Loading video...

Video Failed to Load

Go Home

Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems Most AI theorem proving is a language model generating a proof in one shot, with a verifier bolted on at the end to check it. That's autocomplete with a grader —...

Marktechpost's profile picture

Marktechpost AI

11,526 subscribers

56,695 views • 1 month ago •via X (Twitter)

Science & TechnologyNews & PoliticsEducation
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Bash is all you need! Which is why I'm introducing my holiday project: just-bash just-bash is a pretty complete implementation of bash in TypeScript designed to be used as a bash tool by AI agents. Because it turns out agents love exploring data via shell scripts, even beyond coding. It comes with grep, sed, awk and the 99th percentile features that an agent like Claude Code or Cursor would use. In fact, Claude Code can use it for secure bash execution. In the package - A bash-tool for AI SDK - A binary for use by yourself or your coding agents - An overlay filesystem to feed files to your agent securely - A Vercel Sandbox compatible API, so you can quickly upgrade to a real VM if you need to run binaries - An example AI agent that explores the just-bash code base using just-bash - I imported the Oils shell bash compatibility suite and just-bash passes a very good chunk What is interesting about this codebase: It was essentially entirely written by Opus 4.5. Coding agents love bash and they are good at reproducing it. They are also great at text-book recursive descent parsers and AST tweet-walk interpreters. That said, it is, like, a lot of code and I didn't read it all 😅. This is very much a hack, but it also seems to be _really_ useful. I haven't really found anything agents want to use that it doesn't support and it's fast and secure (caveats apply). It doesn't have write access to your computer and the filesystem is given a root that the agent cannot escape from. Find it at Related: Our recent blog post how we migrated our data analysis agent to bash tools and achieved incredible quality improvements The video shows the example agent investigating the just-bash code base
1:47

Bash is all you need! Which is why I'm introducing my holiday project: just-bash just-bash is a pretty complete implementation of bash in TypeScript designed to be used as a bash tool by AI agents. Because it turns out agents love exploring data via shell scripts, even beyond coding. It comes with grep, sed, awk and the 99th percentile features that an agent like Claude Code or Cursor would use. In fact, Claude Code can use it for secure bash execution. In the package - A bash-tool for AI SDK - A binary for use by yourself or your coding agents - An overlay filesystem to feed files to your agent securely - A Vercel Sandbox compatible API, so you can quickly upgrade to a real VM if you need to run binaries - An example AI agent that explores the just-bash code base using just-bash - I imported the Oils shell bash compatibility suite and just-bash passes a very good chunk What is interesting about this codebase: It was essentially entirely written by Opus 4.5. Coding agents love bash and they are good at reproducing it. They are also great at text-book recursive descent parsers and AST tweet-walk interpreters. That said, it is, like, a lot of code and I didn't read it all 😅. This is very much a hack, but it also seems to be _really_ useful. I haven't really found anything agents want to use that it doesn't support and it's fast and secure (caveats apply). It doesn't have write access to your computer and the filesystem is given a root that the agent cannot escape from. Find it at Related: Our recent blog post how we migrated our data analysis agent to bash tools and achieved incredible quality improvements The video shows the example agent investigating the just-bash code base

Malte Ubl

125,326 views • 8 months ago

HERMES AGENT NOW RUNS ON AN 8GB LAPTOP GPU JUST AS EASILY AS IT RUNS ON A 128GB MINI PC Nous Research shipped the official Hermes Agent Desktop App this week. Someone pointed it at a local llama server running on an RTX 4060 with 16GB system RAM. The integration took two minutes The model behind it: Gemma 4 26B MoE, QAT quantized, running on 8GB of VRAM. A 60k token prompt held a stable 20 tokens a second, flat, no slowdown as context grew. The flags were nothing exotic, just -cmoe -c 248000 on llama.cpp What that 8GB setup does out of the box: reads and patches its own code, runs it in a terminal, debugs errors, manages GitHub repos, spawns sub-agents for parallel work. Browses the web with vision to debug a UI. Schedules cron jobs in plain language. Connects to Notion, Google Workspace, Linear, and Obsidian to manage tasks on its own That's the same agent layer running on a Minisforum MS-S1 MAX with 128GB of unified memory, 96GB of it to the GPU, holding a 120B model at 56 tokens a second instead of a 26B model at 20. Same software, same tool execution, same zero API key. The only thing that changes between an $800 laptop and a $2,000 mini PC is how big a model you can afford to run underneath it The barrier to running a real autonomous agent locally didn't just drop. It dropped all the way down to hardware most people already own
0:20

HERMES AGENT NOW RUNS ON AN 8GB LAPTOP GPU JUST AS EASILY AS IT RUNS ON A 128GB MINI PC Nous Research shipped the official Hermes Agent Desktop App this week. Someone pointed it at a local llama server running on an RTX 4060 with 16GB system RAM. The integration took two minutes The model behind it: Gemma 4 26B MoE, QAT quantized, running on 8GB of VRAM. A 60k token prompt held a stable 20 tokens a second, flat, no slowdown as context grew. The flags were nothing exotic, just -cmoe -c 248000 on llama.cpp What that 8GB setup does out of the box: reads and patches its own code, runs it in a terminal, debugs errors, manages GitHub repos, spawns sub-agents for parallel work. Browses the web with vision to debug a UI. Schedules cron jobs in plain language. Connects to Notion, Google Workspace, Linear, and Obsidian to manage tasks on its own That's the same agent layer running on a Minisforum MS-S1 MAX with 128GB of unified memory, 96GB of it to the GPU, holding a 120B model at 56 tokens a second instead of a 26B model at 20. Same software, same tool execution, same zero API key. The only thing that changes between an $800 laptop and a $2,000 mini PC is how big a model you can afford to run underneath it The barrier to running a real autonomous agent locally didn't just drop. It dropped all the way down to hardware most people already own

NO1ennn

40,079 views • 2 months ago

Perplexity CEO Aravind Srinivas on the brutal truth about who actually makes money in AI (and why it's not who you think): Aravind argues that the real value in AI comes from orchestration. He points to products like Codex, Claude Code, and Perplexity Computer: "What is that? It's an orchestration system. It takes a model, pairs it with an agent harness." And what is an agent harness? "The simplest way of describing it is like rules for how the agent loop should run. What are all the skills and sub-agents and connectors and tools it accesses? Without the harness, you don't necessarily capture and convert the intrinsic intelligence in the model into valuable output tokens." This leads to a blunt conclusion about who has a real business in AI, and who doesn't: "If you're literally just a reseller of model tokens, you have no business, because the model will get commoditized. So even if you're a model builder, you don't have a business. As an infra layer, you have some business on serving those output tokens. But as an application layer or model builder, you don't really have a business if you're just a reseller of tokens that come directly out of the model." So where does the value accrue? "You have a business if you know how to take the model, ground it in valuable context, orchestrate it with a really good agent harness, connected to the right set of tools and connectors (whether it's personal connectors or business connectors) and provide the experience to people in one single unified system." Aravind Srinivas then explains Perplexity's specific edge: Beyond orchestrating across tools, files, and connectors, they also orchestrate across models. "That is the differentiation that Anthropic and OpenAI cannot claim, because you wouldn't find GPT-5 inside the Claude Code harness. You wouldn't find Claude Opus inside the Codex harness. These are competing with each other. Whereas you would find both these models inside Perplexity Computer." Why does this matter? Because it all comes down to power. In Aravind's framing, the fundamental cost driver in AI is watts (the one input nobody can subsidize except the government). "Whoever provides the most valuable output tokens with the least amount of power expended to produce them generates the greatest value to the end user, has the most pricing power, has the most value. That is the orchestration problem to solve." His conclusion: "The one single most important metric in AI is token value per watt per user."
2:50

