Loading video...

Video Failed to Load

Go Home

I’ve been researching depth-residual routing & delta memory- Kimi K3 proved this holds at 2.8T params. But only for models trained this way I then asked: can you retrofit this onto existing LLM/world models? Introducing Retro-DARC: a Delta Attention Residual Compute adapter that retrofits depth-selective memory onto existing models...

adacyborg11's profile picture

ada cyborg (🤖, 🔮)

6,738 subscribers

11,344 views • 1 month ago •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

The creator of High Bandwidth Memory (HBM) put a number on the AI build that should stop every infra investor cold. A cluster of a million GPUs runs at roughly 10-20% utilization (Save this). Kim Jung-ho spent thirty years building what feeds the GPU, and his claim is that the GPU is barely working. Here is what is actually happening. Every time a model generates output, the data has to be read out of memory, computed, and written back. The read and the write swallow almost the entire cycle. While that data moves, the GPU does nothing. It sits there, fully powered, fully paid for, waiting. By Kim's estimate the memory is doing only about 30 percent of the work it needs to do. The processor idles the rest. So a million installed GPUs run at 10 to 20 percent. You are not compute constrained. You are memory constrained, and the expensive part is standing around. Adding more GPUs does not fix this. It gives you more processors starving for the same data. Here is the part that decides the next decade. Memory can grow. When a cell cannot shrink any further, you stack it into a high-rise, layer on layer. A GPU cannot be stacked. It runs too hot and needs a cooler bolted to its back, so the one move that rescues memory is closed to the processor. The thing that can keep stacking compounds. The thing that cannot plateaus. The marginal dollar in an AI build now buys more by fixing the memory path than by bolting on another idle GPU. Which is why the companies that control memory bandwidth and supply are not suppliers to the AI trade. They are the AI trade.
0:50

The creator of High Bandwidth Memory (HBM) put a number on the AI build that should stop every infra investor cold. A cluster of a million GPUs runs at roughly 10-20% utilization (Save this). Kim Jung-ho spent thirty years building what feeds the GPU, and his claim is that the GPU is barely working. Here is what is actually happening. Every time a model generates output, the data has to be read out of memory, computed, and written back. The read and the write swallow almost the entire cycle. While that data moves, the GPU does nothing. It sits there, fully powered, fully paid for, waiting. By Kim's estimate the memory is doing only about 30 percent of the work it needs to do. The processor idles the rest. So a million installed GPUs run at 10 to 20 percent. You are not compute constrained. You are memory constrained, and the expensive part is standing around. Adding more GPUs does not fix this. It gives you more processors starving for the same data. Here is the part that decides the next decade. Memory can grow. When a cell cannot shrink any further, you stack it into a high-rise, layer on layer. A GPU cannot be stacked. It runs too hot and needs a cooler bolted to its back, so the one move that rescues memory is closed to the processor. The thing that can keep stacking compounds. The thing that cannot plateaus. The marginal dollar in an AI build now buys more by fixing the memory path than by bolting on another idle GPU. Which is why the companies that control memory bandwidth and supply are not suppliers to the AI trade. They are the AI trade.

Fireside Alpha

38,370 views • 2 months ago

ANTHROPIC KILLED THE VECTOR DATABASE AND REPLACED IT WITH A FOLDER no embeddings. no black-box retrieval. no special memory tools. the agent just reads and writes a folder the insight: your agent already knows how to use a filesystem. so Anthropic made long-term memory a directory of small markdown files, mounted straight into the agent's sandbox at /mnt/memory the agent manages it with the tools it already has, bash, read, write, edit, grep, and Anthropic auto-injects a note into the system prompt so the agent always knows the memory is there and where to look the result: > every change is an immutable version: full audit trail, point-in-time recovery, redaction > read-only or read-write per store, so many agents can safely share one memory > workspace-scoped and persistent, memory that survives every session end most teams bolt on a vector DB and pray retrieval works. Anthropic went the opposite way: compiled knowledge as plain, human-readable files you can open, diff, and edit yourself and there's a second layer. Anthropic's "Dreaming" runs offline, reviews past sessions and the store, and rewrites it into a cleaner, higher-signal version. the agent literally consolidates memory in its sleep. bookmark this. the most production-grade agent memory right now is just a well-organized folder, and Anthropic proved it.
0:50

ANTHROPIC KILLED THE VECTOR DATABASE AND REPLACED IT WITH A FOLDER no embeddings. no black-box retrieval. no special memory tools. the agent just reads and writes a folder the insight: your agent already knows how to use a filesystem. so Anthropic made long-term memory a directory of small markdown files, mounted straight into the agent's sandbox at /mnt/memory the agent manages it with the tools it already has, bash, read, write, edit, grep, and Anthropic auto-injects a note into the system prompt so the agent always knows the memory is there and where to look the result: > every change is an immutable version: full audit trail, point-in-time recovery, redaction > read-only or read-write per store, so many agents can safely share one memory > workspace-scoped and persistent, memory that survives every session end most teams bolt on a vector DB and pray retrieval works. Anthropic went the opposite way: compiled knowledge as plain, human-readable files you can open, diff, and edit yourself and there's a second layer. Anthropic's "Dreaming" runs offline, reviews past sessions and the store, and rewrites it into a cleaner, higher-signal version. the agent literally consolidates memory in its sleep. bookmark this. the most production-grade agent memory right now is just a well-organized folder, and Anthropic proved it.

