Video wird geladen...

Video konnte nicht geladen werden

Beim Laden dieses Videos ist ein Problem aufgetreten. Dies könnte an einem vorübergehenden Netzwerkproblem liegen oder das Video ist möglicherweise nicht verfügbar.

OpenAI Brad Lightcap says model training cycles are collapsing, driving rapid jumps from GPT-5.1 to 5.4 GPT-5.4 is days old and already doing a $1B run rate revenue, processing 5 trillion tokens daily "by year's end, we'll laugh at today's models. they'll seem pedestrian"

Haider.

66,633 subscribers

54,636 Aufrufe • vor 3 Monaten •via X (Twitter)

Nachrichten & Politik Bildung Wissenschaft & Technologie

Anya Rossi• Live Now

Private livecam show

0 Kommentare

Keine Kommentare verfügbar

Kommentare vom Original-Post werden hier angezeigt

Ähnliche Videos

The value produced by models is getting so much better so fast that old hardware is actually getting *more* expensive to rent. 3 years ago, the best model you could run on a H100 chip was GPT-4. Now, you can run GPT-5.4 on it, which is smaller and cheaper to run while producing much more valuable tokens. w. Dylan Patel

The value produced by models is getting so much better so fast that old hardware is actually getting more expensive to rent. 3 years ago, the best model you could run on a H100 chip was GPT-4. Now, you can run GPT-5.4 on it, which is smaller and cheaper to run while producing much more valuable tokens. w. Dylan Patel

Dwarkesh Patel

78,277 Aufrufe • vor 4 Monaten

Dylan Patel says GPT-5 is not a bigger model and did not gain intelligence through scale but through optimization. He highlights how OpenAI improved efficiency by cutting down wasted tokens compared to earlier o-series models. He adds that GPT-5’s new router system decides whether to use the base model, mini, or thinking, leading to faster and more efficient results. He points out that o3 could think for around 30 seconds, while GPT-5 only thinks for 5–10 seconds, meaning less compute is used per query.

Dylan Patel says GPT-5 is not a bigger model and did not gain intelligence through scale but through optimization. He highlights how OpenAI improved efficiency by cutting down wasted tokens compared to earlier o-series models. He adds that GPT-5’s new router system decides whether to use the base model, mini, or thinking, leading to faster and more efficient results. He points out that o3 could think for around 30 seconds, while GPT-5 only thinks for 5–10 seconds, meaning less compute is used per query.

Wes Roth

87,396 Aufrufe • vor 11 Monaten

Thanksgiving-week treat: an epic conversation on Frontier AI with Lukasz Kaiser -co-author of “Attention Is All You Need” (Transformers) and leading research scientist at OpenAI working on GPT-5.1-era reasoning models. 00:00 – Cold open and intro 01:29 – “AI slowdown” vs a wild week of new frontier models 08:03 – Low-hanging fruit, infra, RL training and better data 11:39 – What is a reasoning model, in plain language 17:02 – Chain-of-thought and training the thinking process with RL 21:39 – Łukasz’s path: from logic and France to Google and Kurzweil 24:20 – Inside the Transformer story and what “attention” really means 28:42 – From Google Brain to OpenAI: culture, scale and GPUs 32:49 – What’s next for pre-training, GPUs and distillation 37:29 – Can we still understand these models? Circuits, sparsity and black boxes 39:42 – GPT-4 → GPT-5 → GPT-5.1: what actually changed 42:40 – Post-training, safety and teaching GPT-5.1 different tones 46:16 – How long should GPT-5.1 think? Reasoning tokens and jagged abilities 47:43 – The five-year-old’s dot puzzle that still breaks frontier models 52:22 – Generalization, child-like learning and whether reasoning is enough 53:48 – Beyond Transformers: ARC, LeCun’s ideas and multimodal bottlenecks 56:10 – GPT-5.1 Codex Max, long-running agents and compaction 1:00:06 – Will foundation models eat most apps? The translation analogy and trust 1:02:34 – What still needs to be solved, and where AI might go next

