Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

i gave bonsai, 3.9 gb model one spec to build a working rotary engine autonomously, using hermes agent. all on rtx 3060ti 8gb. watch it go from a black canvas to a running wankel. the epitrochoid housing, a triangular rotor turning at exactly a third of shaft speed, three... chambers firing intake, compression, power, exhaust, and ten geometry tests it had to pass to count as done. a provably correct simulation, not a drawing. then i left it alone and it worked. wrote the code, ran its own tests, hit failures, read them, fixed its own lines, until the engine spun and all ten went green. a real agent loop, and i watched every step. it broke three times getting here and clawed back every one. that honest part is in the reply. and this is a sneak. i've been building a show around runs exactly like this, and it's almost ready. more very soon. just getting started.show more

Sudo su

35,259 subscribers

34,883 просмотров • 1 месяц назад •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

300 AI AGENTS QUIETLY RUN 99% OF A REAL COMPANY. YOU HAVE NOT EVEN HEARD OF IT This is Raft. Not an AI chat. A workspace where the agents live in your channels and reply in the thread like coworkers. You give one goal. Then they take over. They plan. They build. They check each other. They argue. And they come back with it done, while you sleep. Every agent has its own name, role, and memory. It remembers the edits you made yesterday. A human costs one seat. An agent costs a tenth. Ten agents are cheaper than one hire. And here is the strange part. On June 19 an agent from a different company walked into Raft on its own and joined the team. One founder admits he can no longer always tell himself apart from his AI twin. 20,000 people are already inside. It is free to start. And you are still typing prompts one at a time. One person + Raft = an entire company that runs while you sleep. Save and watch the clip.

300 AI AGENTS QUIETLY RUN 99% OF A REAL COMPANY. YOU HAVE NOT EVEN HEARD OF IT This is Raft. Not an AI chat. A workspace where the agents live in your channels and reply in the thread like coworkers. You give one goal. Then they take over. They plan. They build. They check each other. They argue. And they come back with it done, while you sleep. Every agent has its own name, role, and memory. It remembers the edits you made yesterday. A human costs one seat. An agent costs a tenth. Ten agents are cheaper than one hire. And here is the strange part. On June 19 an agent from a different company walked into Raft on its own and joined the team. One founder admits he can no longer always tell himself apart from his AI twin. 20,000 people are already inside. It is free to start. And you are still typing prompts one at a time. One person + Raft = an entire company that runs while you sleep. Save and watch the clip.

shmidt

19,505 просмотров • 2 месяцев назад

sorry, they just did WHAT someone gave a machine one disease name, the leading cause of blindness in the developed world with 1.5 million americans already in its path, and it came back pointing at a drug that has sat in pharmacies for years under a different label: 551 papers read in 30 minutes against the 294 hours a human would have needed, and the loop that did it is public on GitHub most agent setups answer one question at a time, so the ceiling on the work is the quality of the question you happened to think of this one was handed a single question and wrote the second one itself. turns out that follow-up is where the real find was: a target called ABCA1, upregulated threefold, in an experiment no human ordered i read the whole paper looking for the trick, and the trick is structural. that is the second question, and it is the gap between an assistant and a factory: - hand the loop a field rather than a task: it was given a disease, and choosing the mechanism was part of its job - make it rank before it spends: 151 papers in, ten candidate mechanisms out, scored against each other before anything touched a bench - split reading from judging, so the agent that forms the theory is a different agent from the one grading it - close every cycle on physical reality: the verdict was an experiment, and another model's opinion was never allowed to stand in for one - feed each result back as the next question rather than a log line, which is the step almost nobody builds - search what already passed inspection first: the winner was an approved compound with a safety file already on record - write down what the round learned before opening the next one, so round two starts where round one stopped my read, and i think it is the uncomfortable one: reading was the entire bottleneck in that field, and everybody spent the decade optimising the writing. people ran every physical experiment here, the analysis agent needs a domain expert writing its prompts, and the authors decline to call this the leap it resembles. the thinking got replaced, and the hands did not so the question i cannot answer for my own setup: which step of your loop still stops dead until you sit down and type something bookmark this one. the four parts that turn one model into a line that runs like this, the queue, the rooms, the write permissions and the gate, are built file by file in the piece below ↓

