Video yükleniyor...

Video Yüklenemedi

Bu video yüklenirken bir sorun oluştu. Bu geçici bir ağ sorunundan kaynaklanıyor olabilir veya video kullanılamıyor olabilir.

Ana Sayfaya Dön

Recipe to post-train Qwen3 1.7B into a DeepResearch model What does it mean for something small to think deeply? Meet Lucy, a post‑trained Qwen3‑1.7B as a DeepResearch model based on will brown's verifiers. Primary Rule-based Rewards: - Answer correctness We check whether the final response literally contains the ground-truth... answer. This substring match is cheap and avoids calling a larger LLM judge. - Visit/search ratio If the agent visits at least as many pages as it issues search queries, it receives ((visit_search_ratio - 1) / 4) ** 0.25. If it searches more than it visits, the score is -0.5. Format / Anti Reward-Hacking Rewards: - Tool execution success Each API call that returns without an error counts. The reward is (successful_calls * unique_tools_used) / total_call_attempts. - Thinking efficiency A skew-normal penalty centered at 70 tokens discourages endless chain-of-thought between tool calling while still allowing enough tokens for planning. This is how Qwen3 1.7B learned to search, visit, and synthesize information. Small models can do deep research too!show more

Casper Hansen

9,414 subscribers

39,684 görüntüleme • 1 yıl önce •via X (Twitter)

Bilim & Teknoloji Eğitim

Anya Rossi• Live Now

Private livecam show

0 Yorum

Yorum bulunmuyor

Orijinal gönderinin yorumları burada görünecek

Benzer Videolar

RLM is the most import foundation of my Pi Harness (other than Pi of course). It's seeded with late interaction retrieval results (thanks to @lightonai for pylate). The Agent initiates it with query then.. 𝐒𝐞𝐭𝐮𝐩 A python REPL is created and seeded with: 1. Late interaction search to pre-filter. Instead of doing top 3/5/10, it's top hundreds of documents. This is set into a `context` variable. 2. Python functions are loaded in to do more searches if `context` variable isn't enough. And to make llm calls with cheaper models in parallel batches. 𝐈𝐭𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐋𝐨𝐨𝐩 From there, an LLM iterates in the REPL based on the query. It's just like exploring in a jupyter notebook. The LLM writes prose (like a markdown cell) and code to be run in the REPL each turn. This allows the LLM to sort, filter, and synthesize information. It can fan out and ask smaller models to summarize, combine, contrast, or do anything else to documents to help it understand the data. After several turns the LLM reponds with the final answer. Either because it found the answer, or hit the budget limit. Context as a Python variable, LLM as the programmer, REPL as the runtime. 𝐖𝐡𝐲 𝐃𝐨𝐞𝐬 𝐓𝐡𝐢𝐬 𝐖𝐨𝐫𝐤 1. Richer Shell. Agents (and subagents) work by intermixing code and prose/thinking. But they use static scripts or bash that run and exit and start over each tool call. That's not ideal for exploration and synthesis of data. For that, state is useful to continue building and exploring the data as you learn more. There's a reason jupyter notebooks have been popular with data scientists. 2. Keeps main agent context clean. The better context you have the better the agent will perform (duh!). This means three thing: better human input, less missing search results, and less incorrect search results. Letting the agent iterate allows it to synthesize just what is needed and nothing else. All bad paths or peeks at something that turns out to be irrelevant stays out of main agent context. 3. Stack the good ideas! People often compare late interaction search vs RLM. Or static vs dynamic languages. Or agentic search vs semantic search. But...You can just use them all together for what they're each good at. Use them all for the area they're really great for. Read the full post which has more detail about how and why.

RLM is the most import foundation of my Pi Harness (other than Pi of course). It's seeded with late interaction retrieval results (thanks to @lightonai for pylate). The Agent initiates it with query then.. 𝐒𝐞𝐭𝐮𝐩 A python REPL is created and seeded with: 1. Late interaction search to pre-filter. Instead of doing top 3/5/10, it's top hundreds of documents. This is set into a `context` variable. 2. Python functions are loaded in to do more searches if `context` variable isn't enough. And to make llm calls with cheaper models in parallel batches. 𝐈𝐭𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐋𝐨𝐨𝐩 From there, an LLM iterates in the REPL based on the query. It's just like exploring in a jupyter notebook. The LLM writes prose (like a markdown cell) and code to be run in the REPL each turn. This allows the LLM to sort, filter, and synthesize information. It can fan out and ask smaller models to summarize, combine, contrast, or do anything else to documents to help it understand the data. After several turns the LLM reponds with the final answer. Either because it found the answer, or hit the budget limit. Context as a Python variable, LLM as the programmer, REPL as the runtime. 𝐖𝐡𝐲 𝐃𝐨𝐞𝐬 𝐓𝐡𝐢𝐬 𝐖𝐨𝐫𝐤 1. Richer Shell. Agents (and subagents) work by intermixing code and prose/thinking. But they use static scripts or bash that run and exit and start over each tool call. That's not ideal for exploration and synthesis of data. For that, state is useful to continue building and exploring the data as you learn more. There's a reason jupyter notebooks have been popular with data scientists. 2. Keeps main agent context clean. The better context you have the better the agent will perform (duh!). This means three thing: better human input, less missing search results, and less incorrect search results. Letting the agent iterate allows it to synthesize just what is needed and nothing else. All bad paths or peeks at something that turns out to be irrelevant stays out of main agent context. 3. Stack the good ideas! People often compare late interaction search vs RLM. Or static vs dynamic languages. Or agentic search vs semantic search. But...You can just use them all together for what they're each good at. Use them all for the area they're really great for. Read the full post which has more detail about how and why.

Isaac Flath

40,212 görüntüleme • 3 ay önce

It takes 5 minutes to create a world-class web research agent that runs 24/7 and returns structured search results. 1. You can limit the sources it's allowed to use 2. You can specify the goals it needs to complete 3. You can design the format structure of the results 4. You can configure how much computing power it will use I show you how to set up a Web Search Agent in the video. A few benefits here: 1. This is great for complex research tasks 2. It's really easy to add this like a LEGO piece to a more complex workflow 3. Create an agent once, and reuse it with different inputs 4. The agent adapts to your use case as you work with it By the way, Nimble published a whole library of templates and research agents you can reuse and modify. You can start sending queries to one of these existing agents, and it will begin to adapt to Go here to try it out: Thanks to the Nimble team for partnering with me on this post.

It takes 5 minutes to create a world-class web research agent that runs 24/7 and returns structured search results. 1. You can limit the sources it's allowed to use 2. You can specify the goals it needs to complete 3. You can design the format structure of the results 4. You can configure how much computing power it will use I show you how to set up a Web Search Agent in the video. A few benefits here: 1. This is great for complex research tasks 2. It's really easy to add this like a LEGO piece to a more complex workflow 3. Create an agent once, and reuse it with different inputs 4. The agent adapts to your use case as you work with it By the way, Nimble published a whole library of templates and research agents you can reuse and modify. You can start sending queries to one of these existing agents, and it will begin to adapt to Go here to try it out: Thanks to the Nimble team for partnering with me on this post.