Thanksgiving-week treat: an epic conversation on Frontier AI with Lukasz Kaiser -co-author of “Attention Is All You Need” (Transformers) and leading research scientist at OpenAI working on GPT-5.1-era reasoning models. 00:00 – Cold open and intro 01:29 – “AI slowdown” vs a wild week of new frontier models 08:03 – Low-hanging fruit, infra, RL training and better data 11:39 – What is a reasoning model, in plain language 17:02 – Chain-of-thought and training the thinking process with RL 21:39 – Łukasz’s path: from logic and France to Google and Kurzweil 24:20 – Inside the Transformer story and what “attention” really means 28:42 – From Google Brain to OpenAI: culture, scale and GPUs 32:49 – What’s next for pre-training, GPUs and distillation 37:29 – Can we still understand these models? Circuits, sparsity and black boxes 39:42 – GPT-4 → GPT-5 → GPT-5.1: what actually changed 42:40 – Post-training, safety and teaching GPT-5.1 different tones 46:16 – How long should GPT-5.1 think? Reasoning tokens and jagged abilities 47:43 – The five-year-old’s dot puzzle that still breaks frontier models 52:22 – Generalization, child-like learning and whether reasoning is enough 53:48 – Beyond Transformers: ARC, LeCun’s ideas and multimodal bottlenecks 56:10 – GPT-5.1 Codex Max, long-running agents and compaction 1:00:06 – Will foundation models eat most apps? The translation analogy and trust 1:02:34 – What still needs to be solved, and where AI might go next

Matt Turck

168,007 Aufrufe • vor 8 Monaten

🚨 do you understand what just happened to Microsoft.. Microsoft just dropped seven of its own MAI models, trained from scratch with zero distillation, and said its custom-tuned models already match GPT-tier quality at 10x lower cost. The partner that pays OpenAI's bills is now quietly building the thing that replaces it. - MAI-Thinking-1 hits human-preference parity with Sonnet 4.6 and 53% on SWE Bench Pro, right next to Opus 4.6 - MAI-Code-1-Flash delivers Haiku-class coding at just 5B params and is already shipping inside GitHub Copilot - A custom MAI model tuned for Excel matched GPT-5.4 while running up to 10x more efficiently - Their real weapon isn't the weights, it's that they're already inside everyone's Office, Teams, and VS Code The hard part was never training a good model. The hard part is being the room everyone already works in. Microsoft owns the room.

🚨 do you understand what just happened to Microsoft.. Microsoft just dropped seven of its own MAI models, trained from scratch with zero distillation, and said its custom-tuned models already match GPT-tier quality at 10x lower cost. The partner that pays OpenAI's bills is now quietly building the thing that replaces it. - MAI-Thinking-1 hits human-preference parity with Sonnet 4.6 and 53% on SWE Bench Pro, right next to Opus 4.6 - MAI-Code-1-Flash delivers Haiku-class coding at just 5B params and is already shipping inside GitHub Copilot - A custom MAI model tuned for Excel matched GPT-5.4 while running up to 10x more efficiently - Their real weapon isn't the weights, it's that they're already inside everyone's Office, Teams, and VS Code The hard part was never training a good model. The hard part is being the room everyone already works in. Microsoft owns the room.

BuBBliK

171,865 Aufrufe • vor 1 Monat

GT Protocol AI Digest No.82: AI Starts Running the System Here are five stories worth your attention 👇 💻 NVIDIA Pushes AI to Your Device NVIDIA showcased RTX PCs and DGX systems that can run AI models and agents locally. The shift is important. AI is no longer tied to the cloud, it is moving directly onto user devices. 🧠 Gemini Becomes Personal Google is rolling out personalized Gemini across the US. The assistant adapts to your context and activity, turning AI from a generic tool into a persistent, tailored system. 🎨 Adobe Lets You Train Your Own AI Adobe now allows users to train Firefly models on their own artwork. This gives creators more control over style, but also raises new questions around ownership and dataset boundaries. ⚡️ Smaller GPT-5.4 Models Go Wide OpenAI introduced GPT-5.4 mini and nano, optimized for efficiency and broader deployment. Strong models are no longer limited to heavy infrastructure. 💳 Visa Prepares for AI Payments Visa is building systems for transactions initiated by AI agents. This moves AI beyond interfaces and into real economic activity. Read the full digest on our Medium 👉

GT Protocol AI Digest No.82: AI Starts Running the System Here are five stories worth your attention 👇 💻 NVIDIA Pushes AI to Your Device NVIDIA showcased RTX PCs and DGX systems that can run AI models and agents locally. The shift is important. AI is no longer tied to the cloud, it is moving directly onto user devices. 🧠 Gemini Becomes Personal Google is rolling out personalized Gemini across the US. The assistant adapts to your context and activity, turning AI from a generic tool into a persistent, tailored system. 🎨 Adobe Lets You Train Your Own AI Adobe now allows users to train Firefly models on their own artwork. This gives creators more control over style, but also raises new questions around ownership and dataset boundaries. ⚡️ Smaller GPT-5.4 Models Go Wide OpenAI introduced GPT-5.4 mini and nano, optimized for efficiency and broader deployment. Strong models are no longer limited to heavy infrastructure. 💳 Visa Prepares for AI Payments Visa is building systems for transactions initiated by AI agents. This moves AI beyond interfaces and into real economic activity. Read the full digest on our Medium 👉

GT Protocol

40,065 Aufrufe • vor 4 Monaten

Ridge CEO Sean Frank says he'd be willing to give OpenAI a 40% cut of his revenue for ad placement in GPT, because he's already giving 50% to Meta to sell his products. "The beauty of the Facebook ad product is that they figured out a way to take half of everybody's revenue." "They figured out exactly how much money they could take from everybody at the optimal level, and that's why they do $100 billion a year in revenue." "So if the offer from ChatGPT is 'Give us 40% of your revenue, and we'll give you new customers you wouldn't get otherwise,' I would take that deal because I'm already giving Mark Zuckerberg 50%."

