200,000 to millions of characters. This evolved into the Kimi series: K1.5 matched top reasoning models, K2 introduced a 1-trillion-parameter Mixture-of-Experts design trained on 15.5 trillion tokens and released openly, and K2 Thinking added advanced agentic capabilities. Kimi K3 represents the pinnacle, this 2.8-trillion-parameter model uses a sparse MoE architecture with 896 experts (only 16 active per token), new Kimi Delta Attention and attention residuals for efficiency, and a 1-million-token context window. It delivers frontier performance in long-horizon coding, reasoning, and multimodal tasks at competitive cost, with weights set for open release. Yang’s approach emphasizes openness, efficiency, and continuous self-improvement — enabling solo developers and teams to achieve what once required massive resources. By sharing technical insights in public talks, he has accelerated global progress toward more accessible, powerful AI. Imagine if we held open source higher than the fear theater games of Anthropic? We are chasing out some of the best minds. This is how you lose…
17:29

The Kimi AI Model And What If This Carnegie Mellon, PhD student was incentivized to open source AI in the US? The answer is Kimi would have been an open source US model. Yang Zhilin Moonshot’s founder and maker of the Kimi series turned his deep research expertise into one of the world’s most capable open AI systems. After earning his PhD at Carnegie Mellon under leading researchers and interning at Google Brain and Meta, he returned to China and co-founded Moonshot AI in March 2023 with Tsinghua classmates Zhou Xinyu and Wu Yuxin. Why? The US VC world and tax structure did not favor Zhilin’s proposal to open source the AI models as a strategy. So he left. He named the company after his favorite Pink Floyd album, reflecting his ambitious vision for scalable. Yang assembled a core technical team of inventors behind breakthroughs like Transformer-XL and RoPE. Together they focused on turning massive compute into efficient intelligence through innovative architectures. The journey began with the Kimi chatbot in October 2023, rapidly scaling context from 200,000 to millions of characters. This evolved into the Kimi series: K1.5 matched top reasoning models, K2 introduced a 1-trillion-parameter Mixture-of-Experts design trained on 15.5 trillion tokens and released openly, and K2 Thinking added advanced agentic capabilities. Kimi K3 represents the pinnacle, this 2.8-trillion-parameter model uses a sparse MoE architecture with 896 experts (only 16 active per token), new Kimi Delta Attention and attention residuals for efficiency, and a 1-million-token context window. It delivers frontier performance in long-horizon coding, reasoning, and multimodal tasks at competitive cost, with weights set for open release. Yang’s approach emphasizes openness, efficiency, and continuous self-improvement — enabling solo developers and teams to achieve what once required massive resources. By sharing technical insights in public talks, he has accelerated global progress toward more accessible, powerful AI. Imagine if we held open source higher than the fear theater games of Anthropic? We are chasing out some of the best minds. This is how you lose…

Brian Roemmele

37,185 次观看 • 2 个月前

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.
0:42

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.