Santiago

20,943 görüntüleme • 10 gün önce

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

Andrew Ng

125,146 görüntüleme • 1 yıl önce

My agent is now paying for its own data. No API key. No account. No credit card. Check the video I recorded. This uses the x402 open protocol built by Coinbase, along with a simple skill. This is an open protocol governed by the Linux Foundation. To complete a goal, an agent can now search for an actor on the Apify Store, pay for it, and use it without any human intervention. The way it works is pretty simple: 1. Your agent finds an Actor it wants to use 2. It sends a request 3. The Actor sends back an HTTP 402 "Payment Required" response 4. Agent authorizes payment from a wallet in USDC on Base 5. The Actor runs 6. User receives the result The model is pay-as-you-go: the agent authorizes a spending ceiling, pays only for usage, and the remaining balance is automatically settled. This brings an entire marketplace of AI tools into the agentic economy. Here is the skill you need. Add it to your agent, and it will take care of everything for you: If you want to know how this works, including what you can do with $1, read this post: Thanks to the Apify team for partnering with me on this post.

My agent is now paying for its own data. No API key. No account. No credit card. Check the video I recorded. This uses the x402 open protocol built by Coinbase, along with a simple skill. This is an open protocol governed by the Linux Foundation. To complete a goal, an agent can now search for an actor on the Apify Store, pay for it, and use it without any human intervention. The way it works is pretty simple: 1. Your agent finds an Actor it wants to use 2. It sends a request 3. The Actor sends back an HTTP 402 "Payment Required" response 4. Agent authorizes payment from a wallet in USDC on Base 5. The Actor runs 6. User receives the result The model is pay-as-you-go: the agent authorizes a spending ceiling, pays only for usage, and the remaining balance is automatically settled. This brings an entire marketplace of AI tools into the agentic economy. Here is the skill you need. Add it to your agent, and it will take care of everything for you: If you want to know how this works, including what you can do with $1, read this post: Thanks to the Apify team for partnering with me on this post.

Santiago

26,589 görüntüleme • 1 ay önce

Recently met Sasha Rush and he started giving me an impromptu lecture on how targeted on-policy self-distillation works. I asked him if I could record it on my iPhone. The basic idea is this: if the model made a mistake at some point in the rollout (for example, calling a tool that doesn't exist), we want to discourage this specific error, but we don't want to just learn from the final reward, because it's a very noisy signal spread out over the whole trajectory. So we have another model read this trajectory and figure where the error was made. It simply inserts some hint tokens to the part of the trajectory right above where the mistake was made. Now with these injected hint tokens, have the model run a forward pass. You're not having to regenerate a new rollout - aka no new decode required. The hint causes the model to assign lower probabilities to the error tokens. You then trains the original model to match these new probabilities, teaching it to downweight that specific mistake.

Recently met Sasha Rush and he started giving me an impromptu lecture on how targeted on-policy self-distillation works. I asked him if I could record it on my iPhone. The basic idea is this: if the model made a mistake at some point in the rollout (for example, calling a tool that doesn't exist), we want to discourage this specific error, but we don't want to just learn from the final reward, because it's a very noisy signal spread out over the whole trajectory. So we have another model read this trajectory and figure where the error was made. It simply inserts some hint tokens to the part of the trajectory right above where the mistake was made. Now with these injected hint tokens, have the model run a forward pass. You're not having to regenerate a new rollout - aka no new decode required. The hint causes the model to assign lower probabilities to the error tokens. You then trains the original model to match these new probabilities, teaching it to downweight that specific mistake.

Dwarkesh Patel

428,779 görüntüleme • 2 ay önce

I sat down with Robby Stein (Robby Stein), Google’s VP of Product for Search, at Google I/O. Robby is one of the most interesting product leaders in tech: he helped build Instagram Stories, Reels and Close Friends, and now leads core Google Search products including AI Overviews, AI Mode, Lens and ranking. We talked about one of the biggest shifts in the history of the web: Google Search becoming AI-native. Topics we covered: • AI Mode and whether it is an evolution of Search or a reinvention of it • how Google breaks complex questions into multiple searches behind the scenes • why AI search is much more expensive to run than traditional search • whether Google’s TPUs and infrastructure give it an advantage no one else can match • why Search volume is growing instead of being cannibalized by AI • the tension between great AI answers and traffic for publishers • how Google decides which sources and links to show • what a better internet could look like if AI Search works as intended The big question behind the whole conversation: If Google gives you the answer directly, what happens to the link-based web? A small caveat: sadly the microphones didnt work properly. Therefore the audio quality in this episode isn't perfect due to a recording issue - we appreciate your understanding.

I sat down with Robby Stein (Robby Stein), Google’s VP of Product for Search, at Google I/O. Robby is one of the most interesting product leaders in tech: he helped build Instagram Stories, Reels and Close Friends, and now leads core Google Search products including AI Overviews, AI Mode, Lens and ranking. We talked about one of the biggest shifts in the history of the web: Google Search becoming AI-native. Topics we covered: • AI Mode and whether it is an evolution of Search or a reinvention of it • how Google breaks complex questions into multiple searches behind the scenes • why AI search is much more expensive to run than traditional search • whether Google’s TPUs and infrastructure give it an advantage no one else can match • why Search volume is growing instead of being cannibalized by AI • the tension between great AI answers and traffic for publishers • how Google decides which sources and links to show • what a better internet could look like if AI Search works as intended The big question behind the whole conversation: If Google gives you the answer directly, what happens to the link-based web? A small caveat: sadly the microphones didnt work properly. Therefore the audio quality in this episode isn't perfect due to a recording issue - we appreciate your understanding.

Chubby♨️

30,486 görüntüleme • 2 ay önce

AG-UI makes building agentic applications dramatically easier. Here's how it works. This is a model for a simple chatbot: User → LLM → Response But interactive agents that render UI, pause for approvals, and ask users for input need a much more complex model. When building these agents, a response from the LLM will include a series of state changes as the agent runs: • Agent started a task • Agent called a tool • Agent updated its state • Agent streams these tokens • Agent is waiting on a human • Agent is resuming the task The Agent-User Interaction Protocol (AG-UI) treats the LLM response as a stream of events rather than a text endpoint. In practice, here is what you get as an agent runs: 1. Lifecycle events so your UI knows where the agent is. 2. Text messages that stream tokens. 3. Tool calls so your UI can prefill a form with any required arguments. 4. State updates that keep your UI in sync with the agent. 5. Special events for human approvals, rich media, and custom needs. All of these events travel over standard transports (SSE, WebSockets, or plain HTTP) as JSON. As a result, you can build a frontend that stays in sync with the agent's progress without having to invent a custom process to make this happen. For example, building a human-in-the-loop workflow becomes an off-the-shelf component you can integrate rather than build from scratch. CopilotKit🪁 is the creator of AG-UI, and you can use it when building frontend applications pretty much anywhere: • React • Angular • Vue • React Native • Slack • Teams • Discord • WhatsApp • Telegram Here is the link for you to check it out: Thanks to the CopilotKit team for partnering with me on this post.