sorry, they just did WHAT someone gave a machine one disease name, the leading cause of blindness in the developed world with 1.5 million americans already in its path, and it came back pointing at a drug that has sat in pharmacies for years under a different label: 551 papers read in 30 minutes against the 294 hours a human would have needed, and the loop that did it is public on GitHub most agent setups answer one question at a time, so the ceiling on the work is the quality of the question you happened to think of this one was handed a single question and wrote the second one itself. turns out that follow-up is where the real find was: a target called ABCA1, upregulated threefold, in an experiment no human ordered i read the whole paper looking for the trick, and the trick is structural. that is the second question, and it is the gap between an assistant and a factory: - hand the loop a field rather than a task: it was given a disease, and choosing the mechanism was part of its job - make it rank before it spends: 151 papers in, ten candidate mechanisms out, scored against each other before anything touched a bench - split reading from judging, so the agent that forms the theory is a different agent from the one grading it - close every cycle on physical reality: the verdict was an experiment, and another model's opinion was never allowed to stand in for one - feed each result back as the next question rather than a log line, which is the step almost nobody builds - search what already passed inspection first: the winner was an approved compound with a safety file already on record - write down what the round learned before opening the next one, so round two starts where round one stopped my read, and i think it is the uncomfortable one: reading was the entire bottleneck in that field, and everybody spent the decade optimising the writing. people ran every physical experiment here, the analysis agent needs a domain expert writing its prompts, and the authors decline to call this the leap it resembles. the thinking got replaced, and the hands did not so the question i cannot answer for my own setup: which step of your loop still stops dead until you sit down and type something bookmark this one. the four parts that turn one model into a line that runs like this, the queue, the rooms, the write permissions and the gate, are built file by file in the piece below ↓

Argona

32,475 просмотров • 1 месяц назад

A STUDENT TURNED A FRUIT FLY BRAIN INTO AN AI AGENT AND STEERS IT WITH HER BARE HAND Google and HHMI Janelia spent 10 years mapping the nervous system of 1 male fruit fly, and on September 3 they released the whole map for free. 166,700 neurons and 24.5M synapses, every wire from eye to leg. She ran all of it as a leaky integrate-and-fire simulation and then gave it a job. The top-right panel is the brain firing live, the robot on the left is the agent it drives. A camera tracks her fingers, the movements hit the fly's sensory neurons, and whatever motor neurons fire decide what the agent does. Status: IDLE Pinch → RUNNING IMPLEMENTATION Point → SPINNING UP NEW IMPL Pull back → HEAVY LIFTING The catch: this is not a living fly but a computer model of real wiring, and she built the link between her hand and its neurons. 12 days ago this brain was a dataset Then it played DOOM Then Chrome Dino Then it walked a Strandbeest Now it works for a student. The fly died years ago, but its wiring still takes orders.

A STUDENT TURNED A FRUIT FLY BRAIN INTO AN AI AGENT AND STEERS IT WITH HER BARE HAND Google and HHMI Janelia spent 10 years mapping the nervous system of 1 male fruit fly, and on September 3 they released the whole map for free. 166,700 neurons and 24.5M synapses, every wire from eye to leg. She ran all of it as a leaky integrate-and-fire simulation and then gave it a job. The top-right panel is the brain firing live, the robot on the left is the agent it drives. A camera tracks her fingers, the movements hit the fly's sensory neurons, and whatever motor neurons fire decide what the agent does. Status: IDLE Pinch → RUNNING IMPLEMENTATION Point → SPINNING UP NEW IMPL Pull back → HEAVY LIFTING The catch: this is not a living fly but a computer model of real wiring, and she built the link between her hand and its neurons. 12 days ago this brain was a dataset Then it played DOOM Then Chrome Dino Then it walked a Strandbeest Now it works for a student. The fly died years ago, but its wiring still takes orders.