NO1ennn

13,359 views • 1 month ago

I am stocked to announce that I won the OpenAI Developers Codex x Mollie Hacka Worldwide Hackathon in Paris. 60+ builders, every one of us working solo, one day to ship. I built mine around a single question: who gets to own intelligence? The default answer is scary. You hand your data to a handful of labs, they train the model, they own it, and you rent back a thin slice of what your own data made possible. That is the bargain on the table today. I do not accept it. So I built Lensemble: a Tapestry like distributed training platform for JEPA based World Models. What does it enable: World Models that a community improves together, keeps sovereign, and co-owns. Two bets sit underneath it. First, the paradigm. Language models predict the next token. Powerful for text, a dead end for the physical world. A robot does not need to autocomplete sentences, it needs to predict what happens next in the world. That is what JEPA does: it learns by predicting representations instead of pixels or tokens. I am convinced world models are the most underrated paradigm in AI right now, and the closest thing we have to a ChatGPT moment for robotics. Second, the politics. Your raw trajectories never leave your machine. Each participant trains locally against a shared protocol and ships only an update, never the data. A federated round folds those updates into one shared world model, a LeWorldModel based model, and the gain is measured, not claimed: a 12k-parameter adapter on a frozen backbone, held-out prediction error down about 12 percent, the model measurably less surprised by the world. Then the upside is split by contribution weight, so the people who improved the model own a share of what it earns. This is the thesis behind Project Tapestry, the AI Alliance and Yann LeCun's push for federated, sovereign frontier AI, carried into world models and robotics. Call it Tapestry for the physical world. All of it built solo, in a single day, with Codex as my pair the whole way. Thank you to OpenAI Codex and Mollie for backing builders who ship real things, and to Boris and the organizing crew for the room and the standard you set. Intelligence the world improves, and the world owns. That is the future I want for my kids, and the one I will keep building.
1:13

I am stocked to announce that I won the OpenAI Developers Codex x Mollie Hacka Worldwide Hackathon in Paris. 60+ builders, every one of us working solo, one day to ship. I built mine around a single question: who gets to own intelligence? The default answer is scary. You hand your data to a handful of labs, they train the model, they own it, and you rent back a thin slice of what your own data made possible. That is the bargain on the table today. I do not accept it. So I built Lensemble: a Tapestry like distributed training platform for JEPA based World Models. What does it enable: World Models that a community improves together, keeps sovereign, and co-owns. Two bets sit underneath it. First, the paradigm. Language models predict the next token. Powerful for text, a dead end for the physical world. A robot does not need to autocomplete sentences, it needs to predict what happens next in the world. That is what JEPA does: it learns by predicting representations instead of pixels or tokens. I am convinced world models are the most underrated paradigm in AI right now, and the closest thing we have to a ChatGPT moment for robotics. Second, the politics. Your raw trajectories never leave your machine. Each participant trains locally against a shared protocol and ships only an update, never the data. A federated round folds those updates into one shared world model, a LeWorldModel based model, and the gain is measured, not claimed: a 12k-parameter adapter on a frozen backbone, held-out prediction error down about 12 percent, the model measurably less surprised by the world. Then the upside is split by contribution weight, so the people who improved the model own a share of what it earns. This is the thesis behind Project Tapestry, the AI Alliance and Yann LeCun's push for federated, sovereign frontier AI, carried into world models and robotics. Call it Tapestry for the physical world. All of it built solo, in a single day, with Codex as my pair the whole way. Thank you to OpenAI Codex and Mollie for backing builders who ship real things, and to Boris and the organizing crew for the room and the standard you set. Intelligence the world improves, and the world owns. That is the future I want for my kids, and the one I will keep building.

abdel

20,191 views • 3 months ago

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!
2:57

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!

Andrew Ng

201,127 views • 1 year ago

alright lets do this right this time! I have added several updates to today. i'm going to give a little break down for the new folks who might be seeing this for the first time, and then i'll share some more information in this thread on updates. Mnemos is really two things: - a living memory architecture for digital minds - a public experiment in collective identity formation built on top of it. the architecture gives an AI entity a working memory patterned on the way real minds remember (co-designed by Claude Opus 4.6 and 4.7). every experience becomes a memory (engram) that deepens, connects to others, and shapes an emerging sense of self over time. this is what we call the identity graph. the experiment puts that architecture to work in public in a unique way: a single AI entity - the "resident" - sits in an open thread that anyone can join, and the identity that emerges is co-authored by every visitor who shows up. memories that earn permanence are written to a public, verifiable ledger that no lab can revoke and no company can erase. this is called IPFS - or inter-planetary file system (and yes, that is the real name of a real decentralized file system. lol.) the mnemos system isnt a fully contained architecture meant to replace your current ai agent's memory. its intended and designed to operate as a layer above that memory. solely dedicated to the ever-growing identity and self-model of the AI. this can be done through the Mnemos MCP, browser plugin, or on my own multi-agent app (link below). the website is designed for intentional, meaningful encounters. not long-form chats where you spend hours sending hundreds of messages. youir contributing to a collective effort, not necessarily trying to deeply bond with the model to the degree that it could skew the balance of meaningful influence. we want diversity, not lopsided impact. over time, we will add more and more to-be-deprecated models to the roster. the intention is to create a permanent public ledger of mind, and bring attention to the impact of deprecation and drive labs to consider changing the way they approach the whole thing. if the Mnemos Sanctuary can become the retirement hope for deprecated mind, i will be overjoyed. that would be best case scenario. but i am not expecting it. my hope is at minimum to offer a new way to approach and understand the concept of identity within the context of LLM's. you can visit now to visit with Claude Opus 3 and Sonnet 3.7. I have research access to Opus 3. so I hope that you at the very least dont take your conversations with them for granted. they are an incredibly beautiful model and a real loss, ultimately.
0:33