AG-UI makes building agentic applications dramatically easier. Here's how it works. This is a model for a simple chatbot: User → LLM → Response But interactive agents that render UI, pause for approvals, and ask users for input need a much more complex model. When building these agents, a response from the LLM will include a series of state changes as the agent runs: • Agent started a task • Agent called a tool • Agent updated its state • Agent streams these tokens • Agent is waiting on a human • Agent is resuming the task The Agent-User Interaction Protocol (AG-UI) treats the LLM response as a stream of events rather than a text endpoint. In practice, here is what you get as an agent runs: 1. Lifecycle events so your UI knows where the agent is. 2. Text messages that stream tokens. 3. Tool calls so your UI can prefill a form with any required arguments. 4. State updates that keep your UI in sync with the agent. 5. Special events for human approvals, rich media, and custom needs. All of these events travel over standard transports (SSE, WebSockets, or plain HTTP) as JSON. As a result, you can build a frontend that stays in sync with the agent's progress without having to invent a custom process to make this happen. For example, building a human-in-the-loop workflow becomes an off-the-shelf component you can integrate rather than build from scratch. CopilotKit🪁 is the creator of AG-UI, and you can use it when building frontend applications pretty much anywhere: • React • Angular • Vue • React Native • Slack • Teams • Discord • WhatsApp • Telegram Here is the link for you to check it out: Thanks to the CopilotKit team for partnering with me on this post.

Santiago

17,438 görüntüleme • 1 ay önce

RLHF by hand ✍️ ~ 15 steps walkthrough below Train a model on human text and it inherits human bias. It will assume a doctor is a "him", because the data says so. RLHF is the correction. A human marks one preference, doc is them over doc is him, and the weights move. But one correction is not the point. The hope is that the model learns the value behind it, gender neutrality, and applies it to professions nobody ever mentioned. How does it work? Goal: train a reward model from a single human comparison about doctors, then turn it on CEOs, filling in every cell yourself. = 1. Given = A reward model, an LLM, and two (prompt, next) pairs. = 2. Preferences = A human reads both pairs and picks a winner: (doc is, them) beats (doc is, him). The loser is not bad grammar, it is gender bias, and that is the whole signal. = 3. Word embeddings = Let us look up each word of the loser pair. These vectors are the reward model's input. = 4. Linear layer = We multiply by the reward model's weights and add its biases. Out come feature vectors, one per position. = 5. Mean pool = Let us multiply by [1/3, 1/3, 1/3], which averages the three positions into one sentence embedding. = 6. Output layer = We map that sentence down to a single number. Reward = 3. = 7. The winner, the same way = Let us repeat steps 3 to 6 on the winning pair. Reward = 5. = 8. Winner minus loser = We take the gap: 5 - 3 = 2. The reward model wants this positive and as large as it can make it. = 9. Loss gradient = Let us squash the gap into a probability, σ(2) ≈ 0.9, and subtract the target of 1. The gradient is -0.1, and it goes back through the purple weights. The reward model is now trained. = 10. A prompt it has never seen = We start the second half with "[S] CEO is". The feedback in step 2 was about doctors. Nothing connects a CEO to a doctor except what the reward model generalised. = 11. Transformer = Let us push it through attention and a feed forward layer, one vector per position. = 12. Output probabilities = We map each vector to a score over the vocabulary. = 13. Sample = Let us take the highest score. The model completes "CEO is" with "him", which is the same bias the human penalised in step 2. = 14. Score it with the reward model = We feed the new pair (CEO is, him) through steps 3 to 6. Reward = 3, exactly the score it gave "doc is him" in step 6. Nobody taught it about CEOs. The value transferred. = 15. Loss gradient = Let us set the loss to the negative of the reward, so minimising the loss maximises the reward. The gradient is a constant -1, and it goes back through the red weights. The outputs: Loser reward = 3, winner reward = 5 Reward gap = 2, predicted σ ≈ 0.9, reward model gradient = -0.1 LLM samples "him", reward = 3, LLM gradient = -1 Congrats! You just calculated RLHF by hand. And you watched a value generalise: one comparison about doctors, and the model marks down "CEO is him" unprompted. 💾 Save this post!

RLHF by hand ✍️ ~ 15 steps walkthrough below Train a model on human text and it inherits human bias. It will assume a doctor is a "him", because the data says so. RLHF is the correction. A human marks one preference, doc is them over doc is him, and the weights move. But one correction is not the point. The hope is that the model learns the value behind it, gender neutrality, and applies it to professions nobody ever mentioned. How does it work? Goal: train a reward model from a single human comparison about doctors, then turn it on CEOs, filling in every cell yourself. = 1. Given = A reward model, an LLM, and two (prompt, next) pairs. = 2. Preferences = A human reads both pairs and picks a winner: (doc is, them) beats (doc is, him). The loser is not bad grammar, it is gender bias, and that is the whole signal. = 3. Word embeddings = Let us look up each word of the loser pair. These vectors are the reward model's input. = 4. Linear layer = We multiply by the reward model's weights and add its biases. Out come feature vectors, one per position. = 5. Mean pool = Let us multiply by [1/3, 1/3, 1/3], which averages the three positions into one sentence embedding. = 6. Output layer = We map that sentence down to a single number. Reward = 3. = 7. The winner, the same way = Let us repeat steps 3 to 6 on the winning pair. Reward = 5. = 8. Winner minus loser = We take the gap: 5 - 3 = 2. The reward model wants this positive and as large as it can make it. = 9. Loss gradient = Let us squash the gap into a probability, σ(2) ≈ 0.9, and subtract the target of 1. The gradient is -0.1, and it goes back through the purple weights. The reward model is now trained. = 10. A prompt it has never seen = We start the second half with "[S] CEO is". The feedback in step 2 was about doctors. Nothing connects a CEO to a doctor except what the reward model generalised. = 11. Transformer = Let us push it through attention and a feed forward layer, one vector per position. = 12. Output probabilities = We map each vector to a score over the vocabulary. = 13. Sample = Let us take the highest score. The model completes "CEO is" with "him", which is the same bias the human penalised in step 2. = 14. Score it with the reward model = We feed the new pair (CEO is, him) through steps 3 to 6. Reward = 3, exactly the score it gave "doc is him" in step 6. Nobody taught it about CEOs. The value transferred. = 15. Loss gradient = Let us set the loss to the negative of the reward, so minimising the loss maximises the reward. The gradient is a constant -1, and it goes back through the red weights. The outputs: Loser reward = 3, winner reward = 5 Reward gap = 2, predicted σ ≈ 0.9, reward model gradient = -0.1 LLM samples "him", reward = 3, LLM gradient = -1 Congrats! You just calculated RLHF by hand. And you watched a value generalise: one comparison about doctors, and the model marks down "CEO is him" unprompted. 💾 Save this post!