Ridge CEO Sean Frank says he'd be willing to give OpenAI a 40% cut of his revenue for ad placement in GPT, because he's already giving 50% to Meta to sell his products. "The beauty of the Facebook ad product is that they figured out a way to take half of everybody's revenue." "They figured out exactly how much money they could take from everybody at the optimal level, and that's why they do $100 billion a year in revenue." "So if the offer from ChatGPT is 'Give us 40% of your revenue, and we'll give you new customers you wouldn't get otherwise,' I would take that deal because I'm already giving Mark Zuckerberg 50%."

TBPN

43,375 Aufrufe • vor 6 Monaten

Karpathy told Dwarkesh that a 1 billion parameter model, trained on clean data, could hit the intelligence of today's 1.8 trillion parameter frontier. That is a 1,800x compression claim. The math behind it is more defensible than it sounds. When researchers at frontier labs look at random samples from their training corpus, they see stock ticker symbols, broken HTML, forum spam, autogenerated gibberish. Not Wikipedia. Not the Wall Street Journal. The actual pretraining dataset is mostly noise, and the model is burning parameters to vaguely remember all of it. One estimate pegs Llama 3's information compression at 0.07 bits per token. Well-structured English carries around 1.5 bits per token of real information. The trillion-parameter model is holding a roughly 5% resolution image of the internet it trained on. So when a lab ships a 1.8 trillion parameter model, the overwhelming majority of those weights are handling rough memorization. They are compression overhead for a noisy training set, taking up capacity that could be doing reasoning instead. Karpathy's proposal is to separate the two. Build a cognitive core: a small model that contains only the algorithms for reasoning and problem-solving, stripped of encyclopedic memorization. Pair it with external memory the model queries when it needs a fact. A 1 billion parameter reasoner plus retrieval beats a 1.8 trillion parameter model trying to do both. The data already supports this direction. GPT-4o runs at roughly 200 billion parameters and outperforms the original 1.8 trillion GPT-4. Inference costs for GPT-3.5 level performance fell 280x between 2022 and 2024, driven almost entirely by smaller, cleaner, better-architected models. The trend line is pointing where Karpathy says it should. The real implication for anyone tracking the AI trade: data quality is the actual constraint. The companies winning the next phase will be the ones who figured out what to train on, and what to throw away.

Karpathy told Dwarkesh that a 1 billion parameter model, trained on clean data, could hit the intelligence of today's 1.8 trillion parameter frontier. That is a 1,800x compression claim. The math behind it is more defensible than it sounds. When researchers at frontier labs look at random samples from their training corpus, they see stock ticker symbols, broken HTML, forum spam, autogenerated gibberish. Not Wikipedia. Not the Wall Street Journal. The actual pretraining dataset is mostly noise, and the model is burning parameters to vaguely remember all of it. One estimate pegs Llama 3's information compression at 0.07 bits per token. Well-structured English carries around 1.5 bits per token of real information. The trillion-parameter model is holding a roughly 5% resolution image of the internet it trained on. So when a lab ships a 1.8 trillion parameter model, the overwhelming majority of those weights are handling rough memorization. They are compression overhead for a noisy training set, taking up capacity that could be doing reasoning instead. Karpathy's proposal is to separate the two. Build a cognitive core: a small model that contains only the algorithms for reasoning and problem-solving, stripped of encyclopedic memorization. Pair it with external memory the model queries when it needs a fact. A 1 billion parameter reasoner plus retrieval beats a 1.8 trillion parameter model trying to do both. The data already supports this direction. GPT-4o runs at roughly 200 billion parameters and outperforms the original 1.8 trillion GPT-4. Inference costs for GPT-3.5 level performance fell 280x between 2022 and 2024, driven almost entirely by smaller, cleaner, better-architected models. The trend line is pointing where Karpathy says it should. The real implication for anyone tracking the AI trade: data quality is the actual constraint. The companies winning the next phase will be the ones who figured out what to train on, and what to throw away.