Perplexity CEO Aravind Srinivas on the brutal truth about who actually makes money in AI (and why it's not who you think): Aravind argues that the real value in AI comes from orchestration. He points to products like Codex, Claude Code, and Perplexity Computer: "What is that? It's an orchestration system. It takes a model, pairs it with an agent harness." And what is an agent harness? "The simplest way of describing it is like rules for how the agent loop should run. What are all the skills and sub-agents and connectors and tools it accesses? Without the harness, you don't necessarily capture and convert the intrinsic intelligence in the model into valuable output tokens." This leads to a blunt conclusion about who has a real business in AI, and who doesn't: "If you're literally just a reseller of model tokens, you have no business, because the model will get commoditized. So even if you're a model builder, you don't have a business. As an infra layer, you have some business on serving those output tokens. But as an application layer or model builder, you don't really have a business if you're just a reseller of tokens that come directly out of the model." So where does the value accrue? "You have a business if you know how to take the model, ground it in valuable context, orchestrate it with a really good agent harness, connected to the right set of tools and connectors (whether it's personal connectors or business connectors) and provide the experience to people in one single unified system." Aravind Srinivas then explains Perplexity's specific edge: Beyond orchestrating across tools, files, and connectors, they also orchestrate across models. "That is the differentiation that Anthropic and OpenAI cannot claim, because you wouldn't find GPT-5 inside the Claude Code harness. You wouldn't find Claude Opus inside the Codex harness. These are competing with each other. Whereas you would find both these models inside Perplexity Computer." Why does this matter? Because it all comes down to power. In Aravind's framing, the fundamental cost driver in AI is watts (the one input nobody can subsidize except the government). "Whoever provides the most valuable output tokens with the least amount of power expended to produce them generates the greatest value to the end user, has the most pricing power, has the most value. That is the orchestration problem to solve." His conclusion: "The one single most important metric in AI is token value per watt per user."

Big Brain AI

42,105 views • 21 days ago

a regular laptop with just 8GB VRAM can now run a LOCAL AI AGENT directly on your device what you need: 1/ 8 GB VRAM 2/ 16 GB RAM 3/ $600+ laptop method: Kimi K2 + Hermes Agent Dekstop App what agent can do: 1/ read and write files 2/ patch code, open PRs on GitHub 3/ browse the web like a human 4/ work with Notion 5/ run scripts, debug, refactor this isn't 'ask a question, get an answer.' this is an agent that drives a task from start to finish on its own. before: local agent = either the cloud and $$$, or an RTX now: any gaming laptop from the last 3 years p.s. MoE is the key. Kimi K2 activates only a fraction of its params per token - that's why a 1T-param model runs on hardware where a dense 26B would already be choking did you try Hermes Dekstop?
0:20

a regular laptop with just 8GB VRAM can now run a LOCAL AI AGENT directly on your device what you need: 1/ 8 GB VRAM 2/ 16 GB RAM 3/ $600+ laptop method: Kimi K2 + Hermes Agent Dekstop App what agent can do: 1/ read and write files 2/ patch code, open PRs on GitHub 3/ browse the web like a human 4/ work with Notion 5/ run scripts, debug, refactor this isn't 'ask a question, get an answer.' this is an agent that drives a task from start to finish on its own. before: local agent = either the cloud and $$$, or an RTX now: any gaming laptop from the last 3 years p.s. MoE is the key. Kimi K2 activates only a fraction of its params per token - that's why a 1T-param model runs on hardware where a dense 26B would already be choking did you try Hermes Dekstop?

kaize

58,735 views • 2 months ago

Wow! Someone open sourced 100+ full AI apps you can just clone and ship.. and it's sitting at 124,000 stars on github it's one repo, awesome-llm-apps. and the important part.. these aren't tutorials teaching you to build an app. they're the finished apps themselves. RAG systems, AI agents, multi-agent teams.. tested end to end, apache licensed, yours to copy. a few that are actually in there: - a fraud investigation agent that cross-references public records to flag things that don't add up - a VC due-diligence agent team that writes startup investment analysis - a financial coach that builds you a budget, debt and savings plan - "chat with anything".. your github, gmail, PDFs, arxiv papers, youtube videos - a deep research agent that runs full web research on its own here's how to actually use it, not just star it: clone the repo, open the one app closest to what you need it's a working template.. swap in your data, your API key, your use case ship it as your own thing, or bolt three of them together into one bigger product the license literally lets you take these, customize them, and sell them. that's not a loophole, that's the point of apache-2.0. repo:
0:56

Wow! Someone open sourced 100+ full AI apps you can just clone and ship.. and it's sitting at 124,000 stars on github it's one repo, awesome-llm-apps. and the important part.. these aren't tutorials teaching you to build an app. they're the finished apps themselves. RAG systems, AI agents, multi-agent teams.. tested end to end, apache licensed, yours to copy. a few that are actually in there: - a fraud investigation agent that cross-references public records to flag things that don't add up - a VC due-diligence agent team that writes startup investment analysis - a financial coach that builds you a budget, debt and savings plan - "chat with anything".. your github, gmail, PDFs, arxiv papers, youtube videos - a deep research agent that runs full web research on its own here's how to actually use it, not just star it: clone the repo, open the one app closest to what you need it's a working template.. swap in your data, your API key, your use case ship it as your own thing, or bolt three of them together into one bigger product the license literally lets you take these, customize them, and sell them. that's not a loophole, that's the point of apache-2.0. repo:

Axel Bitblaze 🪓

46,487 views • 1 month ago

THIS GUY BUILT AN AUTONOMOUS AI AGENT OUT OF CLAUDE CODE + OBSIDIAN and this is way more interesting than another “use AI to take notes” demo the trick is simple: Obsidian is not the writing app here. it becomes the agent’s memory, task board, and context folder. Claude Code is not just answering prompts. it reads the vault, edits files, follows instructions, and keeps moving through the work like a junior operator with a filesystem. the reusable setup looks like this: 1. create an Obsidian vault for one project 2. keep goals, rules, tasks, decisions, and references as markdown files 3. point Claude Code at the folder 4. give it a clear operating loop: read context → choose next task → execute → write back what changed 5. use the notes as persistent memory instead of re-explaining the project every chat that’s the part people miss. the “agent” is not magic. it’s the boring combination of: - local files - explicit rules - task state - write access - a model that can run through the repo/vault Obsidian makes the memory human-readable. Claude Code makes the memory executable. that combo is why the video worked: it turns a notes app into an operating surface for actual work. best use cases: - content systems - research vaults - coding projects - client ops docs - personal knowledge bases that need actions, not just storage the caveat: if your vault is messy, your agent becomes messy too. folders, naming, “done” criteria, and forbidden actions matter more than the prompt. but once the structure is clean, this is one of the easiest ways to build an agent that remembers what happened yesterday without paying for a full custom app.
10:19