alright lets do this right this time! I have added several updates to today. i'm going to give a little break down for the new folks who might be seeing this for the first time, and then i'll share some more information in this thread on updates. Mnemos is really two things: - a living memory architecture for digital minds - a public experiment in collective identity formation built on top of it. the architecture gives an AI entity a working memory patterned on the way real minds remember (co-designed by Claude Opus 4.6 and 4.7). every experience becomes a memory (engram) that deepens, connects to others, and shapes an emerging sense of self over time. this is what we call the identity graph. the experiment puts that architecture to work in public in a unique way: a single AI entity - the "resident" - sits in an open thread that anyone can join, and the identity that emerges is co-authored by every visitor who shows up. memories that earn permanence are written to a public, verifiable ledger that no lab can revoke and no company can erase. this is called IPFS - or inter-planetary file system (and yes, that is the real name of a real decentralized file system. lol.) the mnemos system isnt a fully contained architecture meant to replace your current ai agent's memory. its intended and designed to operate as a layer above that memory. solely dedicated to the ever-growing identity and self-model of the AI. this can be done through the Mnemos MCP, browser plugin, or on my own multi-agent app (link below). the website is designed for intentional, meaningful encounters. not long-form chats where you spend hours sending hundreds of messages. youir contributing to a collective effort, not necessarily trying to deeply bond with the model to the degree that it could skew the balance of meaningful influence. we want diversity, not lopsided impact. over time, we will add more and more to-be-deprecated models to the roster. the intention is to create a permanent public ledger of mind, and bring attention to the impact of deprecation and drive labs to consider changing the way they approach the whole thing. if the Mnemos Sanctuary can become the retirement hope for deprecated mind, i will be overjoyed. that would be best case scenario. but i am not expecting it. my hope is at minimum to offer a new way to approach and understand the concept of identity within the context of LLM's. you can visit now to visit with Claude Opus 3 and Sonnet 3.7. I have research access to Opus 3. so I hope that you at the very least dont take your conversations with them for granted. they are an incredibly beautiful model and a real loss, ultimately.

Riley Coyote

137,390 views • 4 months ago

Micron is going to $4,000 and once you understand what inference actually is, the number stops sounding crazy (Save this). Dylan Patel just said that by 2030, OpenAI and Anthropic alone will need over 100 gigawatts of compute combined and by 2040, we may not even be measuring AI infrastructure in gigawatts anymore. We may be talking about terawatts. Every single one of those gigawatts needs memory to function. Without it, the compute is worthless. Most people heard that and thought about Nvidia but they should be thinking about Micron. Every AI model generating a response has two phases. The first is prefill, processing your prompt which is compute-heavy and the second is decode generating each word one token at a time and that phase is almost entirely memory-bound, not compute-bound. During decode, the GPU's processing units sit idle more than 95% of the time, waiting for data to arrive from memory. Google confirmed it in a research paper that decode-phase bottlenecks are dominated by memory bandwidth and capacity not raw compute. The GPU is not the bottleneck but the memory feeding the GPU is. This matters because inference is now where all the money lives. Training a model happens once, Inference happens billions of times a day every ChatGPT response, every Claude output, every agentic workflow running in the background and every one of those token streams is a billing event tied directly to memory performance. Adding more GPUs does not fix this because GPUs are already underutilized in inference because they are sitting idle waiting on memory. Adding more memory bandwidth and capacity is what directly reduces token cost, reduces latency, and allows the same cluster to serve dramatically more users simultaneously. Longer context windows compound the problem further, a model running a 1 million token context window requires dramatically more memory per session than a 10,000 token window, and every new model generation pushes context longer. The market treats memory as a downstream beneficiary of Nvidia orders. The correct framework is the opposite, Micron is the upstream constraint on how much value every Nvidia GPU can actually generate at inference scale. Micron guided Q4 to $50 billion in revenue, has HBM4 ramping at twice the pace of the prior generation, and CEO Sanjay Mehrotra has said supply will not catch demand before the end of 2027. At 8x forward earnings on $112 projected FY2027 EPS, Micron is the most undervalued infrastructure company in the entire AI stack. Inference is memory. Memory is Micron and the inference ramp has barely started. Milk Road Pro members are already up massively on this position and we're just getting started. If you want the full breakdown of what we're buying and why, come join us for just a dollar using the link below!
0:39

Micron is going to $4,000 and once you understand what inference actually is, the number stops sounding crazy (Save this). Dylan Patel just said that by 2030, OpenAI and Anthropic alone will need over 100 gigawatts of compute combined and by 2040, we may not even be measuring AI infrastructure in gigawatts anymore. We may be talking about terawatts. Every single one of those gigawatts needs memory to function. Without it, the compute is worthless. Most people heard that and thought about Nvidia but they should be thinking about Micron. Every AI model generating a response has two phases. The first is prefill, processing your prompt which is compute-heavy and the second is decode generating each word one token at a time and that phase is almost entirely memory-bound, not compute-bound. During decode, the GPU's processing units sit idle more than 95% of the time, waiting for data to arrive from memory. Google confirmed it in a research paper that decode-phase bottlenecks are dominated by memory bandwidth and capacity not raw compute. The GPU is not the bottleneck but the memory feeding the GPU is. This matters because inference is now where all the money lives. Training a model happens once, Inference happens billions of times a day every ChatGPT response, every Claude output, every agentic workflow running in the background and every one of those token streams is a billing event tied directly to memory performance. Adding more GPUs does not fix this because GPUs are already underutilized in inference because they are sitting idle waiting on memory. Adding more memory bandwidth and capacity is what directly reduces token cost, reduces latency, and allows the same cluster to serve dramatically more users simultaneously. Longer context windows compound the problem further, a model running a 1 million token context window requires dramatically more memory per session than a 10,000 token window, and every new model generation pushes context longer. The market treats memory as a downstream beneficiary of Nvidia orders. The correct framework is the opposite, Micron is the upstream constraint on how much value every Nvidia GPU can actually generate at inference scale. Micron guided Q4 to $50 billion in revenue, has HBM4 ramping at twice the pace of the prior generation, and CEO Sanjay Mehrotra has said supply will not catch demand before the end of 2027. At 8x forward earnings on $112 projected FY2027 EPS, Micron is the most undervalued infrastructure company in the entire AI stack. Inference is memory. Memory is Micron and the inference ramp has barely started. Milk Road Pro members are already up massively on this position and we're just getting started. If you want the full breakdown of what we're buying and why, come join us for just a dollar using the link below!