Aakash Gupta

507,952 Aufrufe • vor 3 Monaten

OpenAI Dev Day Just kicked off and it was the biggest day for AI since the launch of GPT-4 in April. Here's a 3 minute video with everything you need to know and a full summary: 1. GPT-4 Turbo is released. It is much better than GPT-4, can accept up to 128k characters and is 3x cheaper. The knowledge cut off is now April 2023 2. Text to Speech is also being released. They offer 6 hyper-realistic voice models that are the best I've heard from any other AI t-t-s software on the market. 3. Copyright Shield. OpenAI will actually defend your legal costs if you get sued for copyright infringement using their API! 4. GPTs. Maybe the most exciting IMHO, you can now create custom "GPTs" by training your own model of GPT and selling it on their store. This will create a massive shakeup of the current AI software landscape and is the first step towards autonomous agents.

OpenAI Dev Day Just kicked off and it was the biggest day for AI since the launch of GPT-4 in April. Here's a 3 minute video with everything you need to know and a full summary: 1. GPT-4 Turbo is released. It is much better than GPT-4, can accept up to 128k characters and is 3x cheaper. The knowledge cut off is now April 2023 2. Text to Speech is also being released. They offer 6 hyper-realistic voice models that are the best I've heard from any other AI t-t-s software on the market. 3. Copyright Shield. OpenAI will actually defend your legal costs if you get sued for copyright infringement using their API! 4. GPTs. Maybe the most exciting IMHO, you can now create custom "GPTs" by training your own model of GPT and selling it on their store. This will create a massive shakeup of the current AI software landscape and is the first step towards autonomous agents.

Mark Gadala-Maria

79,177 Aufrufe • vor 2 Jahren

Andrej Karpathy just made one of the most interesting arguments about AI model design that most people are completely missing. His take is that frontier AI models are not too big because the technology is complex and too big because the training data is garbage. When you or I think of the internet, we picture Wall Street Journal articles, Wikipedia entries, serious writing. That is not what a pretraining dataset looks like. When researchers at frontier labs look at random documents from the actual training corpus, it is stock ticker symbols, broken HTML, spam, gibberish. One estimate puts Llama 3's information compression at just 0.07 bits per token meaning the model has only a hazy recollection of most of what it trained on. So we build trillion parameter models not because we need a trillion parameter brain but because we need a trillion-parameter compression engine to squeeze some intelligence out of a firehose of noise. Most of those parameters are doing memory work, not cognitive work. Karpathy's prediction is separate the two entirely. Build a cognitive core, a model that contains only the algorithms for reasoning and problem-solving, stripped of encyclopedic memorization and pair it with external memory that it can query when it needs facts. He thinks a cognitive core trained on high-quality data could hit genuine intelligence at around one billion parameters. For reference, today's flagship models run between 200 billion and 1.8 trillion parameters with most of that weight dedicated to remembering the internet's slop. The trend is already moving his direction. GPT-4o operates at roughly 200 billion parameters and outperforms the original 1.8 trillion-parameter GPT-4. Inference costs for GPT-3.5-level performance dropped 280-fold between 2022 and 2024 driven almost entirely by smaller, cleaner, better-architected models. The real bottleneck in AI right now is not compute but rather data quality.

Andrej Karpathy just made one of the most interesting arguments about AI model design that most people are completely missing. His take is that frontier AI models are not too big because the technology is complex and too big because the training data is garbage. When you or I think of the internet, we picture Wall Street Journal articles, Wikipedia entries, serious writing. That is not what a pretraining dataset looks like. When researchers at frontier labs look at random documents from the actual training corpus, it is stock ticker symbols, broken HTML, spam, gibberish. One estimate puts Llama 3's information compression at just 0.07 bits per token meaning the model has only a hazy recollection of most of what it trained on. So we build trillion parameter models not because we need a trillion parameter brain but because we need a trillion-parameter compression engine to squeeze some intelligence out of a firehose of noise. Most of those parameters are doing memory work, not cognitive work. Karpathy's prediction is separate the two entirely. Build a cognitive core, a model that contains only the algorithms for reasoning and problem-solving, stripped of encyclopedic memorization and pair it with external memory that it can query when it needs facts. He thinks a cognitive core trained on high-quality data could hit genuine intelligence at around one billion parameters. For reference, today's flagship models run between 200 billion and 1.8 trillion parameters with most of that weight dedicated to remembering the internet's slop. The trend is already moving his direction. GPT-4o operates at roughly 200 billion parameters and outperforms the original 1.8 trillion-parameter GPT-4. Inference costs for GPT-3.5-level performance dropped 280-fold between 2022 and 2024 driven almost entirely by smaller, cleaner, better-architected models. The real bottleneck in AI right now is not compute but rather data quality.