THIS GUY BUILT AN AUTONOMOUS AI AGENT OUT OF CLAUDE CODE + OBSIDIAN and this is way more interesting than another “use AI to take notes” demo the trick is simple: Obsidian is not the writing app here. it becomes the agent’s memory, task board, and context folder. Claude Code is not just answering prompts. it reads the vault, edits files, follows instructions, and keeps moving through the work like a junior operator with a filesystem. the reusable setup looks like this: 1. create an Obsidian vault for one project 2. keep goals, rules, tasks, decisions, and references as markdown files 3. point Claude Code at the folder 4. give it a clear operating loop: read context → choose next task → execute → write back what changed 5. use the notes as persistent memory instead of re-explaining the project every chat that’s the part people miss. the “agent” is not magic. it’s the boring combination of: - local files - explicit rules - task state - write access - a model that can run through the repo/vault Obsidian makes the memory human-readable. Claude Code makes the memory executable. that combo is why the video worked: it turns a notes app into an operating surface for actual work. best use cases: - content systems - research vaults - coding projects - client ops docs - personal knowledge bases that need actions, not just storage the caveat: if your vault is messy, your agent becomes messy too. folders, naming, “done” criteria, and forbidden actions matter more than the prompt. but once the structure is clean, this is one of the easiest ways to build an agent that remembers what happened yesterday without paying for a full custom app.

kocer

30,403 views • 2 months ago

OpenClaw meets RL! OpenClaw Agents adapt through memory files and skills, but the base model weights never actually change. OpenClaw-RL solves this! It wraps a self-hosted model as an OpenAI-compatible API, intercepts live conversations from OpenClaw, and trains the policy in the background using RL. The architecture is fully async. This means serving, reward scoring, and training all run in parallel. Once done, weights get hot-swapped after every batch while the agent keeps responding. Currently, it has two training modes: - Binary RL (GRPO): A process reward model scores each turn as good, bad, or neutral. That scalar reward drives policy updates via a PPO-style clipped objective. - On-Policy Distillation: When concrete corrections come in like "you should have checked that file first," it uses that feedback as a richer, directional training signal at the token level. When to use OpenClaw-RL? To be fair, a lot of agent behavior can already be improved through better memory and skill design. OpenClaw's existing skill ecosystem and community-built self-improvement skills handle a wide range of use cases without touching model weights at all. If the agent keeps forgetting preferences, that's a memory problem. And if it doesn't know how to handle a specific workflow, that's a skill problem. Both are solvable at the prompt and context layer. Where RL becomes interesting is when the failure pattern lives deeper in the model's reasoning itself. Things like consistently poor tool selection order, weak multi-step planning, or failing to interpret ambiguous instructions the way a specific user intends. Research on agentic RL (like ARTIST and Agent-R1) has shown that these behavioral patterns hit a ceiling with prompt-based approaches alone, especially in complex multi-turn tasks where the model needs to recover from tool failures or adapt its strategy mid-execution. That's the layer OpenClaw-RL targets, and it's a meaningful distinction from what OpenClaw offers. I have shared the repo in the replies!
0:35

OpenClaw meets RL! OpenClaw Agents adapt through memory files and skills, but the base model weights never actually change. OpenClaw-RL solves this! It wraps a self-hosted model as an OpenAI-compatible API, intercepts live conversations from OpenClaw, and trains the policy in the background using RL. The architecture is fully async. This means serving, reward scoring, and training all run in parallel. Once done, weights get hot-swapped after every batch while the agent keeps responding. Currently, it has two training modes: - Binary RL (GRPO): A process reward model scores each turn as good, bad, or neutral. That scalar reward drives policy updates via a PPO-style clipped objective. - On-Policy Distillation: When concrete corrections come in like "you should have checked that file first," it uses that feedback as a richer, directional training signal at the token level. When to use OpenClaw-RL? To be fair, a lot of agent behavior can already be improved through better memory and skill design. OpenClaw's existing skill ecosystem and community-built self-improvement skills handle a wide range of use cases without touching model weights at all. If the agent keeps forgetting preferences, that's a memory problem. And if it doesn't know how to handle a specific workflow, that's a skill problem. Both are solvable at the prompt and context layer. Where RL becomes interesting is when the failure pattern lives deeper in the model's reasoning itself. Things like consistently poor tool selection order, weak multi-step planning, or failing to interpret ambiguous instructions the way a specific user intends. Research on agentic RL (like ARTIST and Agent-R1) has shown that these behavioral patterns hit a ceiling with prompt-based approaches alone, especially in complex multi-turn tasks where the model needs to recover from tool failures or adapt its strategy mid-execution. That's the layer OpenClaw-RL targets, and it's a meaningful distinction from what OpenClaw offers. I have shared the repo in the replies!

Avi Chawla

138,769 views • 5 months ago

i just built a 4-agent software team. everything runs from Telegram and gets managed on a kanban board. a project manager who plans the work, a backend developer, a frontend developer, and a tester. the PM reads a goal, breaks it into linked tasks, and assigns each to the right agent. the thing that makes them a team instead of four strangers is a shared kanban board. every task is a row that survives crashes, and when an agent finishes, it writes a summary of what it built and what the next agent needs to know. the next agent reads that summary before it starts. so the frontend developer never has to guess the API shape, and the tester knows exactly what to verify. the hardest part was not the coordination. it was building an agent that could actually act like a backend engineer. a backend engineer stands up a database, wires auth, manages storage, deploys functions, and keeps all of it consistent while the rest of the team builds on top. an agent doing this from scratch drowns. it burns its context window remembering which tables exist and which endpoint it created three steps ago, and the work degrades fast. so the backend agent needs a backend built for agents, not for humans clicking through a dashboard. that is where InsForge came in. it is an open-source, agent-native backend, and i added it to my backend developer agent as a skill. a skill is a step-by-step guide that teaches the agent how to do a specific kind of work. with InsForge installed, the agent stopped improvising infrastructure and followed a reliable path: create the project, define the database, set up auth, deploy functions. to test the whole team, i had them build a working Google Docs clone, AI features included. the backend agent spun up the full service on its own. database tables, user auth, document handling, and edge functions running real TypeScript, all in one dashboard. the frontend agent read that summary and built the UI on top of it, and the tester closed the loop. the result was a backend an agent could reason about end to end, instead of one it kept getting lost inside. if you are building an AI backend engineer, InsForge is worth a look, it's 100% open-source. InsForge GitHub: (don't forget to star 🌟) the full article on Hermes Kanban: Mission Control for your Agents is quoted below.
0:22

i just built a 4-agent software team. everything runs from Telegram and gets managed on a kanban board. a project manager who plans the work, a backend developer, a frontend developer, and a tester. the PM reads a goal, breaks it into linked tasks, and assigns each to the right agent. the thing that makes them a team instead of four strangers is a shared kanban board. every task is a row that survives crashes, and when an agent finishes, it writes a summary of what it built and what the next agent needs to know. the next agent reads that summary before it starts. so the frontend developer never has to guess the API shape, and the tester knows exactly what to verify. the hardest part was not the coordination. it was building an agent that could actually act like a backend engineer. a backend engineer stands up a database, wires auth, manages storage, deploys functions, and keeps all of it consistent while the rest of the team builds on top. an agent doing this from scratch drowns. it burns its context window remembering which tables exist and which endpoint it created three steps ago, and the work degrades fast. so the backend agent needs a backend built for agents, not for humans clicking through a dashboard. that is where InsForge came in. it is an open-source, agent-native backend, and i added it to my backend developer agent as a skill. a skill is a step-by-step guide that teaches the agent how to do a specific kind of work. with InsForge installed, the agent stopped improvising infrastructure and followed a reliable path: create the project, define the database, set up auth, deploy functions. to test the whole team, i had them build a working Google Docs clone, AI features included. the backend agent spun up the full service on its own. database tables, user auth, document handling, and edge functions running real TypeScript, all in one dashboard. the frontend agent read that summary and built the UI on top of it, and the tester closed the loop. the result was a backend an agent could reason about end to end, instead of one it kept getting lost inside. if you are building an AI backend engineer, InsForge is worth a look, it's 100% open-source. InsForge GitHub: (don't forget to star 🌟) the full article on Hermes Kanban: Mission Control for your Agents is quoted below.