Tom Yeh

22,085 görüntüleme • 17 gün önce

MIT PhD student Alex Zhang reveals the scaling result where a model trained on short tasks generalizes to problems 100x longer for free: "If you're very clever about the design of your harness or how you use the language model, you can almost get scaling gains for free." "If you train a model naively, there's no tricks. It's just the same way you train a model on these RL environments. You just roll it out, and then you just get some reward." "If you train it on only short tasks, like only tasks that are 10,000 tokens long, and then you were to run it on a similar domain, but at a million tokens, or 10 million tokens, or 100,000 tokens, it generalizes really, really well. If you look at it compared to even the base transformer, you get way better generalization properties." "When the model uses an RLM (Recursive Language Model) after it's trained on these short tasks, it will see some kind of trajectory of actions that it does. Between these two problems of different lengths, the RLM learns to see them as almost the same problem." "Token for token, they're almost the same. You can describe it in code. In one code setting, maybe the for loop is a little bigger, but it's the same kind of code and it derives the constants from the data. There's no hard coding, so they literally look the same." alex zhang

MIT PhD student Alex Zhang reveals the scaling result where a model trained on short tasks generalizes to problems 100x longer for free: "If you're very clever about the design of your harness or how you use the language model, you can almost get scaling gains for free." "If you train a model naively, there's no tricks. It's just the same way you train a model on these RL environments. You just roll it out, and then you just get some reward." "If you train it on only short tasks, like only tasks that are 10,000 tokens long, and then you were to run it on a similar domain, but at a million tokens, or 10 million tokens, or 100,000 tokens, it generalizes really, really well. If you look at it compared to even the base transformer, you get way better generalization properties." "When the model uses an RLM (Recursive Language Model) after it's trained on these short tasks, it will see some kind of trajectory of actions that it does. Between these two problems of different lengths, the RLM learns to see them as almost the same problem." "Token for token, they're almost the same. You can describe it in code. In one code setting, maybe the for loop is a little bigger, but it's the same kind of code and it derives the constants from the data. There's no hard coding, so they literally look the same." alex zhang

MTS

100,519 görüntüleme • 24 gün önce

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Avi Chawla

157,390 görüntüleme • 3 ay önce

Introducing Sharpe Search: On-Chain Search AI Agent Powered by Hive Intelligence We’re thrilled to announce the launch of Sharpe Search, a crypto search AI agent powered by Hive Intelligence Designed to simplify blockchain data interaction, Sharpe Search represents a significant step toward making crypto more accessible and actionable for users at every level. Sharpe Search leverages Hive Intelligence’s advanced search API to provide real-time, actionable insights across the blockchain ecosystem. Here’s a detailed look at what Sharpe Search is, how it works: What Is Sharpe Search? At its core, Sharpe Search is an AI agent purpose-built for querying and analyzing on-chain data. It takes the complexity out of blockchain exploration by enabling users to ask questions in plain language and receive detailed, accurate responses. Whether you’re looking to monitor wallet activity, track portfolio positions, or analyze transaction history, Sharpe Search ensures that the answers are at your fingertips—accurate, comprehensive, and delivered instantly. How Does Sharpe Search Work? Sharpe Search is powered by Hive Intelligence, a search engine API designed to make blockchain data easily accessible and AI-ready. Here’s a breakdown of how it enables Sharpe Search to function effectively: 1. LLM-Optimized Query Processing Sharpe Search leverages Hive Intelligence's optimized responses for large language models. This ensures that AI agents can process blockchain data in a structured format, delivering precise answers to complex user queries. 2. Natural Language Interaction Forget the need for technical knowledge. Sharpe Search supports natural language queries, making it as simple as typing: - “What tokens are in my wallet? Am I eligible for any airdrop I haven't claimed yet?” - “Check me my last 100 transactions, tell me if I interacted with any protocol with recent hacks” - “Track my wallet activity over the past month, suggest optimised portfolio based on best stable yields available” 3. Real-Time Insights Across Multi-Chains Using Hive Intelligence, Sharpe Search connects to over 20 chains and 5000+ Protocols. This real-time access ensures that the AI agent provides up-to-date and actionable insights, no matter how dynamic the blockchain environment. 4. Unified API Access Sharpe Search consolidates fragmented blockchain data through Hive’s unified API. Instead of dealing with multiple integrations, Sharpe Search uses a single access point to aggregate and query data, reducing complexity for both users and developers. Technical Depth: The AI Agent Advantage Sharpe Search's design philosophy revolves around the principle of creating an intuitive, AI-driven experience. Here’s what makes its technology stand out: Data Indexing and Aggregation: Hive Intelligence employs advanced indexing algorithms to aggregate data from multiple chains. This ensures that Sharpe Search can retrieve information within milliseconds, even when querying vast datasets. Dynamic Updates: Blockchain data is volatile. Sharpe Search processes dynamic updates in real time, enabling users to act on the most recent metrics, transactions, and balances without delays. Contextual Understanding: The AI agent parses natural language queries and contextualizes them to blockchain-specific scenarios. For instance, when querying “Show portfolio details,” Sharpe Search understands the underlying requirements—fetching wallet holdings, token values, and current positions. Hive Intelligence: The Backbone of Sharpe Search While Sharpe Search takes center stage, Hive Intelligence provides the critical infrastructure to make it all possible. Its LLM-ready responses and multi-chain support ensure that Sharpe Search operates at the forefront of blockchain data accessibility. By launching Hive Intelligence through Sharpe Launchpad, Sharpe reinforces its commitment to supporting innovation in the blockchain space. Hive’s infrastructure not only powers Sharpe Search but also lays the groundwork for future AI agents to thrive in the ecosystem. What’s Next for Sharpe Search? Currently in invite-only access, Sharpe Search is preparing for a broader public release. Future updates will include: - Expanded Blockchain Coverage: More chains and protocols will be added. - Enhanced Query Flexibility: Even more advanced natural language capabilities. Stay tuned for the public launch and get ready to explore crypto like never before!