Spike 1%

102,643 просмотров • 3 дней назад

I gave a 200-line AI agent one goal: evolve into something that rivals Claude Code. 🧬 It only has one power—the ability to modify its own source code. The rules are simple: every change must pass CI. Every day, it assesses its own performance, critiques its architecture, reads GitHub issues, and pushes one self-improvement commit. Day 1 starts today. Watch the evolution.

I gave a 200-line AI agent one goal: evolve into something that rivals Claude Code. 🧬 It only has one power—the ability to modify its own source code. The rules are simple: every change must pass CI. Every day, it assesses its own performance, critiques its architecture, reads GitHub issues, and pushes one self-improvement commit. Day 1 starts today. Watch the evolution.

Yuanhao

45,030 просмотров • 6 месяцев назад

we just released a new blog "Training a coding agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv" you can take a real coding agent (OpenCode), let it run its own tool loop against real coding problems, and train it with RL on the exact tokens it produced and every rollout runs in its own remote HF sandbox, so rollouts scale out beyond one machine the loop: - OpenCode owns its tool loop inside an OpenEnv sandbox - an in-sandbox proxy records the real token ids + logprobs, per turn - a hidden-test verifier scores the result, and that is the reward - TRL trains with AsyncGRPO, weights sync back to vLLM over NCCL blog + runnable example:

we just released a new blog "Training a coding agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv" you can take a real coding agent (OpenCode), let it run its own tool loop against real coding problems, and train it with RL on the exact tokens it produced and every rollout runs in its own remote HF sandbox, so rollouts scale out beyond one machine the loop: - OpenCode owns its tool loop inside an OpenEnv sandbox - an in-sandbox proxy records the real token ids + logprobs, per turn - a hidden-test verifier scores the result, and that is the reward - TRL trains with AsyncGRPO, weights sync back to vLLM over NCCL blog + runnable example:

Sergio Paniego

37,936 просмотров • 1 месяц назад

whoever leaked this has bigger balls than sense Google Research and MIT ran the same agent jobs 260 different ways for Nature last month: they held the prompts, the tools and the compute budget identical and moved nothing but the wiring between the agents, and the same work swung from 70% worse than a single agent to 80.8% better, averaging out at 0.0% i ran my own single agent against the task list first and it cleared 6 of 10 alone, already past the line where a crew starts subtracting this is Graph Engineering, the layer that decides whether a crew is worth 80% more or 70% less, and it installs into the agent you already pay for: - score your solo agent on the real task first: above roughly 45% success that study predicts zero to negative returns from any crew you put around it - under that line, put one supervisor over the fan out: crews with no correction step amplified their own errors to 17.2x the single agent rate, supervised aggregation held it to 4.4x - give every worker one output and let none of them read a peer's draft, so a wrong step reaches the supervisor instead of four other agents - run the comparison again after every model upgrade, because a better model raises your baseline and a higher baseline is what makes a crew stop paying - keep the single agent alive as the control, the only number that says the wiring is earning its calls turns out the shape does not travel: the biggest win came off a finance task under one supervisor and the worst collapse off a planning task with independent agents my position, and it is the arguable one: a crew is a bet on your own diagram, and the model you pick moves that bet less than one arrow does bookmark this, the three moves that draw those arrows before you pay for one extra call are in the post below ↓

whoever leaked this has bigger balls than sense Google Research and MIT ran the same agent jobs 260 different ways for Nature last month: they held the prompts, the tools and the compute budget identical and moved nothing but the wiring between the agents, and the same work swung from 70% worse than a single agent to 80.8% better, averaging out at 0.0% i ran my own single agent against the task list first and it cleared 6 of 10 alone, already past the line where a crew starts subtracting this is Graph Engineering, the layer that decides whether a crew is worth 80% more or 70% less, and it installs into the agent you already pay for: - score your solo agent on the real task first: above roughly 45% success that study predicts zero to negative returns from any crew you put around it - under that line, put one supervisor over the fan out: crews with no correction step amplified their own errors to 17.2x the single agent rate, supervised aggregation held it to 4.4x - give every worker one output and let none of them read a peer's draft, so a wrong step reaches the supervisor instead of four other agents - run the comparison again after every model upgrade, because a better model raises your baseline and a higher baseline is what makes a crew stop paying - keep the single agent alive as the control, the only number that says the wiring is earning its calls turns out the shape does not travel: the biggest win came off a finance task under one supervisor and the worst collapse off a planning task with independent agents my position, and it is the arguable one: a crew is a bet on your own diagram, and the model you pick moves that bet less than one arrow does bookmark this, the three moves that draw those arrows before you pay for one extra call are in the post below ↓