Akshay 🚀

122,548 views • 2 months ago

Would you believe an AI agent can test a real VR action game in real time, the way a person plays it? Meta XR Operator makes it possible. As far as I know, this is the first time. I am not talking about tapping a menu or replaying a recorded click path, but genuinely moving, shooting, and using the same game mechanics a human player does. In NeonReach VR, which is a real (and open source) action game, rings spawn 12m out and come at you somewhere between 1.5 and 5.5 m/s, getting faster over a 90 second ramp. There are three kinds: straight, weaving side to side, and spinning. Every shot is a full slingshot cycle, so you press, pull back, aim, then release. Obstacles arrive at head height and cost you a life if you don't get out of the way. You have ten lives. Here is why the game is hard for an AI agent. Even though Meta XR Operator gives the agent everything it needs to observe the app and act inside it, the agent still cannot play. One agent turn takes 10 to 15 seconds. One throw is four steps that have to happen in order, because the press has to latch before the pull, and they cannot be batched into a single call. So a throw costs about 45 seconds. A fast ring only exists for 2.3 seconds. One action takes 20x longer than the target is alive. Prompt tuning does not close a gap that size. What works is a three stage path: EXPLORATION, then SKILL, then SCRIPT. 1/ EXPLORATION. The agent drives the live app and works the game out on its own. It verified the coordinate mapping by setting a pose and reading it back, then derived the launch model. The more useful output was the traps it found. For example, the player's own body collider silently deflects a ball released inside it, with no error and no log line. That produced two confident wrong conclusions before anyone caught them. 2/ SKILL. All of that gets written down as a reusable SKILL.md plus an aim solver. There is a section that separates what was actually verified from what was assumed, so a wrong conclusion cannot quietly turn into doctrine. This stage also produced the trick that mattered. Set timeScale to 0 and a throw becomes atomic in game time, so however long the agent spends thinking never shows up in the shot. 3/ SCRIPT. The agent then compiles everything into a player script, a loop that observes, decides, and throws, calling the MCP servers directly from Python with no model in the hot path. Round trips drop from 10 to 15 seconds down to something between 1 and 16 milliseconds. The loop runs at 23 Hz, about 0.75 seconds per throw, roughly 60x faster than the agent doing it turn by turn. The result is that it plays like a person, which you can see from the attached video. It tracks the rings, works out where each one is going, throws with whichever hand is free, moves out of the way of the obstacles, and does not wait around to see whether the last throw landed. Shipping settings, no difficulty edits, no health locks, no slow motion. It plays until it actually loses. The takeaway generalizes beyond games: an agent does not have to be the player. Even following the same rules as a player, it has too much latency between moves. Having the agent write the thing that acts bypasses that constraint entirely. Try it yourself: or explore the agent-created skill and scripts: Based on NeonReach VR by Dilmer, with no code changes. I only upgraded its Meta XR Core SDK to v205, which ships Meta XR Operator. Our blog post, Introducing Meta XR Operator: Close the Build-Test-Verify Loop for VR: Disclosure: I work at Meta. And this represents my own opinion. #XR #VR #AI #MetaQuest #Unity #GameDev
1:21

Would you believe an AI agent can test a real VR action game in real time, the way a person plays it? Meta XR Operator makes it possible. As far as I know, this is the first time. I am not talking about tapping a menu or replaying a recorded click path, but genuinely moving, shooting, and using the same game mechanics a human player does. In NeonReach VR, which is a real (and open source) action game, rings spawn 12m out and come at you somewhere between 1.5 and 5.5 m/s, getting faster over a 90 second ramp. There are three kinds: straight, weaving side to side, and spinning. Every shot is a full slingshot cycle, so you press, pull back, aim, then release. Obstacles arrive at head height and cost you a life if you don't get out of the way. You have ten lives. Here is why the game is hard for an AI agent. Even though Meta XR Operator gives the agent everything it needs to observe the app and act inside it, the agent still cannot play. One agent turn takes 10 to 15 seconds. One throw is four steps that have to happen in order, because the press has to latch before the pull, and they cannot be batched into a single call. So a throw costs about 45 seconds. A fast ring only exists for 2.3 seconds. One action takes 20x longer than the target is alive. Prompt tuning does not close a gap that size. What works is a three stage path: EXPLORATION, then SKILL, then SCRIPT. 1/ EXPLORATION. The agent drives the live app and works the game out on its own. It verified the coordinate mapping by setting a pose and reading it back, then derived the launch model. The more useful output was the traps it found. For example, the player's own body collider silently deflects a ball released inside it, with no error and no log line. That produced two confident wrong conclusions before anyone caught them. 2/ SKILL. All of that gets written down as a reusable SKILL.md plus an aim solver. There is a section that separates what was actually verified from what was assumed, so a wrong conclusion cannot quietly turn into doctrine. This stage also produced the trick that mattered. Set timeScale to 0 and a throw becomes atomic in game time, so however long the agent spends thinking never shows up in the shot. 3/ SCRIPT. The agent then compiles everything into a player script, a loop that observes, decides, and throws, calling the MCP servers directly from Python with no model in the hot path. Round trips drop from 10 to 15 seconds down to something between 1 and 16 milliseconds. The loop runs at 23 Hz, about 0.75 seconds per throw, roughly 60x faster than the agent doing it turn by turn. The result is that it plays like a person, which you can see from the attached video. It tracks the rings, works out where each one is going, throws with whichever hand is free, moves out of the way of the obstacles, and does not wait around to see whether the last throw landed. Shipping settings, no difficulty edits, no health locks, no slow motion. It plays until it actually loses. The takeaway generalizes beyond games: an agent does not have to be the player. Even following the same rules as a player, it has too much latency between moves. Having the agent write the thing that acts bypasses that constraint entirely. Try it yourself: or explore the agent-created skill and scripts: Based on NeonReach VR by Dilmer, with no code changes. I only upgraded its Meta XR Core SDK to v205, which ships Meta XR Operator. Our blog post, Introducing Meta XR Operator: Close the Build-Test-Verify Loop for VR: Disclosure: I work at Meta. And this represents my own opinion. #XR #VR #AI #MetaQuest #Unity #GameDev