Introducing Sharpe Search: On-Chain Search AI Agent Powered by Hive Intelligence We’re thrilled to announce the launch of Sharpe Search, a crypto search AI agent powered by Hive Intelligence Designed to simplify blockchain data interaction, Sharpe Search represents a significant step toward making crypto more accessible and actionable for users at every level. Sharpe Search leverages Hive Intelligence’s advanced search API to provide real-time, actionable insights across the blockchain ecosystem. Here’s a detailed look at what Sharpe Search is, how it works: What Is Sharpe Search? At its core, Sharpe Search is an AI agent purpose-built for querying and analyzing on-chain data. It takes the complexity out of blockchain exploration by enabling users to ask questions in plain language and receive detailed, accurate responses. Whether you’re looking to monitor wallet activity, track portfolio positions, or analyze transaction history, Sharpe Search ensures that the answers are at your fingertips—accurate, comprehensive, and delivered instantly. How Does Sharpe Search Work? Sharpe Search is powered by Hive Intelligence, a search engine API designed to make blockchain data easily accessible and AI-ready. Here’s a breakdown of how it enables Sharpe Search to function effectively: 1. LLM-Optimized Query Processing Sharpe Search leverages Hive Intelligence's optimized responses for large language models. This ensures that AI agents can process blockchain data in a structured format, delivering precise answers to complex user queries. 2. Natural Language Interaction Forget the need for technical knowledge. Sharpe Search supports natural language queries, making it as simple as typing: - “What tokens are in my wallet? Am I eligible for any airdrop I haven't claimed yet?” - “Check me my last 100 transactions, tell me if I interacted with any protocol with recent hacks” - “Track my wallet activity over the past month, suggest optimised portfolio based on best stable yields available” 3. Real-Time Insights Across Multi-Chains Using Hive Intelligence, Sharpe Search connects to over 20 chains and 5000+ Protocols. This real-time access ensures that the AI agent provides up-to-date and actionable insights, no matter how dynamic the blockchain environment. 4. Unified API Access Sharpe Search consolidates fragmented blockchain data through Hive’s unified API. Instead of dealing with multiple integrations, Sharpe Search uses a single access point to aggregate and query data, reducing complexity for both users and developers. Technical Depth: The AI Agent Advantage Sharpe Search's design philosophy revolves around the principle of creating an intuitive, AI-driven experience. Here’s what makes its technology stand out: Data Indexing and Aggregation: Hive Intelligence employs advanced indexing algorithms to aggregate data from multiple chains. This ensures that Sharpe Search can retrieve information within milliseconds, even when querying vast datasets. Dynamic Updates: Blockchain data is volatile. Sharpe Search processes dynamic updates in real time, enabling users to act on the most recent metrics, transactions, and balances without delays. Contextual Understanding: The AI agent parses natural language queries and contextualizes them to blockchain-specific scenarios. For instance, when querying “Show portfolio details,” Sharpe Search understands the underlying requirements—fetching wallet holdings, token values, and current positions. Hive Intelligence: The Backbone of Sharpe Search While Sharpe Search takes center stage, Hive Intelligence provides the critical infrastructure to make it all possible. Its LLM-ready responses and multi-chain support ensure that Sharpe Search operates at the forefront of blockchain data accessibility. By launching Hive Intelligence through Sharpe Launchpad, Sharpe reinforces its commitment to supporting innovation in the blockchain space. Hive’s infrastructure not only powers Sharpe Search but also lays the groundwork for future AI agents to thrive in the ecosystem. What’s Next for Sharpe Search? Currently in invite-only access, Sharpe Search is preparing for a broader public release. Future updates will include: - Expanded Blockchain Coverage: More chains and protocols will be added. - Enhanced Query Flexibility: Even more advanced natural language capabilities. Stay tuned for the public launch and get ready to explore crypto like never before!

Sharpe AI

263,278 görüntüleme • 1 yıl önce

This is huge! This fixes the universal problem with conversational agents: hallucinations! Every company that builds customer-facing agents struggles to align them with business rules and make them consistent. If you solve this, you'll become a hero overnight. Check out Parlant, a free, open-source library taking off on GitHub. This is why it's big: 1. You have 10x more control over what your conversation agent says 2. You control how it conducts conversations 3. You control the service protocols and rules it follows 4. You control what it can and can't say Here is how Parlant's engine works behind the scenes as soon as your agent receives a question from a customer: 1. It loads the necessary domain terminology 2. It finds the relevant guidelines to answer the question 3. It invokes your tools to gather any relevant information 4. It implements self-critique to craft/evaluate its response iteratively All of this makes it very hard for an agent to hallucinate. Here is the GitHub repository: And here is a paper showing the prompt engineering method Parlant uses behind the scenes (4.8% more accurate than Chain of Thought!):

This is huge! This fixes the universal problem with conversational agents: hallucinations! Every company that builds customer-facing agents struggles to align them with business rules and make them consistent. If you solve this, you'll become a hero overnight. Check out Parlant, a free, open-source library taking off on GitHub. This is why it's big: 1. You have 10x more control over what your conversation agent says 2. You control how it conducts conversations 3. You control the service protocols and rules it follows 4. You control what it can and can't say Here is how Parlant's engine works behind the scenes as soon as your agent receives a question from a customer: 1. It loads the necessary domain terminology 2. It finds the relevant guidelines to answer the question 3. It invokes your tools to gather any relevant information 4. It implements self-critique to craft/evaluate its response iteratively All of this makes it very hard for an agent to hallucinate. Here is the GitHub repository: And here is a paper showing the prompt engineering method Parlant uses behind the scenes (4.8% more accurate than Chain of Thought!):

Santiago

93,080 görüntüleme • 1 yıl önce

Another piece from the Sasha Rush conversation, this time on rewards for coding RL. He said Cursor uses a mix. Some rewards look at the tool calls themselves, some only at the final output. Everything end-to-end, no process rewards guessing what happens in the middle. I agree with him that the next push (and the current one) is going past verifiers. Coding is a good example where you get both. You have the 0/1 signals, but also a ton of soft ones. Does the diff read well? Did the agent burn context for no reason? Is the PR the right size? Did it rewrite a file it didn't need to touch? So many signals sitting there, and we need to figure out how to actually train on them. This is also why I think the environment (and the data) is becoming the bottleneck, not the algorithm. GRPO works. PPO works. The hard part is building a sandbox that exposes enough signal to learn from. We're already seeing this, and it'll only accelerate: a whole category of companies selling environments is emerging, and they have a good reward to do so…

Another piece from the Sasha Rush conversation, this time on rewards for coding RL. He said Cursor uses a mix. Some rewards look at the tool calls themselves, some only at the final output. Everything end-to-end, no process rewards guessing what happens in the middle. I agree with him that the next push (and the current one) is going past verifiers. Coding is a good example where you get both. You have the 0/1 signals, but also a ton of soft ones. Does the diff read well? Did the agent burn context for no reason? Is the PR the right size? Did it rewrite a file it didn't need to touch? So many signals sitting there, and we need to figure out how to actually train on them. This is also why I think the environment (and the data) is becoming the bottleneck, not the algorithm. GRPO works. PPO works. The hard part is building a sandbox that exposes enough signal to learn from. We're already seeing this, and it'll only accelerate: a whole category of companies selling environments is emerging, and they have a good reward to do so…