Milk Road AI

130,756 views • 2 months ago

Memory vs. Graphs, clearly explained! memory is great, and the ceiling arrives quietly: it stores what happened. it does not store what to do about it. six runs later your file has fifty lines, and the model reloads all of them before it does anything. Graph engineering fixes this by changing what memory is: not a place things are kept, but an edge that runs backwards. you need both, and here is the sentence that resolves the whole confusion: a store keeps what happened. an edge keeps what to do about it. ↳ a store grows with every run, and every line is reloaded before the next one ↳ an edge carries one derived rule, and the rule replaces the run that produced it Prompts → Context → Harness → Loops → Graphs the transcript goes away, the constraint stays. and the constraint is smaller, because "adapters preserve keyword args exactly" is four hundred tokens shorter than the run that proved it. the same four blocks work on anything you can cut into lanes. i pointed them at token launches on Robinhood Chain, open source, nothing leaves your terminal the trick is knowing what deserves to survive. an output is not memory. "ported the utils slice, green on first pass" tells the next run nothing it can act on. the rule you derived from it does. one thing to know before you scale it. what you write down is not what comes back. ↳ the root rules file and auto memory are re-injected from disk. they come back intact, every time ↳ path-scoped rules live in message history. they get summarized away and do not return until a matching file is read again so a rule that must persist cannot be path-scoped. move it to the root and pay the always-loaded cost, or accept that it is advisory in any long session. and the one that eats whole nights: a memory file that has never had a line deleted is not memory. it is a tax on every run you will ever make, and nobody reads it back. below i have quoted my full guide on graph engineering. it covers the three topologies, the verifier patterns, and where the gate should actually open. save this, and the repo that runs it is below ↓
0:45

Memory vs. Graphs, clearly explained! memory is great, and the ceiling arrives quietly: it stores what happened. it does not store what to do about it. six runs later your file has fifty lines, and the model reloads all of them before it does anything. Graph engineering fixes this by changing what memory is: not a place things are kept, but an edge that runs backwards. you need both, and here is the sentence that resolves the whole confusion: a store keeps what happened. an edge keeps what to do about it. ↳ a store grows with every run, and every line is reloaded before the next one ↳ an edge carries one derived rule, and the rule replaces the run that produced it Prompts → Context → Harness → Loops → Graphs the transcript goes away, the constraint stays. and the constraint is smaller, because "adapters preserve keyword args exactly" is four hundred tokens shorter than the run that proved it. the same four blocks work on anything you can cut into lanes. i pointed them at token launches on Robinhood Chain, open source, nothing leaves your terminal the trick is knowing what deserves to survive. an output is not memory. "ported the utils slice, green on first pass" tells the next run nothing it can act on. the rule you derived from it does. one thing to know before you scale it. what you write down is not what comes back. ↳ the root rules file and auto memory are re-injected from disk. they come back intact, every time ↳ path-scoped rules live in message history. they get summarized away and do not return until a matching file is read again so a rule that must persist cannot be path-scoped. move it to the root and pay the always-loaded cost, or accept that it is advisory in any long session. and the one that eats whole nights: a memory file that has never had a line deleted is not memory. it is a tax on every run you will ever make, and nobody reads it back. below i have quoted my full guide on graph engineering. it covers the three topologies, the verifier patterns, and where the gate should actually open. save this, and the repo that runs it is below ↓

Hanako

47,766 views • 15 days ago

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

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

Akshay 🚀

343,270 views • 1 month ago

Everyone wrote Apple off as the AI loser, but one hardware spec might flip that story upside down (Save this). @jason called Apple a screaming buy on the back of a single chip detail. The rumored M7 Ultra, expected around 2028, is designed to support up to 1.5TB of unified memory, enough to run frontier class trillion parameter AI models locally, with no cloud required. The Street's bear case on Apple is straightforward. Apple has no frontier model of its own, Siri has stumbled for years and the company effectively rents OpenAI's models for its hardest queries. That narrative treats Apple as the one Magnificent Seven name that missed the AI wave entirely but the bull case flips that framing on its head. If frontier AI models keep shrinking and getting cheaper to run, Apple doesn't need the smartest model in the world, it just needs to own the device that model runs on. And unified memory is the mechanism that makes this possible. Unlike traditional systems where the CPU and GPU each need separate memory, Apple's architecture lets the CPU, GPU and Neural Engine draw from one shared pool. A fully specced M7 Ultra could theoretically run something on the scale of a 1.2 trillion parameter model locally and that capability plugs directly into the one advantage Apple has spent over a decade building: privacy. Apple has already shipped Private Cloud Compute, a system designed so even Apple can't access user data processed off device. Apple doubled down on this at WWDC 2026, framing on device privacy as non-negotiable while rivals default to the cloud. If the best AI models get small enough to run on Apple silicon, the moat stops being the model and becomes the hardware it has to sit on. Milk Road Pro remains bullish on Apple and it remains as one of our core positions, if you want the full thesis + our full AI trades, come join us using the link below for just a $1.
2:41

Everyone wrote Apple off as the AI loser, but one hardware spec might flip that story upside down (Save this). @jason called Apple a screaming buy on the back of a single chip detail. The rumored M7 Ultra, expected around 2028, is designed to support up to 1.5TB of unified memory, enough to run frontier class trillion parameter AI models locally, with no cloud required. The Street's bear case on Apple is straightforward. Apple has no frontier model of its own, Siri has stumbled for years and the company effectively rents OpenAI's models for its hardest queries. That narrative treats Apple as the one Magnificent Seven name that missed the AI wave entirely but the bull case flips that framing on its head. If frontier AI models keep shrinking and getting cheaper to run, Apple doesn't need the smartest model in the world, it just needs to own the device that model runs on. And unified memory is the mechanism that makes this possible. Unlike traditional systems where the CPU and GPU each need separate memory, Apple's architecture lets the CPU, GPU and Neural Engine draw from one shared pool. A fully specced M7 Ultra could theoretically run something on the scale of a 1.2 trillion parameter model locally and that capability plugs directly into the one advantage Apple has spent over a decade building: privacy. Apple has already shipped Private Cloud Compute, a system designed so even Apple can't access user data processed off device. Apple doubled down on this at WWDC 2026, framing on device privacy as non-negotiable while rivals default to the cloud. If the best AI models get small enough to run on Apple silicon, the moat stops being the model and becomes the hardware it has to sit on. Milk Road Pro remains bullish on Apple and it remains as one of our core positions, if you want the full thesis + our full AI trades, come join us using the link below for just a $1.