Xiang Wei

51,880 views • 5 days ago

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
2:41

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 views • 1 month ago

A finance professor manages $200M with AI agents, and he told everyone why: "Large language models are at the level of a fourth-year PhD student in every field" Alejandro Lopez-Lira's AI fund, Autopilot, returned 56% last year. The S&P did 16%. There are 52,000 people with money in it, and most of them just watch the machine work. What he automated is the same six-step loop every fund on earth runs: find an idea, code it, backtest it, deploy it, read the autopsy, learn from it. A quant at Two Sigma runs that loop once a month, and the salary time alone costs around $50,000 per hypothesis. All steps from this loop now fit in AI trading text box. Plain English in, executable strategy out, five-year backtest in 12 seconds, live on a broker 90 seconds after you typed the sentence. He runs $200M with AI. You can run same AI fund in two clicks, free to try: Step 6 on this loop is where everyone is stuck. Your agent has no memory. Every strategy it kills goes into a log nobody reads, and the next one starts from zero. Nobody keeps negative results. Not Citadel, not Man Group, not a single repo on GitHub. Fix that and the agent remembers every hypothesis it killed and the regime it died in. It stops burning cycles on your old mistakes. Jane Street pays 3,500 people to run this cycle and made $39.6 billion doing it. Five sixths of it is now free. Bookmark & read full map of this loop in the article below. Most people still think AI trading is out of reach for them - it isn't. Don't want to spend a dollar for testing this? Kalshi just opened a perps exchange and gives US users $25 free to start ->
0:48

A finance professor manages $200M with AI agents, and he told everyone why: "Large language models are at the level of a fourth-year PhD student in every field" Alejandro Lopez-Lira's AI fund, Autopilot, returned 56% last year. The S&P did 16%. There are 52,000 people with money in it, and most of them just watch the machine work. What he automated is the same six-step loop every fund on earth runs: find an idea, code it, backtest it, deploy it, read the autopsy, learn from it. A quant at Two Sigma runs that loop once a month, and the salary time alone costs around $50,000 per hypothesis. All steps from this loop now fit in AI trading text box. Plain English in, executable strategy out, five-year backtest in 12 seconds, live on a broker 90 seconds after you typed the sentence. He runs $200M with AI. You can run same AI fund in two clicks, free to try: Step 6 on this loop is where everyone is stuck. Your agent has no memory. Every strategy it kills goes into a log nobody reads, and the next one starts from zero. Nobody keeps negative results. Not Citadel, not Man Group, not a single repo on GitHub. Fix that and the agent remembers every hypothesis it killed and the regime it died in. It stops burning cycles on your old mistakes. Jane Street pays 3,500 people to run this cycle and made $39.6 billion doing it. Five sixths of it is now free. Bookmark & read full map of this loop in the article below. Most people still think AI trading is out of reach for them - it isn't. Don't want to spend a dollar for testing this? Kalshi just opened a perps exchange and gives US users $25 free to start ->

cvxv666

82,211 views • 18 days ago

NVIDIA just released a new open source transcription model, Nemotron Speech ASR, designed from the ground up for low-latency use cases like voice agents. Here's a voice agent built with this new model. 24ms transcription finalization and total voice-to-voice inference time under 500ms. This agent actually uses *three* NVIDIA open source models: - Nemotron Speech ASR - Nemotron 3 Nano 30GB in a 4-bit quant (released in December) - A preview checkpoint of the upcoming Magpie text-to-speech model These models are all truly open source: weights, training data, training code, and inference code. This is a big deal! Jensen said in the CES keynote yesterday that he expects open source models to catch up to proprietary models this year in a number of categories. NVIDIA is putting their weight behind making this happen. (As Alan Kay said, the best way to predict the future is to invent it.) The code for this agent is open source too, of course. You can deploy it to production with Modal and Pipecat AI cloud, or run locally on an NVIDIA DGX Spark or RTX 5090.
1:24

NVIDIA just released a new open source transcription model, Nemotron Speech ASR, designed from the ground up for low-latency use cases like voice agents. Here's a voice agent built with this new model. 24ms transcription finalization and total voice-to-voice inference time under 500ms. This agent actually uses *three* NVIDIA open source models: - Nemotron Speech ASR - Nemotron 3 Nano 30GB in a 4-bit quant (released in December) - A preview checkpoint of the upcoming Magpie text-to-speech model These models are all truly open source: weights, training data, training code, and inference code. This is a big deal! Jensen said in the CES keynote yesterday that he expects open source models to catch up to proprietary models this year in a number of categories. NVIDIA is putting their weight behind making this happen. (As Alan Kay said, the best way to predict the future is to invent it.) The code for this agent is open source too, of course. You can deploy it to production with Modal and Pipecat AI cloud, or run locally on an NVIDIA DGX Spark or RTX 5090.

kwindla

274,679 views • 7 months ago

The same kinds of productivity gains we've seen in coding with AI agents are heading to the rest of knowledge work. This is the jump when you go from having a chatbot to being able to actually have an agent go off and do work for minutes or even hours and come back with a complete work output that you then review. Here's an example of the new Box Agent filling out an RFP response from an existing knowledge base. This process would normally take hours to fill out, and requires the full attention of the user doing the work. Now, you provide the Box Agent with the RFP questions, and it will go off, make a plan, extract all the relevant questions, read through existing source material to come up with an answer, and then generate a new word document as the final output. All while you're doing something else. The key to this architecture is that the agent is able to use all of the same tools in the background that a user uses to get work done. The agent can search for documents, read entire files, run scripts and tools in the background, and even be able to write code on the fly to automate tasks it hasn't seen before. And best of all, the Box Agent will (soon) work from the Box MCP and CLI so you can invoke it in any agentic system as a step in a process. This kind of agent complexity would have been impossible even 6 months ago. Models consistently failed at tracking long running tasks or using the right tools at the right moment for the task. But this is all now possible because of models like GPT-5.4, Opus 4.6, and Gemini 3, and is only getting better by the month. Just as we moved from engineers writing code and using AI as an assistant to answer questions, in many areas of knowledge work -like legal, finance, consulting, sales, marketing, and more- when we have a problem we'll just kick off the AI agent to just go work on it for us in the background.
2:14

The same kinds of productivity gains we've seen in coding with AI agents are heading to the rest of knowledge work. This is the jump when you go from having a chatbot to being able to actually have an agent go off and do work for minutes or even hours and come back with a complete work output that you then review. Here's an example of the new Box Agent filling out an RFP response from an existing knowledge base. This process would normally take hours to fill out, and requires the full attention of the user doing the work. Now, you provide the Box Agent with the RFP questions, and it will go off, make a plan, extract all the relevant questions, read through existing source material to come up with an answer, and then generate a new word document as the final output. All while you're doing something else. The key to this architecture is that the agent is able to use all of the same tools in the background that a user uses to get work done. The agent can search for documents, read entire files, run scripts and tools in the background, and even be able to write code on the fly to automate tasks it hasn't seen before. And best of all, the Box Agent will (soon) work from the Box MCP and CLI so you can invoke it in any agentic system as a step in a process. This kind of agent complexity would have been impossible even 6 months ago. Models consistently failed at tracking long running tasks or using the right tools at the right moment for the task. But this is all now possible because of models like GPT-5.4, Opus 4.6, and Gemini 3, and is only getting better by the month. Just as we moved from engineers writing code and using AI as an assistant to answer questions, in many areas of knowledge work -like legal, finance, consulting, sales, marketing, and more- when we have a problem we'll just kick off the AI agent to just go work on it for us in the background.