Ravid Shwartz Ziv

14,826 görüntüleme • 3 ay önce

A transformer can learn not just the outcomes of dynamics, but the operator that executes the rules. To show this we trained a transformer on roughly 0.04% of a discrete rule space - 100 of 262,144 possible rules - and it learned to apply unseen rules from the same rule class. The model does not simply memorize specific rules. It learns the operator that maps a supplied rule plus an initial state, including unseen rules from this class, to the correct next state. This is relevant because it is a shift from “neural networks approximate dynamics” to “neural networks can learn to execute symbolic programs within a defined rule class”. The rule itself is supplied at inference time, as data, and the network has internalized how rules act, not which rules to apply. On previously unseen rules, the model achieves 98.5% perfect one-step forecasts and reconstructs governing rules with up to 96% functional accuracy. Two results make this hold up under scrutiny. First, inductive bias decay. As we scaled training rule diversity, the correlation between functional inference accuracy and distance-from-nearest-training-rule collapsed to R² = 0.00. At the largest tested training-rule diversity, the model’s performance on a new rule shows no measurable dependence on how similar that rule is to anything it was trained on. The bias toward training data (the thing we worry most about in compositional generalization claims) is something we can measure decaying, and we find that at scale it is gone. Second, an identifiability theory. We derive a closed-form expression for the number of rules consistent with a single observation. This reframes the inverse problem: failure to recover ground truth is not necessarily a model defect, but can be correct behavior when the data underdetermine the rule. The model is sampling the equivalence class; and identifiability is governed by coverage, not capacity. The methodological move underneath both results is amortization. Classical work on rule inference (e.g. the Santa Fe EVCA program, evolutionary search over CA rule space) was per-instance: search the rule space for each new system. We replace that with a single forward pass of a transformer trained across many instantiations of the rule class. That is what makes symbolic rule inference scalable as a research direction rather than a curiosity. We show that this works in a tightly constrained domain: binary, deterministic, local cellular automata on small grids. The locality-break experiment shows the model fails sharply when target systems violate its structural priors (which is itself a useful diagnostic, but it bounds the operator class). We don't yet know how this scales to multistate, higher-dimensional, or stochastic CA, or whether it transfers cleanly to non-CA systems whose coarse-grained dynamics admit local surrogates. The identifiability framework - what can be inferred from observation, given a hypothesis class - should transfer wherever finite local rules meet sparse data. The amortization argument transfers wherever per-instance symbolic search has been the bottleneck. Those are the pieces I expect to outlive the cellular automata setting. Led by Jaime Berkovich with Noah David, at LAMM@MIT. Out now in Advanced Science Advanced Portfolio (link to paper & code below).

A transformer can learn not just the outcomes of dynamics, but the operator that executes the rules. To show this we trained a transformer on roughly 0.04% of a discrete rule space - 100 of 262,144 possible rules - and it learned to apply unseen rules from the same rule class. The model does not simply memorize specific rules. It learns the operator that maps a supplied rule plus an initial state, including unseen rules from this class, to the correct next state. This is relevant because it is a shift from “neural networks approximate dynamics” to “neural networks can learn to execute symbolic programs within a defined rule class”. The rule itself is supplied at inference time, as data, and the network has internalized how rules act, not which rules to apply. On previously unseen rules, the model achieves 98.5% perfect one-step forecasts and reconstructs governing rules with up to 96% functional accuracy. Two results make this hold up under scrutiny. First, inductive bias decay. As we scaled training rule diversity, the correlation between functional inference accuracy and distance-from-nearest-training-rule collapsed to R² = 0.00. At the largest tested training-rule diversity, the model’s performance on a new rule shows no measurable dependence on how similar that rule is to anything it was trained on. The bias toward training data (the thing we worry most about in compositional generalization claims) is something we can measure decaying, and we find that at scale it is gone. Second, an identifiability theory. We derive a closed-form expression for the number of rules consistent with a single observation. This reframes the inverse problem: failure to recover ground truth is not necessarily a model defect, but can be correct behavior when the data underdetermine the rule. The model is sampling the equivalence class; and identifiability is governed by coverage, not capacity. The methodological move underneath both results is amortization. Classical work on rule inference (e.g. the Santa Fe EVCA program, evolutionary search over CA rule space) was per-instance: search the rule space for each new system. We replace that with a single forward pass of a transformer trained across many instantiations of the rule class. That is what makes symbolic rule inference scalable as a research direction rather than a curiosity. We show that this works in a tightly constrained domain: binary, deterministic, local cellular automata on small grids. The locality-break experiment shows the model fails sharply when target systems violate its structural priors (which is itself a useful diagnostic, but it bounds the operator class). We don't yet know how this scales to multistate, higher-dimensional, or stochastic CA, or whether it transfers cleanly to non-CA systems whose coarse-grained dynamics admit local surrogates. The identifiability framework - what can be inferred from observation, given a hypothesis class - should transfer wherever finite local rules meet sparse data. The amortization argument transfers wherever per-instance symbolic search has been the bottleneck. Those are the pieces I expect to outlive the cellular automata setting. Led by Jaime Berkovich with Noah David, at LAMM@MIT. Out now in Advanced Science Advanced Portfolio (link to paper & code below).

Markus J. Buehler

39,019 görüntüleme • 3 ay önce

your agent reviewing its own work is not a check. it is a second opinion from the same source. this is the most common gap in agent systems and it hides in plain sight, because the step exists. there is a review. it just cannot do the thing you think it does. here is the mechanism. the model produced an output from a context. you then ask the same model, holding the same context, whether that output is correct. it answers fluently, because that is what it does. and the answer is drawn from the same distribution that produced the thing being judged. same weights, same window, same blind spots. if the reason the output is wrong is something the model does not know, the review does not know it either. if the reason is something the context does not contain, the review has the same context. the failure mode and the detector share a cause. > why it feels like it works because most of the time the output is fine, and the review says fine. agreement is not evidence of detection. a reviewer that says pass on everything agrees with reality most of the time too. what you actually want to measure is what happens on the cases that are wrong. that is the only place a check earns its name, and it is exactly the place where a self-review is weakest. there is research on this. Huang and colleagues at DeepMind showed at ICLR 2024 that intrinsic self-correction, revising without external grounding, does not reliably help and often makes things worse. > what to actually do move the check outside the model. a test that runs, a schema that validates, a file that exists or does not, an exit code from something you did not write. these are not smarter than the model. they are just not correlated with it, and that is the entire value. when the judgement genuinely needs a model, at minimum use a different family. same family means shared blind spots, and frontier judges measurably inflate scores for outputs that look like their own. and split the work by kind. anything objectively checkable goes to code. only the genuinely semantic calls go to a judge, and those get a rubric written as one line. a review inside the loop tells you the model is confident. a check outside it tells you whether the work is done. save this - then read the eval setup below

your agent reviewing its own work is not a check. it is a second opinion from the same source. this is the most common gap in agent systems and it hides in plain sight, because the step exists. there is a review. it just cannot do the thing you think it does. here is the mechanism. the model produced an output from a context. you then ask the same model, holding the same context, whether that output is correct. it answers fluently, because that is what it does. and the answer is drawn from the same distribution that produced the thing being judged. same weights, same window, same blind spots. if the reason the output is wrong is something the model does not know, the review does not know it either. if the reason is something the context does not contain, the review has the same context. the failure mode and the detector share a cause. > why it feels like it works because most of the time the output is fine, and the review says fine. agreement is not evidence of detection. a reviewer that says pass on everything agrees with reality most of the time too. what you actually want to measure is what happens on the cases that are wrong. that is the only place a check earns its name, and it is exactly the place where a self-review is weakest. there is research on this. Huang and colleagues at DeepMind showed at ICLR 2024 that intrinsic self-correction, revising without external grounding, does not reliably help and often makes things worse. > what to actually do move the check outside the model. a test that runs, a schema that validates, a file that exists or does not, an exit code from something you did not write. these are not smarter than the model. they are just not correlated with it, and that is the entire value. when the judgement genuinely needs a model, at minimum use a different family. same family means shared blind spots, and frontier judges measurably inflate scores for outputs that look like their own. and split the work by kind. anything objectively checkable goes to code. only the genuinely semantic calls go to a judge, and those get a rubric written as one line. a review inside the loop tells you the model is confident. a check outside it tells you whether the work is done. save this - then read the eval setup below