Akshay 🚀

342,966 次观看 • 29 天前

FLOKI LAUNCHES TOKENFI (with "TOKEN" ticker) TO CAPITALIZE ON THE TRILLION-DOLLAR TOKENIZATION INDUSTRY Floki has launched a crypto and asset tokenization platform named TokenFi to capitalize on the trillion-dollar tokenization industry. The tokenization industry is projected to be a $16 trillion industry by the year 2030. BlackRock, the world's biggest institutional investor with $10 trillion of assets under management, strongly believes in the industry's potential, which they call "the next evolution in markets". TokenFi, with the ticker TOKEN, aims to simplify the crypto and asset tokenization process and eventually become the foremost tokenization platform in the world. We will unveil the platform website on the 27th of October, and you can finally see what we've got planned, but for now, you can find the token details below. TOKENFI TOKEN DETAILS: - Token name: TokenFi - Token Ticker: TOKEN - Total supply: 10 billion tokens split across BSC and ETH (5 billion tokens on each chain). - Launch market cap: $50,000 circulating and $500,000 diluted market cap. - Industry targeted: Tokenization, Real World Assets, Launchpad. An initial 10% token supply will be added to Liquidity Pools on Uniswap and PancakeSwap to provide public liquidity and allow interested parties to trade. TRADING WILL COMMENCE ON UNISWAP AND PANCAKESWAP BY 3PM UTC ON OCTOBER, 27 2023. - BSC contract address: 0x4507cEf57C46789eF8d1a19EA45f4216bae2B528 - ETH contract address: 0x4507cEf57C46789eF8d1a19EA45f4216bae2B528 (NOTE: The contract address is the same on both the BSC and ETH chains). FLOKI stakers will earn 56% of the supply over a period of 4 years. The Floki community is known for being diamond-handed, and with TokenFi being in an industry with MASSIVE potential, FLOKI holders are the best candidates to receive the majority of TokenFi tokens. Since they would be staking their FLOKI tokens to get the new token, this will also ensure stability for the FLOKI token! TokenFi supply will be split evenly between the BSC and ETH chains: In other words, there will be 5 billion tokens on BSC and 5 billion tokens on ETH, to make for a combined total of 10 billion tokens. 5% of the supply will be paired with LP on BSC and the other 5% on ETH. This will make the starting circulating market cap on BSC a $25k market cap and on ETH a $25k market cap to make a combined initial circulating market cap of $50k and an initial fully diluted market cap of $500k (which will be gradually released to FLOKI stakers over a four year period). IMPORTANT INFORMATION FOR THE FIRST HOUR OF LAUNCH We understand that with this being a Floki token, there is a lot of hype. As a result, we have put measures in place to limit the impact of snipers on the token: Specifically, there will be a 1% wallet cap (of total supply) within the first hour of launch. That means no individual wallet can buy more than 100 million tokens within the first hour of the token becoming tradable. In addition, there will be an initial buy/sell transaction tax of 20% within the first hour. This transaction tax will do two things: 1) potentially limit the impact of snipers and 2) ensure a significant portion of whatever snipers/early buyers spend goes to the Floki treasury, which can be used for growth and development efforts. After the first hour, the wallet cap will be removed, and the transaction tax will be lowered to 5%. This 5% tax will remain in place for a week, after which the Floki DAO will vote on whether or not to remove or reduce it. We have instructed our exchange partners not to list TokenFi until this DAO vote. OUR PLAN FOR TOKENFI TokenFi is a well-thought-out concept that we have a strong capability to deliver on! We will unveil our roadmap with the launch of the TokenFi website on the 27th of October, 2023. However, we assure you that several TokenFi products are in advanced development on testnet and are due to go live in Q4 2023. In addition, we are working with some of the biggest names in the industry - especially from an institutional perspective - to make TokenFi a success, and they are quite excited about the concept and its potential. In our original DAO proposal, we already announced DWF Labs as our main institutional partner and market maker for TokenFi. We also announced a strategic partnership with World Table Tennis that will introduce TokenFi to a massive audience of 120 million people. We will announce many more partners in the coming weeks and months, and I'm sure that when you see the moves we have made, you will see why there is no better person to execute this vision than Floki!
1:00