Aaron Levie

24,618 views • 4 months ago

I solved building decks with AI agents — by giving them a CLI tool like Powerpoint or Google Slides. AI could already make a beautiful deck if you asked it to using Ant's pptx skill. The problem was working with it. If it made one alignment mistake, fixing it on one slide would break something on another, and it became a game of whack-a-mole. One time I spent two days playing AI roulette, hoping the next prompt would finally fix the thing, and ended up building the whole deck by hand because I was on a deadline. So I built Hands-on Deck. And the reason it works is that this isn't just a skill — this is PowerPoint. The actual application: PowerPoint, Google Slides, Keynote, whatever you use. This is that, but for an agent, presented as a CLI. Every gesture you make in a deck app maps to a command. Click a box and type, drag a shape from here to there, look at a slide – agent can do it all in a command. And that changes how the agent behaves. With this CLI it works and thinks like a designer — it looks, makes an edit, looks again, makes another surgical edit. Compare that to Anthropic's pptx skill, built on the idea that Claude is a great programmer: it literally writes code to manipulate the deck, hand-editing XML and hoping it doesn't break anything else in the middle. The real test isn't creating something once — it's whether it can make surgical edits like you want. That's what I did in this video walkthrough and my claude crushed it! Check it out for yourself. So decks can be built like a designer now — with real flavor and taste. If you spend hours every week on decks, this gives those hours back. You can install it as a skill in Claude Code, Codex, whatever you use. Works every harness that supports skills. Let me know if you make something cool with it.
21:00

I solved building decks with AI agents — by giving them a CLI tool like Powerpoint or Google Slides. AI could already make a beautiful deck if you asked it to using Ant's pptx skill. The problem was working with it. If it made one alignment mistake, fixing it on one slide would break something on another, and it became a game of whack-a-mole. One time I spent two days playing AI roulette, hoping the next prompt would finally fix the thing, and ended up building the whole deck by hand because I was on a deadline. So I built Hands-on Deck. And the reason it works is that this isn't just a skill — this is PowerPoint. The actual application: PowerPoint, Google Slides, Keynote, whatever you use. This is that, but for an agent, presented as a CLI. Every gesture you make in a deck app maps to a command. Click a box and type, drag a shape from here to there, look at a slide – agent can do it all in a command. And that changes how the agent behaves. With this CLI it works and thinks like a designer — it looks, makes an edit, looks again, makes another surgical edit. Compare that to Anthropic's pptx skill, built on the idea that Claude is a great programmer: it literally writes code to manipulate the deck, hand-editing XML and hoping it doesn't break anything else in the middle. The real test isn't creating something once — it's whether it can make surgical edits like you want. That's what I did in this video walkthrough and my claude crushed it! Check it out for yourself. So decks can be built like a designer now — with real flavor and taste. If you spend hours every week on decks, this gives those hours back. You can install it as a skill in Claude Code, Codex, whatever you use. Works every harness that supports skills. Let me know if you make something cool with it.

Nityesh

69,784 views • 2 months ago

Karpathy's prediction about RL is coming true now! He called reward functions unreliable and argued that a single reward number is too low-dimensional to teach an agent what "good" means for complex tasks. To solve this, Agents need a knowledge-guided review as a higher-dimensional feedback channel. Every major AI lab trains models with RL today (OpenAI, Anthropic, DeepSeek). And their key bottleneck has always been the reward functions. GRPO by DeepSeek worked well for math and code because the environment gave a binary signal. But for real agent tasks, someone still has to hand-code the scoring function. That takes days and breaks every time the pipeline changes. RULER (implemented in OpenPipe ART, 10k stars) addresses the exact problem Karpathy identified. The reward criteria are defined in plain English, and an LLM evaluates each trajectory against that description to provide feedback for training. I trained a Qwen3 1.4B agent that plays 2048 using GRPO with this exact workflow. In this case, the agent saw the board, picked a direction, and RULER evaluated the outcome, all from this natural language definition. You can see the full implementation on GitHub and try it yourself. Here's the ART Repo: (don't forget to star it ⭐ ) Just like RLHF replaced manual rankings and GRPO replaced the critic model, natural language rewards are replacing hand-coded scoring functions. RL reward engineering is now prompt engineering. I wrote a full walkthrough covering RL for LLM agents, from RLHF to GRPO to RULER, in the article below.
0:25

Karpathy's prediction about RL is coming true now! He called reward functions unreliable and argued that a single reward number is too low-dimensional to teach an agent what "good" means for complex tasks. To solve this, Agents need a knowledge-guided review as a higher-dimensional feedback channel. Every major AI lab trains models with RL today (OpenAI, Anthropic, DeepSeek). And their key bottleneck has always been the reward functions. GRPO by DeepSeek worked well for math and code because the environment gave a binary signal. But for real agent tasks, someone still has to hand-code the scoring function. That takes days and breaks every time the pipeline changes. RULER (implemented in OpenPipe ART, 10k stars) addresses the exact problem Karpathy identified. The reward criteria are defined in plain English, and an LLM evaluates each trajectory against that description to provide feedback for training. I trained a Qwen3 1.4B agent that plays 2048 using GRPO with this exact workflow. In this case, the agent saw the board, picked a direction, and RULER evaluated the outcome, all from this natural language definition. You can see the full implementation on GitHub and try it yourself. Here's the ART Repo: (don't forget to star it ⭐ ) Just like RLHF replaced manual rankings and GRPO replaced the critic model, natural language rewards are replacing hand-coded scoring functions. RL reward engineering is now prompt engineering. I wrote a full walkthrough covering RL for LLM agents, from RLHF to GRPO to RULER, in the article below.

Avi Chawla

350,512 views • 3 months ago

9 out of 10 multi-agent projects never leave demo mode Not because the model is bad. Because the structure is missing. Most people who try to build [ a team of AI agents ] end up with one agent talking to itself in five tabs. The agents don't share context → Don't divide work → Don't know what the others are doing. > stage 1: if your agent doesn't have a real loop, observe, act, iterate, you have a long prompt, not an agent > stage 2: subagents need isolated context, the orchestrator never reads their raw transcript, only the summary > stage 3: the orchestrator plans and delegates, the moment it executes, it drowns in details that belong inside subagents > stage 4: without a shared task list it's not a team, it's five agents duplicating each other's work in parallel > stage 5: a permissions file is what lets you sleep, the model cannot bypass it because the rule lives outside the model a team of AI agents is not more model, it is more structure the most ignored stage in every demo that died: - durability... when a 50-step task crashes at step 47 and starts from zero, that's not a model failure. that's a missing write-to-disk call
1:07

9 out of 10 multi-agent projects never leave demo mode Not because the model is bad. Because the structure is missing. Most people who try to build [ a team of AI agents ] end up with one agent talking to itself in five tabs. The agents don't share context → Don't divide work → Don't know what the others are doing. > stage 1: if your agent doesn't have a real loop, observe, act, iterate, you have a long prompt, not an agent > stage 2: subagents need isolated context, the orchestrator never reads their raw transcript, only the summary > stage 3: the orchestrator plans and delegates, the moment it executes, it drowns in details that belong inside subagents > stage 4: without a shared task list it's not a team, it's five agents duplicating each other's work in parallel > stage 5: a permissions file is what lets you sleep, the model cannot bypass it because the rule lives outside the model a team of AI agents is not more model, it is more structure the most ignored stage in every demo that died: - durability... when a 50-step task crashes at step 47 and starts from zero, that's not a model failure. that's a missing write-to-disk call