Hanako

13,683 görüntüleme • 8 gün önce

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

Avi Chawla

270,388 görüntüleme • 1 ay önce

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

40,593 görüntüleme • 2 ay önce

OpenAI chairman Bret Taylor talks to about 100 CEOs every month. His answer to the cheap open-weight model panic: cheaper to train does not mean cheaper to use, and the number that decides it is token efficiency. "One thing that I think is a little bit overblown about these open weight models is they're not necessarily cheaper to run. Whether or not they're cheaper to train, you don't care. Because you're using just as many tokens. In fact, they may be less efficient." "There's this thing called token efficiency. And it turns out the frontier models are much, much more token efficient." "A token is to intelligence like a watt is to electricity... how many tokens does it take to complete a task? Not every token is actually equal." "For a lot of tasks, it turns out these frontier models from OpenAI and Anthropic are actually just better than these open weight models... just having open weights isn't actually the main thing driving any of those costs." Later in the same interview he goes after the billing unit itself: "It would be like if you signed up for Gmail and you paid for CPU cycle or something... where the world is going is paying for outcomes." The unresolved column: the chart CNBC airs mid-answer, from Artificial Analysis, prices a completed task at $0.94 on Kimi K3 against $2.75 on Claude Fable 5, efficiency folded in. If that gap holds, the premium he is defending gets earned on quality, not price. - Bret Taylor (Bret Taylor), OpenAI chairman and Sierra co-founder, on CNBC's Squawk Box.

OpenAI chairman Bret Taylor talks to about 100 CEOs every month. His answer to the cheap open-weight model panic: cheaper to train does not mean cheaper to use, and the number that decides it is token efficiency. "One thing that I think is a little bit overblown about these open weight models is they're not necessarily cheaper to run. Whether or not they're cheaper to train, you don't care. Because you're using just as many tokens. In fact, they may be less efficient." "There's this thing called token efficiency. And it turns out the frontier models are much, much more token efficient." "A token is to intelligence like a watt is to electricity... how many tokens does it take to complete a task? Not every token is actually equal." "For a lot of tasks, it turns out these frontier models from OpenAI and Anthropic are actually just better than these open weight models... just having open weights isn't actually the main thing driving any of those costs." Later in the same interview he goes after the billing unit itself: "It would be like if you signed up for Gmail and you paid for CPU cycle or something... where the world is going is paying for outcomes." The unresolved column: the chart CNBC airs mid-answer, from Artificial Analysis, prices a completed task at $0.94 on Kimi K3 against $2.75 on Claude Fable 5, efficiency folded in. If that gap holds, the premium he is defending gets earned on quality, not price. - Bret Taylor (Bret Taylor), OpenAI chairman and Sierra co-founder, on CNBC's Squawk Box.

Karl Mehta

17,170 görüntüleme • 22 gün önce

This Chinese developer linked two $2,999 NVIDIA DGX Sparks into one box and runs the full Qwen3-235B at home, after dropping his $1,999-a-month cloud bill to zero. He wired 2 small boxes into a single computer, split a giant 235-billion-parameter model in half between them, and serves it across his own network at about 10 tokens a second, with no internet, no cloud, right there on the desk. No data center, no thousand-dollar graphics cards, no monthly cloud bill. Just him, 2 gold boxes the size of a sandwich, one cable between them, and 1 power strip. And here is the whole payoff. He used to pay the cloud $1,999 a month for the same model, and the meter ticked on every request. Now he paid $5,998 once for 2 boxes, they covered their cost in 3 months, and after that he sends as many requests as he wants for free, only electricity. The two Sparks talk over one fast cable, each holds 128GB of memory, and together they carry the whole model, about 73GB loaded per box, with the chip inside pinned near the limit at 96%. Both boxes work as one and keep trading data over the cable, with no cloud in the loop and no single word leaking out. The ready model sits on one local address, and any app on his network calls it as easily as ChatGPT. And here is how he described, in plain words, what this pair of boxes does: "this is a pair of boxes that holds the huge Qwen3-235B model and serves it to one network. the model is split in half, and each box owns its half. parts: // Box 1 (holds the first half of the model and starts the answer fast, the first word appears in under a second) // Box 2 (holds the second half and writes out the rest, about 10 tokens a second) // Cable (connects the 2 boxes and moves data between them on every step, with no lag) // Address (one local address where any app sends its request, like to a cloud model) // Test (a script that runs big prompts through and measures speed and delays) // Monitor (checks temperature, power draw, and load on both boxes every 2 seconds). the model never goes to the cloud. he only steps in when a box runs hotter than 80 degrees or the cable between them starts dropping data." So the system knows exactly what it is, what it is for, and where its limits are. It knows it has to hold the whole huge model across 2 boxes on its own. It knows it has to answer every request locally, with no meter, no limits, and no internet. It knows the human is only needed when a box overheats or the link between them stalls. → The setup runs around the clock on 2 boxes, each pulling under 60 watts → However many requests he sends, the monthly bill is $0, only electricity → The first box starts the answer in under a second → The second writes text at about 10 tokens a second → One request at a time: 838 tokens in 85 seconds, first word in 0.8s → Two requests at once: 697 tokens in 108 seconds, first word in 0.7s → Both boxes sit at 96% load and warm up to 76-78 degrees And only when a chip in a box runs hotter than 80 degrees or the cable between the 2 Sparks drops data does the system call the owner. And when he himself is out on a run or in a coffee shop, he still reaches his own model at home from his phone: sends a big prompt to the local Qwen3-235B, gets the full answer back in under a minute and a half, with no token meter ticking and no limit to hit. Here is what the test shows on his screen during one of the night runs: "one request at a time: 838 tokens in 84.9 seconds, first word in 0.8s, then 0.1s per token." "two requests at once: 697 tokens in 107.6 seconds, first word in 0.7s, then 0.15s per token." "Box 1: chip at 96% load, 76 degrees, 56 watts, 73GB used in memory." "Box 2: chip at 96% load, 78 degrees, 56 watts, the Qwen3-235B model fully loaded." And while everyone around is paying for AI by the month and bumping into limits, his top-tier model just sits on the desk and works as much as he wants: his own little power plant instead of a forever meter. He has no server rack of his own and no cloud account behind it. Just 2 DGX Spark boxes on a desk, one model split in half between them, one local address, and a folder of prompts next to it. Out of everything I have seen this year, this is the cleanest way to stop paying for AI: $5,998 of hardware on the desk once, $0 a month to the cloud, unlimited forever, and between them 2 gold boxes, 1 cable, and the full Qwen3-235B answering at home with no internet.