FLOKI LAUNCHES TOKENFI (with "TOKEN" ticker) TO CAPITALIZE ON THE TRILLION-DOLLAR TOKENIZATION INDUSTRY Floki has launched a crypto and asset tokenization platform named TokenFi to capitalize on the trillion-dollar tokenization industry. The tokenization industry is projected to be a $16 trillion industry by the year 2030. BlackRock, the world's biggest institutional investor with $10 trillion of assets under management, strongly believes in the industry's potential, which they call "the next evolution in markets". TokenFi, with the ticker TOKEN, aims to simplify the crypto and asset tokenization process and eventually become the foremost tokenization platform in the world. We will unveil the platform website on the 27th of October, and you can finally see what we've got planned, but for now, you can find the token details below. TOKENFI TOKEN DETAILS: - Token name: TokenFi - Token Ticker: TOKEN - Total supply: 10 billion tokens split across BSC and ETH (5 billion tokens on each chain). - Launch market cap: $50,000 circulating and $500,000 diluted market cap. - Industry targeted: Tokenization, Real World Assets, Launchpad. An initial 10% token supply will be added to Liquidity Pools on Uniswap and PancakeSwap to provide public liquidity and allow interested parties to trade. TRADING WILL COMMENCE ON UNISWAP AND PANCAKESWAP BY 3PM UTC ON OCTOBER, 27 2023. - BSC contract address: 0x4507cEf57C46789eF8d1a19EA45f4216bae2B528 - ETH contract address: 0x4507cEf57C46789eF8d1a19EA45f4216bae2B528 (NOTE: The contract address is the same on both the BSC and ETH chains). FLOKI stakers will earn 56% of the supply over a period of 4 years. The Floki community is known for being diamond-handed, and with TokenFi being in an industry with MASSIVE potential, FLOKI holders are the best candidates to receive the majority of TokenFi tokens. Since they would be staking their FLOKI tokens to get the new token, this will also ensure stability for the FLOKI token! TokenFi supply will be split evenly between the BSC and ETH chains: In other words, there will be 5 billion tokens on BSC and 5 billion tokens on ETH, to make for a combined total of 10 billion tokens. 5% of the supply will be paired with LP on BSC and the other 5% on ETH. This will make the starting circulating market cap on BSC a $25k market cap and on ETH a $25k market cap to make a combined initial circulating market cap of $50k and an initial fully diluted market cap of $500k (which will be gradually released to FLOKI stakers over a four year period). IMPORTANT INFORMATION FOR THE FIRST HOUR OF LAUNCH We understand that with this being a Floki token, there is a lot of hype. As a result, we have put measures in place to limit the impact of snipers on the token: Specifically, there will be a 1% wallet cap (of total supply) within the first hour of launch. That means no individual wallet can buy more than 100 million tokens within the first hour of the token becoming tradable. In addition, there will be an initial buy/sell transaction tax of 20% within the first hour. This transaction tax will do two things: 1) potentially limit the impact of snipers and 2) ensure a significant portion of whatever snipers/early buyers spend goes to the Floki treasury, which can be used for growth and development efforts. After the first hour, the wallet cap will be removed, and the transaction tax will be lowered to 5%. This 5% tax will remain in place for a week, after which the Floki DAO will vote on whether or not to remove or reduce it. We have instructed our exchange partners not to list TokenFi until this DAO vote. OUR PLAN FOR TOKENFI TokenFi is a well-thought-out concept that we have a strong capability to deliver on! We will unveil our roadmap with the launch of the TokenFi website on the 27th of October, 2023. However, we assure you that several TokenFi products are in advanced development on testnet and are due to go live in Q4 2023. In addition, we are working with some of the biggest names in the industry - especially from an institutional perspective - to make TokenFi a success, and they are quite excited about the concept and its potential. In our original DAO proposal, we already announced DWF Labs as our main institutional partner and market maker for TokenFi. We also announced a strategic partnership with World Table Tennis that will introduce TokenFi to a massive audience of 120 million people. We will announce many more partners in the coming weeks and months, and I'm sure that when you see the moves we have made, you will see why there is no better person to execute this vision than Floki!

FLOKI

1,955,825 次观看 • 2 年前

Grok Bot might be the first tool that lets one non-technical person run an entire business with a team of AI agents. My friend Billy runs his whole newsletter business on Grok Bot agents, and I think we're about to see 100,000+ businesses like his. BEST PRACTICES: 1. The agents run on a shared cloud computer, so running your newsletter, your X, and your receipts all in one place creates context bloat and burns tokens fast. One mission per setup. 2. Start with a Chief of Staff. Give it access to your existing docs (Notion, Slack, Gmail), have it audit the business, then tell you the top three agents to build first to drive revenue. 3. Perfect a task with the Chief of Staff before spinning up a new agent. Have it do the outbound sales once, review it, and only then say "now build a bot that does exactly that." You earn each new hire by proving the task works first. 4. Constraints are the feature. You get a limited number of agents, one thread per bot, like DMs with a teammate. It forces you to stay mission-oriented instead of spinning up a bot for every random idea. 5. You make the decisions, not the agent. Billy's team spent three weeks unable to pick where content should live. At some point you say "we're doing Notion, no more tinkering" and move on. 6. Run week one with no new agents. Build the team, learn to fly the plane, just execute. Week three is when you find the real gaps and expand, someone to man the inbox, someone for the Shopify shop. 7. Then add routines so it works while you sleep. Ask your Chief of Staff what recurring jobs would move the business forward overnight, and it builds the automations that run without you. Thanks to Billy Howell for sharing the sauce on The Startup Ideas Podcast (SIP) 🧃 (follow for more). Grokbot is really cool. Watch below:
44:20