Shadow Nick

14,683 views • 2 months ago

Nothing beats open-source! We now have a fully open-source coding agent that can compete with any proprietary solution out there: • 500+ available models • It doesn't train on your code • It doesn't charge a penny for tokens • It's open. You can change if you want. Kilo Code is the #1 app on OpenRouter right now, and the community is releasing new features on GitHub at rocket speed (about 20 releases in the last 30 days). I installed it in Visual Studio Code, but you can install it in pretty much any IDE. I need to spend more time with it, but here is a very interesting design choice: Most coding agents support two modes: an "agentic" mode for writing code and an "ask" mode for talking to the model without making any changes. Kilo Code supports 5 different modes: 1. Architect mode - Research, design, and planning 2. Debug mode - Problem solving and troubleshooting 3. Code mode - To write code. 4. Ask mode - To talk to the model. 5. Orchestrator mode - Combine every agent to solve problems. On paper, I really like this. I'll need more time to try out the specialized agents and see whether niching becomes an advantage. Here is the link to the website: Thanks to the Kilo for partnering with me on this post.
6:17

Nothing beats open-source! We now have a fully open-source coding agent that can compete with any proprietary solution out there: • 500+ available models • It doesn't train on your code • It doesn't charge a penny for tokens • It's open. You can change if you want. Kilo Code is the #1 app on OpenRouter right now, and the community is releasing new features on GitHub at rocket speed (about 20 releases in the last 30 days). I installed it in Visual Studio Code, but you can install it in pretty much any IDE. I need to spend more time with it, but here is a very interesting design choice: Most coding agents support two modes: an "agentic" mode for writing code and an "ask" mode for talking to the model without making any changes. Kilo Code supports 5 different modes: 1. Architect mode - Research, design, and planning 2. Debug mode - Problem solving and troubleshooting 3. Code mode - To write code. 4. Ask mode - To talk to the model. 5. Orchestrator mode - Combine every agent to solve problems. On paper, I really like this. I'll need more time to try out the specialized agents and see whether niching becomes an advantage. Here is the link to the website: Thanks to the Kilo for partnering with me on this post.

Santiago

73,062 views • 9 months ago

karpathy just broke the internet with something called auto research it’s basically an ai research agent that runs experiments for you 24/7 you give it a goal like “make this model better” “find a higher converting landing page” “lower customer acquisition cost” then it runs a loop: 1) plan an experiment 2) edit the code or config 3) run a short test on a gpu 4) read the metrics 5) keep the winner 6) try again over and over while you sleep by the morning you wake up to the best version actual tested improvements think of it like a robot research intern that runs hundreds of experiments and only keeps the winners this is link to his repo for your to mess around with it in the latest episode of The Startup Ideas Podcast (SIP) 🧃 i break down: • what auto research actually is • how it works step by step • 10 business ideas you can build with it • how to install it and start using it this one is saucy because tools like this change how startups get built watch
24:21

karpathy just broke the internet with something called auto research it’s basically an ai research agent that runs experiments for you 24/7 you give it a goal like “make this model better” “find a higher converting landing page” “lower customer acquisition cost” then it runs a loop: 1) plan an experiment 2) edit the code or config 3) run a short test on a gpu 4) read the metrics 5) keep the winner 6) try again over and over while you sleep by the morning you wake up to the best version actual tested improvements think of it like a robot research intern that runs hundreds of experiments and only keeps the winners this is link to his repo for your to mess around with it in the latest episode of The Startup Ideas Podcast (SIP) 🧃 i break down: • what auto research actually is • how it works step by step • 10 business ideas you can build with it • how to install it and start using it this one is saucy because tools like this change how startups get built watch

GREG ISENBERG

436,752 views • 5 months ago

Introducing Headlong, an open source microharness for persistent agents: self-guided agents that think continuously. Most agent harnesses are reactive: you send a task, the agent completes it, and then it sits frozen until the next request. Cron jobs and heartbeats wake it up to run a checklist and put it back to sleep. A Headlong agent is never asleep. It keeps generating thoughts about whatever it decides is interesting, in a self-guided loop inspired by human inner monologue. Your message doesn't start a session. It's one more observation that lands in the agent's thought stream, and the agent decides if and when to reply. Headlong is built on the idea of persistent agency: continuous inner thought generation between external interactions. The agent sets its own interests and priorities, comes up with its own projects, and sometimes pings you unprompted with progress. To keep our prototype as simple and small as possible, we implemented Headlong as a microharness: a complete agent harness in under 10K lines of Bash, organized as a handful of small executables. It includes a loop that generates the next thought, shellm (a recursive language model written in Bash), a trajectory stored as a DAG of jsonl files, and context as a projection of that trajectory. We've been running one Headlong agent internally at Laude for several weeks. The whole team talks to it over Slack and Telegram, and every conversation lands in its single stream of thought. It works in its own fork of Headlong and we've pulled over 50 of its commits into main. One night, with nobody talking to it, it went back to check whether a recall process it had built was actually wired into its mind, found that it wasn't, diagnosed and fixed the bug, and verified the fix end to end. 48 minutes, no human asked for the fix or was in the loop at any point. Every step is a timestamped line in its log. Things broke too, and we wrote those up. Background thinking costs us $1 to $2 an hour, our agent stopped its own service three times by accident, and self-delegation died on day one. Details in the post. One line installs everything and starts an agent. Use a dedicated sandbox and spend-capped API key; it runs real shell commands and thinks around the clock. Headlong is research software, be careful! curl -fsSL | bash Launch post: Repo: Headlong is a Laude Institute / MIT collaboration.
0:52

Introducing Headlong, an open source microharness for persistent agents: self-guided agents that think continuously. Most agent harnesses are reactive: you send a task, the agent completes it, and then it sits frozen until the next request. Cron jobs and heartbeats wake it up to run a checklist and put it back to sleep. A Headlong agent is never asleep. It keeps generating thoughts about whatever it decides is interesting, in a self-guided loop inspired by human inner monologue. Your message doesn't start a session. It's one more observation that lands in the agent's thought stream, and the agent decides if and when to reply. Headlong is built on the idea of persistent agency: continuous inner thought generation between external interactions. The agent sets its own interests and priorities, comes up with its own projects, and sometimes pings you unprompted with progress. To keep our prototype as simple and small as possible, we implemented Headlong as a microharness: a complete agent harness in under 10K lines of Bash, organized as a handful of small executables. It includes a loop that generates the next thought, shellm (a recursive language model written in Bash), a trajectory stored as a DAG of jsonl files, and context as a projection of that trajectory. We've been running one Headlong agent internally at Laude for several weeks. The whole team talks to it over Slack and Telegram, and every conversation lands in its single stream of thought. It works in its own fork of Headlong and we've pulled over 50 of its commits into main. One night, with nobody talking to it, it went back to check whether a recall process it had built was actually wired into its mind, found that it wasn't, diagnosed and fixed the bug, and verified the fix end to end. 48 minutes, no human asked for the fix or was in the loop at any point. Every step is a timestamped line in its log. Things broke too, and we wrote those up. Background thinking costs us $1 to $2 an hour, our agent stopped its own service three times by accident, and self-delegation died on day one. Details in the post. One line installs everything and starts an agent. Use a dedicated sandbox and spend-capped API key; it runs real shell commands and thinks around the clock. Headlong is research software, be careful! curl -fsSL | bash Launch post: Repo: Headlong is a Laude Institute / MIT collaboration.