Milk Road AI

200,325 Aufrufe • vor 3 Monaten

Factory (Factory) CEO Matan Grinberg (Matan Grinberg) says "In the next 12 months, 90% of tokens will be going to open models." Enterprises are already shifting rapidly.. "At the beginning of the year, <1% of their tokens were going to open models." "Around March, it crossed 1%... & in May it crossed 10%." "These open models are incredibly performant, incredibly cheap, incredibly fast." While frontier closed models aren't going away: "They'll still have a place, but I think that place will be shrinking—at least in token share." And by the end of 2026? "At least in the enterprise, we'll probably cross the 50% threshold."

Factory (Factory) CEO Matan Grinberg (Matan Grinberg) says "In the next 12 months, 90% of tokens will be going to open models." Enterprises are already shifting rapidly.. "At the beginning of the year, <1% of their tokens were going to open models." "Around March, it crossed 1%... & in May it crossed 10%." "These open models are incredibly performant, incredibly cheap, incredibly fast." While frontier closed models aren't going away: "They'll still have a place, but I think that place will be shrinking—at least in token share." And by the end of 2026? "At least in the enterprise, we'll probably cross the 50% threshold."

Molly O’Shea

120,683 Aufrufe • vor 17 Tagen

5 patterns in Text Arena's price–performance Pareto frontier since 2023: 1. GPT-4-level quality is now ~500x lower cost. - From a ~$50 blended price per million tokens in 2023 to ~$0.10 today. 2. The higher-price end is both better and lower-priced since 2023. - The leading Arena score has climbed ~170 points (1,330 → 1,500). While the price of the higher-end frontier models dropped from ~$50 to ~$20 per million tokens. 3. The low-cost end gained the most. - Under $0.20 per million tokens, the best available model went from ~1,000 Arena score in 2023 to ~1,440 today. 4. The low-cost/top performance gap has nearly closed. - In 2023, sub-$0.20 models trailed the leader by ~350 Arena points. Today, ~60. 5. The cast has rotated quite a bit. - - OpenAI set the 2023–24 benchmark. - AI at Meta strengthened the low-cost end in 2024. - Google DeepMind drove the 2025 jump. - Anthropic holds the peak in 2026. - xAI and Chinese labs like DeepSeek AI, Z.ai, Kimi.ai, Xiaomi MiMo, and Qwen are continuing to push the mid-price frontier.

5 patterns in Text Arena's price–performance Pareto frontier since 2023: 1. GPT-4-level quality is now ~500x lower cost. - From a ~$50 blended price per million tokens in 2023 to ~$0.10 today. 2. The higher-price end is both better and lower-priced since 2023. - The leading Arena score has climbed ~170 points (1,330 → 1,500). While the price of the higher-end frontier models dropped from ~$50 to ~$20 per million tokens. 3. The low-cost end gained the most. - Under $0.20 per million tokens, the best available model went from ~1,000 Arena score in 2023 to ~1,440 today. 4. The low-cost/top performance gap has nearly closed. - In 2023, sub-$0.20 models trailed the leader by ~350 Arena points. Today, ~60. 5. The cast has rotated quite a bit. - - OpenAI set the 2023–24 benchmark. - AI at Meta strengthened the low-cost end in 2024. - Google DeepMind drove the 2025 jump. - Anthropic holds the peak in 2026. - xAI and Chinese labs like DeepSeek AI, Z.ai, Kimi.ai, Xiaomi MiMo, and Qwen are continuing to push the mid-price frontier.

Arena.ai

58,133 Aufrufe • vor 2 Monaten

Game Arena from Generality, Inc. is the largest LLM strategy game tournament to date. Games are great for measuring LLMs on instruction following, long-horizon planning, and problem-solving. In fact, models that are Olympiad-level at math and coding often struggle to make accurate moves in Game Arena (median illegal-move rate: 11.4%). Game environments also resist contamination and saturation: every run is unique, so there’s no risk of training-data leakage, and the bar keeps rising as models improve. Game Arena has digitized hundreds of board games into fresh new environments, and today it’s debuting its first tournament, with models like GPT-5 High, Claude Opus 4.1, and DeepSeek V3.1 from OpenAI, Anthropic, Qwen, DeepSeek, Google, and more going head-to-head. You can watch game replays and full model reasonings on Congrats on the launch, @sanerc110 and Kayla Lee!