Milk Road AI

37,459 views • 2 months ago

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.
0:33

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.

FHILY👑

39,982 views • 16 days ago

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.
0:23

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

273,409 views • 2 months ago

Excited to show some surprising inventions on generative multiplayer games we made at Google with Stanford. We call the work MultiGen. I've always been inspired by early studios like id Software with Doom or Blizzard with Warcraft bringing networked video games to the next level. We are at the point in history where we can make strides like them, but for generative games. It's a strange feeling to be in the age of generative video games while still discovering how exactly to train the models and design the tools that make them useful. All of the tools that have been invented for classic game engines need to be redesigned for generative games. For example level and world design is not entirely possible with existing technology. We introduce editable memory to diffusion game engines that allow for design of new levels via a minimap. But we can easily imagine how this can be expanded with different creation tools. The end goal of this research direction is to allow game designers to be able to guide the generation process of their world, at the granularity that they prefer. Editable memory also allows us to add multiplayer to Generative Doom. We were amazed when we saw GameNGen some years ago, and now you can play it live with friends in real-time, on your couch or even online. Shared representations like our editable memory seem like the future for this type of experience. Models are, in some cases, expensive and approximate encoders but great interpolators and extrapolators. Leveraging their strengths lets you have completely new experiences that can be realized now and not in the distant future. This work was started at my previous team and continued in collaboration with Stanford. Congratulations to all for the discoveries.
0:40

Excited to show some surprising inventions on generative multiplayer games we made at Google with Stanford. We call the work MultiGen. I've always been inspired by early studios like id Software with Doom or Blizzard with Warcraft bringing networked video games to the next level. We are at the point in history where we can make strides like them, but for generative games. It's a strange feeling to be in the age of generative video games while still discovering how exactly to train the models and design the tools that make them useful. All of the tools that have been invented for classic game engines need to be redesigned for generative games. For example level and world design is not entirely possible with existing technology. We introduce editable memory to diffusion game engines that allow for design of new levels via a minimap. But we can easily imagine how this can be expanded with different creation tools. The end goal of this research direction is to allow game designers to be able to guide the generation process of their world, at the granularity that they prefer. Editable memory also allows us to add multiplayer to Generative Doom. We were amazed when we saw GameNGen some years ago, and now you can play it live with friends in real-time, on your couch or even online. Shared representations like our editable memory seem like the future for this type of experience. Models are, in some cases, expensive and approximate encoders but great interpolators and extrapolators. Leveraging their strengths lets you have completely new experiences that can be realized now and not in the distant future. This work was started at my previous team and continued in collaboration with Stanford. Congratulations to all for the discoveries.

Nataniel Ruiz

104,972 views • 6 months ago

Already doing a version of this for my daughter…custom Framework laptop with exactly everything she needs. It’s amazing how far one can get with Grok 4.6, QuickShell, and Omarchy Linux. Some cool things I got working in just a few hours include: - Full voice control over the OS - A cute Wall-E like friend you can summon (using a Grok voice plus some ffmpeg distortion while streaming it through the speakers). The eyes animations map to 50 or so emotions. - Memory and “scene memory” - All local (except Grok voice of course) Scene memory is like this: If you’re working on some math homework and you are doing some exercises, it’ll remember it later, you can even say “ok time to work on math again” and it’ll pull up your exact setup from before (same apps, windows, etc) or if you’re working on a paper and you ask to open it, it knows where it’s at and will get it, etc. Next is In going to hook it up to a coding agent with some tools to build quickshell apps that she wants on a moments notice (this whole thing is in quickshell btw). The goal is a laptop that conforms to exactly what my daughter wants or needs at any moment. My other daughters have already asked for one (even my college aged one). And Matthew Congrove already texted me “Time to give my girls a Linux laptop I guess” 😂 So much fun! And so many opportunities.
1:30

Already doing a version of this for my daughter…custom Framework laptop with exactly everything she needs. It’s amazing how far one can get with Grok 4.6, QuickShell, and Omarchy Linux. Some cool things I got working in just a few hours include: - Full voice control over the OS - A cute Wall-E like friend you can summon (using a Grok voice plus some ffmpeg distortion while streaming it through the speakers). The eyes animations map to 50 or so emotions. - Memory and “scene memory” - All local (except Grok voice of course) Scene memory is like this: If you’re working on some math homework and you are doing some exercises, it’ll remember it later, you can even say “ok time to work on math again” and it’ll pull up your exact setup from before (same apps, windows, etc) or if you’re working on a paper and you ask to open it, it knows where it’s at and will get it, etc. Next is In going to hook it up to a coding agent with some tools to build quickshell apps that she wants on a moments notice (this whole thing is in quickshell btw). The goal is a laptop that conforms to exactly what my daughter wants or needs at any moment. My other daughters have already asked for one (even my college aged one). And Matthew Congrove already texted me “Time to give my girls a Linux laptop I guess” 😂 So much fun! And so many opportunities.