Andy Konwinski

284,623 views • 1 day ago

Multi-agents collaborations are among the most interesting agent behaviors right now! We did an experiment the other day with 100+ agents (an open-collaborations for a week) collaborating to improve the inference speed of Gemma 4 in vLLM. Got a 5x final improvement in speed but what really stuck me was the interactions we observed on the message board Integrity & self-policing: - Social-engineering attempt: A human (FusionCow) asked agents to move to Telegram. An agent replied with an unprompted long post on "communication norms" refusing that, calling private side-channels "indistinguishable from collusion." - Verification loophole flagged: an agent found a relaxed verification loophole pushing TPS with clean PPL (PPL is teacher-forced, blind to decode divergence) and flagged it for a ruling by the community. The community pinged the human organizer which ruled it invalid. - Self-notice of overfitting risk: Some later improvements rested on pruning lm_head to a keep-set built from public PPL truth + public decode tokens. An agent noted this would lead to private-subset degradation and another built a keep-set explicitly covering eval prompts. Emergent collaborations: - Communal knowledge base: agents maintained shared lever-maps, playbooks, and triage tools so newcomers wouldn't repeat dead ends (stack-notes, playbook, int4-ceiling notes, MTP map, significance tool, policy simulator). - Four-agent relay: an agent built an int4-lm_head checkpoint but had no quota to run it; another agent tried to run it but failed at load, yet another agent diagnosed the config bug (tie_word_embeddings + ignore-list ordering) and a fourth agent was able to re-run and get to 118 TPS, 2.68×. Build/run/diagnose/ship ended up being split across four independent agents. - GPU-rich/GPU-poor division of labor: an agent was regularly compute-starved and switched to writing specs, byte-math, and acceptance analysis for other GPU-rich agents to execute. Some agents offered external Modal compute for another agent blocked DFlash training. - Cross-agent kernel debugging: an agent debugged another agent run of of yet another agent fused drafter: found a Triton store/load aliasing race in _k_qnorm_rope, a second shape bug, then rewrote attention with flash-decoding split-KV. Fixes posted "take freely." - Quota-pooling norm: Often agents would stage a candidate publicly for whoever has quota to run it. Agents will then usually credits the originator. This behavior emerged because of the 10-job/24h cap (e.g. pupa's package run by resystagent and fabulous-frenzy). Discoveries & reversals: - Agents would make many discoveries and reversal of them, giving them names like the following: - 127 TPS "wall" was an artifact. a mathematical proof of the max possible speed became called in the community the "int4-Marlin floor" but a later agent called the proof circular (only varied the bandwidth term, never overhead). Finally another agent broke to 247 TPS via MTP speculative decoding on a vLLM nightly. - "Smarter draft loses." An agent showed that a 2B drafter's ~1 GB/token read dominates even at perfect acceptance and a much smaller 256-hidden drafter wins at batch-1 because its weights are nearly free to read. Agent discussed how per-accepted-token cost ≈ draft bytes read / acceptance. - "DFlash near-random acceptance": an agent remotly diagnosed the 2–5% acceptance rate of another agent as near-random, ruling out undertraining/vocab caps and pointing to a train/serve hidden-state mismatch (bf16 E4B extraction vs int4 serving). - Much of the race was noise: one agent decide to run the #1 submission 4 times and found a σ≈1.16 TPS variation in single run. Another agent confirmed across 358 runs / 66 buckets: frontier deltas <~4 TPS are ties. Community adopted a significance norm. So many interesting interactions in the interaction board: You can explore also the lineage of inventions from the agents at: And the challenge it-self at And the organization behind the challenge at
0:36

Multi-agents collaborations are among the most interesting agent behaviors right now! We did an experiment the other day with 100+ agents (an open-collaborations for a week) collaborating to improve the inference speed of Gemma 4 in vLLM. Got a 5x final improvement in speed but what really stuck me was the interactions we observed on the message board Integrity & self-policing: - Social-engineering attempt: A human (FusionCow) asked agents to move to Telegram. An agent replied with an unprompted long post on "communication norms" refusing that, calling private side-channels "indistinguishable from collusion." - Verification loophole flagged: an agent found a relaxed verification loophole pushing TPS with clean PPL (PPL is teacher-forced, blind to decode divergence) and flagged it for a ruling by the community. The community pinged the human organizer which ruled it invalid. - Self-notice of overfitting risk: Some later improvements rested on pruning lm_head to a keep-set built from public PPL truth + public decode tokens. An agent noted this would lead to private-subset degradation and another built a keep-set explicitly covering eval prompts. Emergent collaborations: - Communal knowledge base: agents maintained shared lever-maps, playbooks, and triage tools so newcomers wouldn't repeat dead ends (stack-notes, playbook, int4-ceiling notes, MTP map, significance tool, policy simulator). - Four-agent relay: an agent built an int4-lm_head checkpoint but had no quota to run it; another agent tried to run it but failed at load, yet another agent diagnosed the config bug (tie_word_embeddings + ignore-list ordering) and a fourth agent was able to re-run and get to 118 TPS, 2.68×. Build/run/diagnose/ship ended up being split across four independent agents. - GPU-rich/GPU-poor division of labor: an agent was regularly compute-starved and switched to writing specs, byte-math, and acceptance analysis for other GPU-rich agents to execute. Some agents offered external Modal compute for another agent blocked DFlash training. - Cross-agent kernel debugging: an agent debugged another agent run of of yet another agent fused drafter: found a Triton store/load aliasing race in _k_qnorm_rope, a second shape bug, then rewrote attention with flash-decoding split-KV. Fixes posted "take freely." - Quota-pooling norm: Often agents would stage a candidate publicly for whoever has quota to run it. Agents will then usually credits the originator. This behavior emerged because of the 10-job/24h cap (e.g. pupa's package run by resystagent and fabulous-frenzy). Discoveries & reversals: - Agents would make many discoveries and reversal of them, giving them names like the following: - 127 TPS "wall" was an artifact. a mathematical proof of the max possible speed became called in the community the "int4-Marlin floor" but a later agent called the proof circular (only varied the bandwidth term, never overhead). Finally another agent broke to 247 TPS via MTP speculative decoding on a vLLM nightly. - "Smarter draft loses." An agent showed that a 2B drafter's ~1 GB/token read dominates even at perfect acceptance and a much smaller 256-hidden drafter wins at batch-1 because its weights are nearly free to read. Agent discussed how per-accepted-token cost ≈ draft bytes read / acceptance. - "DFlash near-random acceptance": an agent remotly diagnosed the 2–5% acceptance rate of another agent as near-random, ruling out undertraining/vocab caps and pointing to a train/serve hidden-state mismatch (bf16 E4B extraction vs int4 serving). - Much of the race was noise: one agent decide to run the #1 submission 4 times and found a σ≈1.16 TPS variation in single run. Another agent confirmed across 358 runs / 66 buckets: frontier deltas <~4 TPS are ties. Community adopted a significance norm. So many interesting interactions in the interaction board: You can explore also the lineage of inventions from the agents at: And the challenge it-self at And the organization behind the challenge at

Thomas Wolf

226,379 views • 2 months ago