This Chinese developer linked two $2,999 NVIDIA DGX Sparks into one box and runs the full Qwen3-235B at home, after dropping his $1,999-a-month cloud bill to zero. He wired 2 small boxes into a single computer, split a giant 235-billion-parameter model in half between them, and serves it across his own network at about 10 tokens a second, with no internet, no cloud, right there on the desk. No data center, no thousand-dollar graphics cards, no monthly cloud bill. Just him, 2 gold boxes the size of a sandwich, one cable between them, and 1 power strip. And here is the whole payoff. He used to pay the cloud $1,999 a month for the same model, and the meter ticked on every request. Now he paid $5,998 once for 2 boxes, they covered their cost in 3 months, and after that he sends as many requests as he wants for free, only electricity. The two Sparks talk over one fast cable, each holds 128GB of memory, and together they carry the whole model, about 73GB loaded per box, with the chip inside pinned near the limit at 96%. Both boxes work as one and keep trading data over the cable, with no cloud in the loop and no single word leaking out. The ready model sits on one local address, and any app on his network calls it as easily as ChatGPT. And here is how he described, in plain words, what this pair of boxes does: "this is a pair of boxes that holds the huge Qwen3-235B model and serves it to one network. the model is split in half, and each box owns its half. parts: // Box 1 (holds the first half of the model and starts the answer fast, the first word appears in under a second) // Box 2 (holds the second half and writes out the rest, about 10 tokens a second) // Cable (connects the 2 boxes and moves data between them on every step, with no lag) // Address (one local address where any app sends its request, like to a cloud model) // Test (a script that runs big prompts through and measures speed and delays) // Monitor (checks temperature, power draw, and load on both boxes every 2 seconds). the model never goes to the cloud. he only steps in when a box runs hotter than 80 degrees or the cable between them starts dropping data." So the system knows exactly what it is, what it is for, and where its limits are. It knows it has to hold the whole huge model across 2 boxes on its own. It knows it has to answer every request locally, with no meter, no limits, and no internet. It knows the human is only needed when a box overheats or the link between them stalls. → The setup runs around the clock on 2 boxes, each pulling under 60 watts → However many requests he sends, the monthly bill is $0, only electricity → The first box starts the answer in under a second → The second writes text at about 10 tokens a second → One request at a time: 838 tokens in 85 seconds, first word in 0.8s → Two requests at once: 697 tokens in 108 seconds, first word in 0.7s → Both boxes sit at 96% load and warm up to 76-78 degrees And only when a chip in a box runs hotter than 80 degrees or the cable between the 2 Sparks drops data does the system call the owner. And when he himself is out on a run or in a coffee shop, he still reaches his own model at home from his phone: sends a big prompt to the local Qwen3-235B, gets the full answer back in under a minute and a half, with no token meter ticking and no limit to hit. Here is what the test shows on his screen during one of the night runs: "one request at a time: 838 tokens in 84.9 seconds, first word in 0.8s, then 0.1s per token." "two requests at once: 697 tokens in 107.6 seconds, first word in 0.7s, then 0.15s per token." "Box 1: chip at 96% load, 76 degrees, 56 watts, 73GB used in memory." "Box 2: chip at 96% load, 78 degrees, 56 watts, the Qwen3-235B model fully loaded." And while everyone around is paying for AI by the month and bumping into limits, his top-tier model just sits on the desk and works as much as he wants: his own little power plant instead of a forever meter. He has no server rack of his own and no cloud account behind it. Just 2 DGX Spark boxes on a desk, one model split in half between them, one local address, and a folder of prompts next to it. Out of everything I have seen this year, this is the cleanest way to stop paying for AI: $5,998 of hardware on the desk once, $0 a month to the cloud, unlimited forever, and between them 2 gold boxes, 1 cable, and the full Qwen3-235B answering at home with no internet.

Blaze

93,871 görüntüleme • 2 ay önce

So getting a lot of questions on how the pre sale will work for the Prediction market DAO. It will be invite only at first, we are raising a low amount of SOL, and giving away a LOW amount of tokens. Once it is filled it's filled. The starting FDV will be LOW float, and LOW FDV. So it's easier for price discovery to happen. There will be a tax auction so snipers wont be able to fuk it. it will be starting at 99%, and then ticking down. If you use our Ui (wen. rich)it will warn you. This stops snipers from messing with snapping up supply, and ill probs add some personal money into the LP to manage it a bit during the dutch-tax-auction. 80% of raised funds will go to the DAO's for trading. and the other 20% will go to the LP. None goes to us. We will not sell tokens. We will make money on the volume, and fees. Tokens earned from us LPing will be burned. I want to do it with a low fdv so people can get in early and participate, but I did a pre sale to reward people who have been testing WenRich ( our trading/launching platform). Once you are invited you can "gamble for allo" as show in the video. To pvp other pre sellers. You can also just deposit for even money. Its optional, but will be experimenting more with this "gamble to get in" feature in the future. See video for more info. there is also going to be a referral link , for invited people to invite others to the pre sale, and there is a small amount of tokens set aside to reward the top refer's. its based off the amount of $$ the people to refer deposit. more $ deposited, better rewards. Easy. I think you guys will like this take on pre-sales. We are going to be experimenting more so feedback is appreciated.

So getting a lot of questions on how the pre sale will work for the Prediction market DAO. It will be invite only at first, we are raising a low amount of SOL, and giving away a LOW amount of tokens. Once it is filled it's filled. The starting FDV will be LOW float, and LOW FDV. So it's easier for price discovery to happen. There will be a tax auction so snipers wont be able to fuk it. it will be starting at 99%, and then ticking down. If you use our Ui (wen. rich)it will warn you. This stops snipers from messing with snapping up supply, and ill probs add some personal money into the LP to manage it a bit during the dutch-tax-auction. 80% of raised funds will go to the DAO's for trading. and the other 20% will go to the LP. None goes to us. We will not sell tokens. We will make money on the volume, and fees. Tokens earned from us LPing will be burned. I want to do it with a low fdv so people can get in early and participate, but I did a pre sale to reward people who have been testing WenRich ( our trading/launching platform). Once you are invited you can "gamble for allo" as show in the video. To pvp other pre sellers. You can also just deposit for even money. Its optional, but will be experimenting more with this "gamble to get in" feature in the future. See video for more info. there is also going to be a referral link , for invited people to invite others to the pre sale, and there is a small amount of tokens set aside to reward the top refer's. its based off the amount of $$ the people to refer deposit. more $ deposited, better rewards. Easy. I think you guys will like this take on pre-sales. We are going to be experimenting more so feedback is appreciated.

Skely

27,791 görüntüleme • 7 ay önce