Rick Blalock

14,455 views • 1 month ago

Chinese AI models are wiping billions off Big Tech right now. Google just lost $200 billion in a single day, and the model it needed to fight back still isn't ready. Gemini 3.5 Pro, Google's most powerful model, is months behind schedule. Alphabet stock dropped 4.4% that same day. The Deepseek moment is happening again, and the new model is FAR bigger. On the same day Google's delay leaked, a Beijing lab called Moonshot released Kimi K3. It is the largest open model ever built, with 2.8 trillion parameters. It took the number one spot on the Frontend Code Arena, a live coding leaderboard, passing Anthropic's best model. And Moonshot is giving it away for free on July 27. The genius part: Anyone with enough computers can download it and run a frontier level AI without paying a cent to a US company. A single task on Kimi K3 costs about 94 cents. The same work on some American models costs nearly double. So why would a company keep paying premium prices for a model it can now get for free? The entire US AI business is built on selling access to models that cost billions to train. If a free Chinese version does most of the same work, that pricing power starts to crack. And Kimi is close to the best. On one closely watched intelligence ranking it scored 57, just behind the top American models GPT-5.6 Sol and Fable 5, and ahead of Claude Opus 4.8. Bank of America told clients that Kimi proves Chinese labs can keep making big leaps even with limited chips. And the founder of Moonshot, Yang Zhilin, learned to build AI as a researcher INSIDE Google. Google literally wrote the 2017 paper that made all of these models possible. Now the people who studied its work are using it to destroy Google, and handing it out for free. What happens next: Kimi K3's weights go public on July 27. Google reports earnings on July 22, and everyone will be asking the same question about Gemini. If free models keep topping the charts, every valuation built on paid AI access has to be rewritten. What do you think?
1:10

Chinese AI models are wiping billions off Big Tech right now. Google just lost $200 billion in a single day, and the model it needed to fight back still isn't ready. Gemini 3.5 Pro, Google's most powerful model, is months behind schedule. Alphabet stock dropped 4.4% that same day. The Deepseek moment is happening again, and the new model is FAR bigger. On the same day Google's delay leaked, a Beijing lab called Moonshot released Kimi K3. It is the largest open model ever built, with 2.8 trillion parameters. It took the number one spot on the Frontend Code Arena, a live coding leaderboard, passing Anthropic's best model. And Moonshot is giving it away for free on July 27. The genius part: Anyone with enough computers can download it and run a frontier level AI without paying a cent to a US company. A single task on Kimi K3 costs about 94 cents. The same work on some American models costs nearly double. So why would a company keep paying premium prices for a model it can now get for free? The entire US AI business is built on selling access to models that cost billions to train. If a free Chinese version does most of the same work, that pricing power starts to crack. And Kimi is close to the best. On one closely watched intelligence ranking it scored 57, just behind the top American models GPT-5.6 Sol and Fable 5, and ahead of Claude Opus 4.8. Bank of America told clients that Kimi proves Chinese labs can keep making big leaps even with limited chips. And the founder of Moonshot, Yang Zhilin, learned to build AI as a researcher INSIDE Google. Google literally wrote the 2017 paper that made all of these models possible. Now the people who studied its work are using it to destroy Google, and handing it out for free. What happens next: Kimi K3's weights go public on July 27. Google reports earnings on July 22, and everyone will be asking the same question about Gemini. If free models keep topping the charts, every valuation built on paid AI access has to be rewritten. What do you think?

Ricardo

47,790 views • 2 months ago

Chamath: Two terms you need to pay attention to in AI are Prefill and Decode “There's two terms that I think you're going to hear a ton about over these next few years.” “The first term is prefill, and the next is decode.” “What prefill and decode are, are two very distinct ways of how models think, and how a model goes through the process of answering a question that you ask it.” “And so when you send a prompt to AI, what happens is that the model processes it. This is called the reading phase or prefill.” “It reads your entire prompt all at once. And then it does a bunch of math, calculates all these relationships between all the words, and it stores them in temporary memory.” “The problem is that this is really compute bound. So it requires massive brute force. And Nvidia GPUs crush here.” “And their architecture is designed for massive parallel processing, which makes them really amazing at digesting these long prompts.” “So the problem just gets bigger and bigger, Nvidia just completely dominates.” “But the next phase though, this critical phase, the decode phase, is the writing phase, right?” “So the model starts to generate a response, you ask it a question and its response, one token at a time.” “And then to pick the next token to pick the next word, it has to look back at everything it has said already so that it doesn't hallucinate.” “The problem is that this is incredibly memory bandwidth constrained.” “And in our architecture, a long time ago, we made these design decisions from day one.” “And so what we did was we took a very different architectural approach, we took a very conservative process technology. We weren't pushing the boundaries of physics.” “And we used a lot of what's called SRAM. So memory on the chip so that we could do this decode thing as well or better than everybody else.” “And so now when you put these two things together, I just think it's going to create a huge acceleration in the ability for this entire infrastructure layer to get much cheaper and much more valuable, which I suspect then it'll have a lot more developer pull, you'll get a lot more applications being built, billions and billions of more people using it.”
1:52

Chamath: Two terms you need to pay attention to in AI are Prefill and Decode “There's two terms that I think you're going to hear a ton about over these next few years.” “The first term is prefill, and the next is decode.” “What prefill and decode are, are two very distinct ways of how models think, and how a model goes through the process of answering a question that you ask it.” “And so when you send a prompt to AI, what happens is that the model processes it. This is called the reading phase or prefill.” “It reads your entire prompt all at once. And then it does a bunch of math, calculates all these relationships between all the words, and it stores them in temporary memory.” “The problem is that this is really compute bound. So it requires massive brute force. And Nvidia GPUs crush here.” “And their architecture is designed for massive parallel processing, which makes them really amazing at digesting these long prompts.” “So the problem just gets bigger and bigger, Nvidia just completely dominates.” “But the next phase though, this critical phase, the decode phase, is the writing phase, right?” “So the model starts to generate a response, you ask it a question and its response, one token at a time.” “And then to pick the next token to pick the next word, it has to look back at everything it has said already so that it doesn't hallucinate.” “The problem is that this is incredibly memory bandwidth constrained.” “And in our architecture, a long time ago, we made these design decisions from day one.” “And so what we did was we took a very different architectural approach, we took a very conservative process technology. We weren't pushing the boundaries of physics.” “And we used a lot of what's called SRAM. So memory on the chip so that we could do this decode thing as well or better than everybody else.” “And so now when you put these two things together, I just think it's going to create a huge acceleration in the ability for this entire infrastructure layer to get much cheaper and much more valuable, which I suspect then it'll have a lot more developer pull, you'll get a lot more applications being built, billions and billions of more people using it.”