Game Arena from Generality, Inc. is the largest LLM strategy game tournament to date. Games are great for measuring LLMs on instruction following, long-horizon planning, and problem-solving. In fact, models that are Olympiad-level at math and coding often struggle to make accurate moves in Game Arena (median illegal-move rate: 11.4%). Game environments also resist contamination and saturation: every run is unique, so there’s no risk of training-data leakage, and the bar keeps rising as models improve. Game Arena has digitized hundreds of board games into fresh new environments, and today it’s debuting its first tournament, with models like GPT-5 High, Claude Opus 4.1, and DeepSeek V3.1 from OpenAI, Anthropic, Qwen, DeepSeek, Google, and more going head-to-head. You can watch game replays and full model reasonings on Congrats on the launch, @sanerc110 and Kayla Lee!

Y Combinator

21,862 Aufrufe • vor 9 Monaten

Happy GPT-5 Day to all who celebrate. Here's what I expect: It will be agentic-first (this is all but confirmed by OpenAI) It will be heavily multimodal (audio, video, music, tool use, and more) It will be a reasoning-first model as well (math, planning, etc) Directionally, I suspect we're looking at moving towards a convergence of "everything to everything" end-to-end models. Sort of like how Tesla improved FSD by doing neural networks for end-to-end perception-planning-control. Likewise, I suspect that, for the foreseeable future on general purpose models, we're going to converge towards "all in one" models. They will just tend to get smarter and smarter the more modalities we stuff into them (world models, math, image, text, video, audio, etc).

Happy GPT-5 Day to all who celebrate. Here's what I expect: It will be agentic-first (this is all but confirmed by OpenAI) It will be heavily multimodal (audio, video, music, tool use, and more) It will be a reasoning-first model as well (math, planning, etc) Directionally, I suspect we're looking at moving towards a convergence of "everything to everything" end-to-end models. Sort of like how Tesla improved FSD by doing neural networks for end-to-end perception-planning-control. Likewise, I suspect that, for the foreseeable future on general purpose models, we're going to converge towards "all in one" models. They will just tend to get smarter and smarter the more modalities we stuff into them (world models, math, image, text, video, audio, etc).

David Shapiro (L/0)

19,305 Aufrufe • vor 11 Monaten

Found a way to run Kimi K3 for FREE with daily credits on 😳 it's a 2.8T parameter model. 1M context. beats Fable 5 on coding benchmarks. what you get for $0: -2.8 trillion parameters -1M context window -beats Fable 5 on SWE Marathon (42 vs 35) -beats Fable 5 on Program Bench (77.8) -topped BrowseComp (91.2) -native vision understanding What this replaces: -Paid frontier API costs for coding agents -Fable 5 at $30/M output tokens -GPT-5.6 Sol at $25/M output tokens how to use it (2 min): 1. Go to 2. Sign up with email 3. Start using K3 — it's the default model Also available on desktop app, CLI, and mobile. Important: Free daily credits reset. Full open weights dropping July 27. Rate limits apply on free tier. use this before the free tier gets crowded

Found a way to run Kimi K3 for FREE with daily credits on 😳 it's a 2.8T parameter model. 1M context. beats Fable 5 on coding benchmarks. what you get for $0: -2.8 trillion parameters -1M context window -beats Fable 5 on SWE Marathon (42 vs 35) -beats Fable 5 on Program Bench (77.8) -topped BrowseComp (91.2) -native vision understanding What this replaces: -Paid frontier API costs for coding agents -Fable 5 at $30/M output tokens -GPT-5.6 Sol at $25/M output tokens how to use it (2 min): 1. Go to 2. Sign up with email 3. Start using K3 — it's the default model Also available on desktop app, CLI, and mobile. Important: Free daily credits reset. Full open weights dropping July 27. Rate limits apply on free tier. use this before the free tier gets crowded

painn

246,003 Aufrufe • vor 7 Tagen

GPT-5 is live in Cline. We've been working with OpenAI to get this model ready, and here's our take: it's disciplined, persistent, & highly competent. It's collaborative in planning & and a diligent operator while acting. It plans thoroughly, asks optioned follow-ups when needed, & then gets out of the way and ships code. On long tasks it keeps going before pausing to check in. It follows instructions to the letter. And most importantly -- it writes good code. GPT-5 is like "The Wolf" from Pulp Fiction. Comes in, assesses the situation, then executes. Here's what you can expect from GPT-5 in Cline: > verbose while planning; terse while executing > asks a lot of good clarification questions, & frequently provides options when appropriate > strong context retention and persistence over long horizons (256k context window) > good at diff-style edits and multi-file changes (we'll monitor as more usage data comes in) > quiet in Act mode -- writes code without yapping Metaprompting is another strength. We tested early with OpenAI and used GPT-5 to tune our own prompt for GPT-5. Here's a pattern we like: “Answer from your own perspective: what changes or additions would help you better follow this prompt? Here is the prompt (or snippet): [snippet]. Users have complained about X and Y. What minimal edits would you make while keeping the rest intact?” Do you need to change any of your existing patterns in Cline? No -- it's good out of the box. Give a clear goal and constraints, let it plan, then let it cook. Expect more clarifying questions than most models. Pricing: $1.25/M input tokens (+90% cache), $10/M output. Roughly half of Sonnet 4 ($3/$15). Want to try GPT-5? Use it in Cline today for pure, unfiltered inference via the OpenAI, Cline, or OpenRouter providers. (fyi -- GPT-5 one-shotted this browser DAW below on the prompt "build something impressive to show me what you're capable of")