Grok Bot might be the first tool that lets one non-technical person run an entire business with a team of AI agents. My friend Billy runs his whole newsletter business on Grok Bot agents, and I think we're about to see 100,000+ businesses like his. BEST PRACTICES: 1. The agents run on a shared cloud computer, so running your newsletter, your X, and your receipts all in one place creates context bloat and burns tokens fast. One mission per setup. 2. Start with a Chief of Staff. Give it access to your existing docs (Notion, Slack, Gmail), have it audit the business, then tell you the top three agents to build first to drive revenue. 3. Perfect a task with the Chief of Staff before spinning up a new agent. Have it do the outbound sales once, review it, and only then say "now build a bot that does exactly that." You earn each new hire by proving the task works first. 4. Constraints are the feature. You get a limited number of agents, one thread per bot, like DMs with a teammate. It forces you to stay mission-oriented instead of spinning up a bot for every random idea. 5. You make the decisions, not the agent. Billy's team spent three weeks unable to pick where content should live. At some point you say "we're doing Notion, no more tinkering" and move on. 6. Run week one with no new agents. Build the team, learn to fly the plane, just execute. Week three is when you find the real gaps and expand, someone to man the inbox, someone for the Shopify shop. 7. Then add routines so it works while you sleep. Ask your Chief of Staff what recurring jobs would move the business forward overnight, and it builds the automations that run without you. Thanks to Billy Howell for sharing the sauce on The Startup Ideas Podcast (SIP) 🧃 (follow for more). Grokbot is really cool. Watch below:

GREG ISENBERG

4,026,671 次观看 • 1 个月前

Destiny 2 left a permanent mark on my life. The memories I’ve made with this game, both as a player and as someone lucky enough to work on it, will forever stay with me. Not just because of late nights running King’s Fall with friends, or grinding out the Crucible Glorious Seal in solo queue like a complete maniac, but because Destiny 2 challenged me, shaped me, and pushed me toward becoming the creative I am today. My journey into digital art and photography started back in 2007, when I began taking screenshots in Halo 3 and entering Bungie community art contests. The relationships I built during that time eventually opened the door to an opportunity with Bungie’s Gameplay Capture team in 2014, helping the team capture footage for a new game called Destiny. I had no idea then that a commendably short two-week contract would turn into an incredible 12-year journey with this franchise. I originally came to Bungie hoping to pursue environmental concept art, but along the way I had the opportunity to work on marketing art and quickly fell in love with it. Creating marketing art for a franchise like Destiny challenged me in so many different ways. It forced me to expand my technical, creative, and communication skill sets, constantly adapt, and keep learning new tools and workflows just to keep pace with the live-service beast that was Destiny 2. It means a lot to be able to look back and say with confidence that I had a visual impact on Destiny 2; but the most meaningful part of it all was getting to work alongside and learn from so many incredible artists at Bungie, and seeing firsthand just how much care, effort, and humanity it takes to make work like this possible. It’s hard to fully capture how much energy, skill, and collaboration goes into every visual part of a game like Destiny 2. That work is shaped by artists from different backgrounds, experiences, disciplines, and perspectives. Each of them, including me, left a small part of themselves in what they created. To me, that’s what makes a game like Destiny 2 feel truly meaningful and memorable. Especially now, in a world increasingly saturated with content and driven by instant output and gratification through AI, I keep coming back to the value of process. For me, and for so many of my peers, it was never only about arriving at the final image. It was about the journey it took to get there. The late nights. The iteration. The problem-solving. The trust. The shared pursuit of trying to make something special. To my peers, I’m deeply proud of what we built together, but even more grateful for how we built it. We challenged each other, inspired each other, and kept showing up for one another through every high and low, as Destiny 2 has had many. That kind of shared effort leaves a lasting mark. What we made mattered. What we gave mattered. And the impact of what we built together will stay with us, and this community, for a long time. Shoutout to the current and former members of Creative Studios, the VizD team, and the many Bungie developers and marketers who helped shape this chapter of my life. I’m especially grateful to the teammates who believed in me, encouraged me to embrace failure and new beginnings as essential parts of artistic growth, and pushed me to take on challenges even when they felt beyond my reach. You showed me that the strength of a team will always surpass that of any individual hero. As my work on #Destiny2 comes to a close and I look toward the future, I plan to spend the next few weeks sharing some of the pieces I had the chance to create or art direct that mean the most to me. For now, I’ll leave you all with a collage of some of my personal favorite pieces to work on across Destiny 2’s lifetime. These projects mean so much to me because many of them started as personal passion projects or late-night concept sketches, inspired by playing early builds of Destiny 2 and by the incredible work of our development team.
0:20