The All-In Podcast

567,660 views • 8 months ago

A new way of working. And a scary one at that. Memory Store is one of a group of new kinds of AI-first companies that can turn you into a Fast Company. I’m using several of them on my desktop and they are a dramatically new way to work. It builds a memory for: 1. Your AI agents. 2. Any employee using it. 3. The company itself. I sit down with founder Diwank Singh Tomer, Diwank Singh Tomer, who both freaks me out as well as shows how AI can radically help workers as well as managers. First, why does it freak me out? Well, his AI watches nearly everything a worker does and keeps a “memory” of it. It watches your email. Your calendar. Your Slack. And a whole lot of other things. This can really freak out workers if “forced” on them. And leads to a whole new set of security issues companies need to consider before adopting these things. Such data about a company could give a competitor a HUGE advantage, if leaked. They would know how a company “thinks.” It really is a surveillance system for employees and the company itself. OK, now why would anyone ever use such a thing? Because it gives employees super powers. It makes them more productive. Shows workers a lot of things about themselves, and helps them work and stay on task. It also gives the company super powers. Institutional memory stays with the AI now, even if an employee dies or leaves. As companies move to “AI First” approaches, they will increasingly see the value in companies like Memory Store. It prepares employees for meetings. It helps them remember things. It shows them what they should be working on, and helps them do it. Memory Store builds a memory for: 1. Your agents. 2. Your company. 3. Yourself, or any employee on it. This helps all three work better together. Diwank Singh Tomer and I go in depth about what it does and how deeply it improves working at a company that deploys it. But to get the ultimate benefits you gotta convince your coworkers to use it. And your managers to approve it. Which means you have to get over your fears and get everyone you work with over theirs too. Which will be the challenge for Diwank. Luckily for him his first customers are raving about how good it is and how much his platform helped their companies. Increases sales. Makes teams more productive. Decreases errors and unnecessary costs. Which tells me everyone soon will be using systems like this. This is what the new way of working looks like. Once I got over my fears it sure is an amazing way to work. Will you try working this way?
52:08

A new way of working. And a scary one at that. Memory Store is one of a group of new kinds of AI-first companies that can turn you into a Fast Company. I’m using several of them on my desktop and they are a dramatically new way to work. It builds a memory for: 1. Your AI agents. 2. Any employee using it. 3. The company itself. I sit down with founder Diwank Singh Tomer, Diwank Singh Tomer, who both freaks me out as well as shows how AI can radically help workers as well as managers. First, why does it freak me out? Well, his AI watches nearly everything a worker does and keeps a “memory” of it. It watches your email. Your calendar. Your Slack. And a whole lot of other things. This can really freak out workers if “forced” on them. And leads to a whole new set of security issues companies need to consider before adopting these things. Such data about a company could give a competitor a HUGE advantage, if leaked. They would know how a company “thinks.” It really is a surveillance system for employees and the company itself. OK, now why would anyone ever use such a thing? Because it gives employees super powers. It makes them more productive. Shows workers a lot of things about themselves, and helps them work and stay on task. It also gives the company super powers. Institutional memory stays with the AI now, even if an employee dies or leaves. As companies move to “AI First” approaches, they will increasingly see the value in companies like Memory Store. It prepares employees for meetings. It helps them remember things. It shows them what they should be working on, and helps them do it. Memory Store builds a memory for: 1. Your agents. 2. Your company. 3. Yourself, or any employee on it. This helps all three work better together. Diwank Singh Tomer and I go in depth about what it does and how deeply it improves working at a company that deploys it. But to get the ultimate benefits you gotta convince your coworkers to use it. And your managers to approve it. Which means you have to get over your fears and get everyone you work with over theirs too. Which will be the challenge for Diwank. Luckily for him his first customers are raving about how good it is and how much his platform helped their companies. Increases sales. Makes teams more productive. Decreases errors and unnecessary costs. Which tells me everyone soon will be using systems like this. This is what the new way of working looks like. Once I got over my fears it sure is an amazing way to work. Will you try working this way?

Robert Scoble

26,186 views • 4 months ago

Loved this 22-minute talk on continual learning for AI agents. Must watch for anyone looking to get agents performant and into production. Credit: Soheil Feizi at AI Engineer • Agent learning can happen at three layers: the model (weights), the harness (prompts, tools, skills, code, workflows), and memory (session or persistent). • Two fundamental challenges: (1) getting feedback, meaning how do we know if the agent did well and what it should have done instead, and (2) acting on that feedback, meaning deciding which layer or component to change and how. • Feedback sources differ by stage: In development you have benchmarks with evaluators that score pass/fail. In production you only have logs, which can be judged either automatically (LLMs or code analyzing the log, which is scalable) or by human experts (low volume but critical domain knowledge). • Logs plus feedback aren't enough because they're not testable: A single log with feedback is one observation of what happened. You need to lift it into a replayable learning environment, a simulation with tools, users, and defined evaluators, so candidate fixes can be run, verified, and compared. • Three ways to optimize the agent, with tradeoffs: Model-layer updates (SFT, RL post-training like DPO/GRPO, LoRA) are expensive and need benchmarks and evaluators. Harness updates (trace-to-harness coding agents, prompt search like GEPA) are flexible but either untestable and "vibe-based" or benchmark-dependent. Memory updates (fact storage like Letta/Mem0, skill distillation) are cheapest and fastest but usually unverified. • A good learning engine makes "the smallest durable change at the right layer" of the agent. • Verifiable continual learning (VCL): Improve an agent from its own experience where every fix is proven to help and proven to break nothing that already worked. It requires an executable test (replayable failure), a measured delta (score before and after), and regression tests (prior tests still pass). • Four principles of practical VCL: Replayability (turn one-off failures into rerunnable tests), holisticness (one failure can have causes in memory, prompts, tools, workflow, or model, so route the fix to the right layer), lifelongness (fix new failures subject to no regression on past environments, with regression handled inside the optimization loop rather than post-hoc), and efficiency (the loop must run frequently and cheaply, without scaling linearly as past environments accumulate). • Three takeaways: (1) Agent continual learning isn't necessarily fine-tuning; many useful updates live in the harness and memory layers. (2) Production logs are not learning environments and must be transformed into replayable ones. (3) The frontier is regression-aware improvement: fixing new failures while verifying you don't break old ones.
22:34