GPT-5 is live in Cline. We've been working with OpenAI to get this model ready, and here's our take: it's disciplined, persistent, & highly competent. It's collaborative in planning & and a diligent operator while acting. It plans thoroughly, asks optioned follow-ups when needed, & then gets out of the way and ships code. On long tasks it keeps going before pausing to check in. It follows instructions to the letter. And most importantly -- it writes good code. GPT-5 is like "The Wolf" from Pulp Fiction. Comes in, assesses the situation, then executes. Here's what you can expect from GPT-5 in Cline: > verbose while planning; terse while executing > asks a lot of good clarification questions, & frequently provides options when appropriate > strong context retention and persistence over long horizons (256k context window) > good at diff-style edits and multi-file changes (we'll monitor as more usage data comes in) > quiet in Act mode -- writes code without yapping Metaprompting is another strength. We tested early with OpenAI and used GPT-5 to tune our own prompt for GPT-5. Here's a pattern we like: “Answer from your own perspective: what changes or additions would help you better follow this prompt? Here is the prompt (or snippet): [snippet]. Users have complained about X and Y. What minimal edits would you make while keeping the rest intact?” Do you need to change any of your existing patterns in Cline? No -- it's good out of the box. Give a clear goal and constraints, let it plan, then let it cook. Expect more clarifying questions than most models. Pricing: $1.25/M input tokens (+90% cache), $10/M output. Roughly half of Sonnet 4 ($3/$15). Want to try GPT-5? Use it in Cline today for pure, unfiltered inference via the OpenAI, Cline, or OpenRouter providers. (fyi -- GPT-5 one-shotted this browser DAW below on the prompt "build something impressive to show me what you're capable of")

Cline

63,496 Aufrufe • vor 11 Monaten

How GPT-5 thinks, with OpenAI VP of Research Jerry Tworek 00:00 - Intro 01:01 - What Reasoning Actually Means in AI 02:32 - Chain of Thought: Models Thinking in Words 05:25 - How Models Decide How Long to Think 07:24 - Evolution from o1 to o3 to GPT-5 11:00 - The Road to OpenAI: Growing up in Poland, Dropping out of School, Trading 20:32 - Working on Robotics and Rubik's Cube Solving 23:02 - A Day in the Life: Talking to Researchers 24:06 - How Research Priorities Are Determined 26:53 - OpenAI's Culture of Transparency 29:32 - Balancing Research with Shipping Fast 31:52 - Using OpenAI's Own Tools Daily 32:43 - Pre-Training Plus RL: The Modern AI Stack 35:10 - Reinforcement Learning 101: Training Dogs 40:17 - The Evolution of Deep Reinforcement Learning 42:09 - When GPT-4 Seemed Underwhelming at First 45:39 - How RLHF Made GPT-4 Actually Useful 48:02 - Unsupervised vs Supervised Learning 49:59 - GRPO and How DeepSeek Accelerated US Research 53:05 - What It Takes to Scale Reinforcement Learning 55:36 - Agentic AI and Long-Horizon Thinking 59:19 - Alignment as an RL Problem 1:01:11 - Winning ICPC World Finals Without Specific Training 1:05:53 - Applying RL Beyond Math and Coding 1:09:15 - The Path from Here to AGI 1:12:23 - Pure RL vs Language Models