Destiny 2 left a permanent mark on my life. The memories I’ve made with this game, both as a player and as someone lucky enough to work on it, will forever stay with me. Not just because of late nights running King’s Fall with friends, or grinding out the Crucible Glorious Seal in solo queue like a complete maniac, but because Destiny 2 challenged me, shaped me, and pushed me toward becoming the creative I am today. My journey into digital art and photography started back in 2007, when I began taking screenshots in Halo 3 and entering Bungie community art contests. The relationships I built during that time eventually opened the door to an opportunity with Bungie’s Gameplay Capture team in 2014, helping the team capture footage for a new game called Destiny. I had no idea then that a commendably short two-week contract would turn into an incredible 12-year journey with this franchise. I originally came to Bungie hoping to pursue environmental concept art, but along the way I had the opportunity to work on marketing art and quickly fell in love with it. Creating marketing art for a franchise like Destiny challenged me in so many different ways. It forced me to expand my technical, creative, and communication skill sets, constantly adapt, and keep learning new tools and workflows just to keep pace with the live-service beast that was Destiny 2. It means a lot to be able to look back and say with confidence that I had a visual impact on Destiny 2; but the most meaningful part of it all was getting to work alongside and learn from so many incredible artists at Bungie, and seeing firsthand just how much care, effort, and humanity it takes to make work like this possible. It’s hard to fully capture how much energy, skill, and collaboration goes into every visual part of a game like Destiny 2. That work is shaped by artists from different backgrounds, experiences, disciplines, and perspectives. Each of them, including me, left a small part of themselves in what they created. To me, that’s what makes a game like Destiny 2 feel truly meaningful and memorable. Especially now, in a world increasingly saturated with content and driven by instant output and gratification through AI, I keep coming back to the value of process. For me, and for so many of my peers, it was never only about arriving at the final image. It was about the journey it took to get there. The late nights. The iteration. The problem-solving. The trust. The shared pursuit of trying to make something special. To my peers, I’m deeply proud of what we built together, but even more grateful for how we built it. We challenged each other, inspired each other, and kept showing up for one another through every high and low, as Destiny 2 has had many. That kind of shared effort leaves a lasting mark. What we made mattered. What we gave mattered. And the impact of what we built together will stay with us, and this community, for a long time. Shoutout to the current and former members of Creative Studios, the VizD team, and the many Bungie developers and marketers who helped shape this chapter of my life. I’m especially grateful to the teammates who believed in me, encouraged me to embrace failure and new beginnings as essential parts of artistic growth, and pushed me to take on challenges even when they felt beyond my reach. You showed me that the strength of a team will always surpass that of any individual hero. As my work on #Destiny2 comes to a close and I look toward the future, I plan to spend the next few weeks sharing some of the pieces I had the chance to create or art direct that mean the most to me. For now, I’ll leave you all with a collage of some of my personal favorite pieces to work on across Destiny 2’s lifetime. These projects mean so much to me because many of them started as personal passion projects or late-night concept sketches, inspired by playing early builds of Destiny 2 and by the incredible work of our development team.

Biwald

73,319 次观看 • 3 个月前