Loved this 22-minute talk on continual learning for AI agents. Must watch for anyone looking to get agents performant and into production. Credit: Soheil Feizi at AI Engineer • Agent learning can happen at three layers: the model (weights), the harness (prompts, tools, skills, code, workflows), and memory (session or persistent). • Two fundamental challenges: (1) getting feedback, meaning how do we know if the agent did well and what it should have done instead, and (2) acting on that feedback, meaning deciding which layer or component to change and how. • Feedback sources differ by stage: In development you have benchmarks with evaluators that score pass/fail. In production you only have logs, which can be judged either automatically (LLMs or code analyzing the log, which is scalable) or by human experts (low volume but critical domain knowledge). • Logs plus feedback aren't enough because they're not testable: A single log with feedback is one observation of what happened. You need to lift it into a replayable learning environment, a simulation with tools, users, and defined evaluators, so candidate fixes can be run, verified, and compared. • Three ways to optimize the agent, with tradeoffs: Model-layer updates (SFT, RL post-training like DPO/GRPO, LoRA) are expensive and need benchmarks and evaluators. Harness updates (trace-to-harness coding agents, prompt search like GEPA) are flexible but either untestable and "vibe-based" or benchmark-dependent. Memory updates (fact storage like Letta/Mem0, skill distillation) are cheapest and fastest but usually unverified. • A good learning engine makes "the smallest durable change at the right layer" of the agent. • Verifiable continual learning (VCL): Improve an agent from its own experience where every fix is proven to help and proven to break nothing that already worked. It requires an executable test (replayable failure), a measured delta (score before and after), and regression tests (prior tests still pass). • Four principles of practical VCL: Replayability (turn one-off failures into rerunnable tests), holisticness (one failure can have causes in memory, prompts, tools, workflow, or model, so route the fix to the right layer), lifelongness (fix new failures subject to no regression on past environments, with regression handled inside the optimization loop rather than post-hoc), and efficiency (the loop must run frequently and cheaply, without scaling linearly as past environments accumulate). • Three takeaways: (1) Agent continual learning isn't necessarily fine-tuning; many useful updates live in the harness and memory layers. (2) Production logs are not learning environments and must be transformed into replayable ones. (3) The frontier is regression-aware improvement: fixing new failures while verifying you don't break old ones.

Alex Lieberman

20,417 views • 2 months ago

Hermes agent just left the terminal. 𝗛𝗲𝗿𝗺𝗲𝘀 𝗗𝗲𝘀𝗸𝘁𝗼𝗽 dropped yesterday. native app for macOS, Windows, and Linux. for months Hermes was the agent that learned your projects, wrote its own skills, and built a model of who you are. all of it buried in terminal logs. now it has a window. the important part is that it's not a wrapper. it runs the same agent core, the same sessions, memory, and skills as the CLI. you can start a task in the terminal and finish it in the app without anything resetting. the state is shared across every interface, not copied between them. what the GUI actually adds: → streaming chat that shows live tool calls and inline reasoning instead of a spinner → a preview rail that renders pages, code, and images right beside the conversation → an artifacts panel that collects every file the agent has ever produced → remote gateway mode, so you can point the app at a VPS and run the heavy work elsewhere → skills, cron, profiles, and gateways managed point-and-click instead of through YAML → voice mode, drag-drop files, and inline image generation remote gateway mode is the one worth slowing down on. the agent runs 24/7 on a $5 server while you control it from your laptop like a local app. other agent UIs are chatboxes with a logo. this one shows the autonomy instead of hiding it, so you watch the skills load, the tools fire, and the artifacts pile up as it works. it was teased in Jensen's GTC keynote. MIT licensed, local-first, no telemetry. if you already run Hermes, download it and everything is already there. your chats, memory, and skills carry straight over. i wrote a full masterclass on Hermes Agent that walks through the SOUL. md identity layer, the three-tier memory system, the self-evolving skills loop, and how to run three specialized agents 24/7. desktop is the interface that finally does all of it justice. the article is quoted below.
1:18

Hermes agent just left the terminal. 𝗛𝗲𝗿𝗺𝗲𝘀 𝗗𝗲𝘀𝗸𝘁𝗼𝗽 dropped yesterday. native app for macOS, Windows, and Linux. for months Hermes was the agent that learned your projects, wrote its own skills, and built a model of who you are. all of it buried in terminal logs. now it has a window. the important part is that it's not a wrapper. it runs the same agent core, the same sessions, memory, and skills as the CLI. you can start a task in the terminal and finish it in the app without anything resetting. the state is shared across every interface, not copied between them. what the GUI actually adds: → streaming chat that shows live tool calls and inline reasoning instead of a spinner → a preview rail that renders pages, code, and images right beside the conversation → an artifacts panel that collects every file the agent has ever produced → remote gateway mode, so you can point the app at a VPS and run the heavy work elsewhere → skills, cron, profiles, and gateways managed point-and-click instead of through YAML → voice mode, drag-drop files, and inline image generation remote gateway mode is the one worth slowing down on. the agent runs 24/7 on a $5 server while you control it from your laptop like a local app. other agent UIs are chatboxes with a logo. this one shows the autonomy instead of hiding it, so you watch the skills load, the tools fire, and the artifacts pile up as it works. it was teased in Jensen's GTC keynote. MIT licensed, local-first, no telemetry. if you already run Hermes, download it and everything is already there. your chats, memory, and skills carry straight over. i wrote a full masterclass on Hermes Agent that walks through the SOUL. md identity layer, the three-tier memory system, the self-evolving skills loop, and how to run three specialized agents 24/7. desktop is the interface that finally does all of it justice. the article is quoted below.

Akshay 🚀

51,540 views • 3 months ago