How GPT-5 thinks, with OpenAI VP of Research Jerry Tworek 00:00 - Intro 01:01 - What Reasoning Actually Means in AI 02:32 - Chain of Thought: Models Thinking in Words 05:25 - How Models Decide How Long to Think 07:24 - Evolution from o1 to o3 to GPT-5 11:00 - The Road to OpenAI: Growing up in Poland, Dropping out of School, Trading 20:32 - Working on Robotics and Rubik's Cube Solving 23:02 - A Day in the Life: Talking to Researchers 24:06 - How Research Priorities Are Determined 26:53 - OpenAI's Culture of Transparency 29:32 - Balancing Research with Shipping Fast 31:52 - Using OpenAI's Own Tools Daily 32:43 - Pre-Training Plus RL: The Modern AI Stack 35:10 - Reinforcement Learning 101: Training Dogs 40:17 - The Evolution of Deep Reinforcement Learning 42:09 - When GPT-4 Seemed Underwhelming at First 45:39 - How RLHF Made GPT-4 Actually Useful 48:02 - Unsupervised vs Supervised Learning 49:59 - GRPO and How DeepSeek Accelerated US Research 53:05 - What It Takes to Scale Reinforcement Learning 55:36 - Agentic AI and Long-Horizon Thinking 59:19 - Alignment as an RL Problem 1:01:11 - Winning ICPC World Finals Without Specific Training 1:05:53 - Applying RL Beyond Math and Coding 1:09:15 - The Path from Here to AGI 1:12:23 - Pure RL vs Language Models

Matt Turck

451,229 Aufrufe • vor 9 Monaten

OpenAI are testing a new model on the Web Dev Arena Arena under the name 'Anonymous Chatbot 0717'. I can't believe I'm gonna say this, but it is genuinely at a completely different level of front end coding - far better than Sonnet, o3, Gemini 2.5 Pro, or Grok 4. To test it, I ran a great prompt borrowed from the amazing The Feature Crew YouTube channel, asking models to create a procedurally generated planet with Three.js. Take a look yourselves, but I'm pretty astonished by how big the jump is. I have featured the new model twice just because its implementations have been so interesting. Of course, this is only one test, but OpenAI models have always been a bit 'meh' at front-end work, and they seem to have finally overtaken everyone else on that front. We'll see when it comes out. Credit to Chetaslua for discovering the model

OpenAI are testing a new model on the Web Dev Arena Arena under the name 'Anonymous Chatbot 0717'. I can't believe I'm gonna say this, but it is genuinely at a completely different level of front end coding - far better than Sonnet, o3, Gemini 2.5 Pro, or Grok 4. To test it, I ran a great prompt borrowed from the amazing The Feature Crew YouTube channel, asking models to create a procedurally generated planet with Three.js. Take a look yourselves, but I'm pretty astonished by how big the jump is. I have featured the new model twice just because its implementations have been so interesting. Of course, this is only one test, but OpenAI models have always been a bit 'meh' at front-end work, and they seem to have finally overtaken everyone else on that front. We'll see when it comes out. Credit to Chetaslua for discovering the model

Peter Gostev

495,815 Aufrufe • vor 1 Jahr

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

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

Jim Fan

991,756 Aufrufe • vor 2 Jahren

🚨 OPENAI'S "SPUD" IS ALREADY IN THE WILD (AND IT RIVALS MYTHOS) Brad Gerstner just went on the All-In Podcast and confirmed that OpenAI's highly anticipated "Spud" model (expected to be GPT-5.5) is already being tested behind closed doors and the early reviews are insane. Addressing the recent wave of skepticism surrounding the company, Gerstner claims we are currently at "peak OpenAI FUD," warning that: "It would be seriously foolish to count out OpenAI... it starts with great researchers and great models. And I think when you see the Spud model they're about ready to release, I think it's going to be an excellent model, shows that they're firmly on the wave." When Jason Calacanis presses him on whether anyone has actually gotten their hands on the new model yet, Gerstner confirms the quiet rollout, stating that: "People are using Spud, right? So it is being previewed." When asked exactly what those early testers are saying about its capabilities, Gerstner drops a massive comparison, revealing that: "They're telling us that it's an incredible model on par with Mythos, right? And that it's a very usable model in terms of how it's packaged."

🚨 OPENAI'S "SPUD" IS ALREADY IN THE WILD (AND IT RIVALS MYTHOS) Brad Gerstner just went on the All-In Podcast and confirmed that OpenAI's highly anticipated "Spud" model (expected to be GPT-5.5) is already being tested behind closed doors and the early reviews are insane. Addressing the recent wave of skepticism surrounding the company, Gerstner claims we are currently at "peak OpenAI FUD," warning that: "It would be seriously foolish to count out OpenAI... it starts with great researchers and great models. And I think when you see the Spud model they're about ready to release, I think it's going to be an excellent model, shows that they're firmly on the wave." When Jason Calacanis presses him on whether anyone has actually gotten their hands on the new model yet, Gerstner confirms the quiet rollout, stating that: "People are using Spud, right? So it is being previewed." When asked exactly what those early testers are saying about its capabilities, Gerstner drops a massive comparison, revealing that: "They're telling us that it's an incredible model on par with Mythos, right? And that it's a very usable model in terms of how it's packaged."

Chris

138,250 Aufrufe • vor 3 Monaten

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. :-)

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

40,319 Aufrufe • vor 2 Monaten