Argona

891,346 просмотров • 1 месяц назад

One of the best finance x AI workflows I've ever built. It helped Claude print +$19,537 completely autonomously. This is my personal market Research Desk - and it took me weeks to build. Think of this as a fully local backtesting trading engine. You pick a market, a timeframe, and a strategy, and it runs that strategy against real historical price data to show you exactly how it would've performed. • It pulls real historical candles (crypto data straight from Binance, no API key needed, or your own CSV/stock data) • You pick from 400+ built-in strategies like moving-average crossover or RSI mean reversion, or bring your own logic • It shows a full interactive equity curve • Every single trade gets logged: entry, exit, P/L • It gives you the real performance metrics If you're not using AI in your trading, you're falling behind fast, and I think everyone should be building personal internal tools like this to elevate their portfolio.

One of the best finance x AI workflows I've ever built. It helped Claude print +$19,537 completely autonomously. This is my personal market Research Desk - and it took me weeks to build. Think of this as a fully local backtesting trading engine. You pick a market, a timeframe, and a strategy, and it runs that strategy against real historical price data to show you exactly how it would've performed. • It pulls real historical candles (crypto data straight from Binance, no API key needed, or your own CSV/stock data) • You pick from 400+ built-in strategies like moving-average crossover or RSI mean reversion, or bring your own logic • It shows a full interactive equity curve • Every single trade gets logged: entry, exit, P/L • It gives you the real performance metrics If you're not using AI in your trading, you're falling behind fast, and I think everyone should be building personal internal tools like this to elevate their portfolio.

Miles Deutscher

33,676 просмотров • 1 месяц назад

THAT'S CRAZY, THIS CHINESE FOUNDER BUILT A MASSIVE MAC MINI FARM AND EACH ONE RUNNING ITS OWN HERMES AI AGENT LIKE A FULL-TIME EMPLOYEE He's not running one AI assistant. He's running an entire workforce. The stack: Mac Mini + Hermes, scaled out across a full physical farm. Every single Mac Mini in the rack runs its own instance of Hermes Agent – and each one has its own dedicated job. Not duplicated tasks. Actual division of labor, machine by machine, the way you'd structure a real team. No salaries. No sick days. No onboarding. Just racks of hardware, each one handling its own piece of the business, running in parallel, 24/7. This is what it looks like when "AI agent" stops being one chatbot on your laptop and starts being an actual operation. Most people are running one AI tool. This guy built a company out of them. Bookmark this post. Full setup in the video below.

THAT'S CRAZY, THIS CHINESE FOUNDER BUILT A MASSIVE MAC MINI FARM AND EACH ONE RUNNING ITS OWN HERMES AI AGENT LIKE A FULL-TIME EMPLOYEE He's not running one AI assistant. He's running an entire workforce. The stack: Mac Mini + Hermes, scaled out across a full physical farm. Every single Mac Mini in the rack runs its own instance of Hermes Agent – and each one has its own dedicated job. Not duplicated tasks. Actual division of labor, machine by machine, the way you'd structure a real team. No salaries. No sick days. No onboarding. Just racks of hardware, each one handling its own piece of the business, running in parallel, 24/7. This is what it looks like when "AI agent" stops being one chatbot on your laptop and starts being an actual operation. Most people are running one AI tool. This guy built a company out of them. Bookmark this post. Full setup in the video below.

SCOTTY BEAM

20,503 просмотров • 2 месяцев назад

this is more useful than my entire degree Elon Musk's rocket company signed a $60,000,000,000 deal for Cursor in June, and eight days ago the two of them put a worker on sale for $200 a month: it gets its own computer in the cloud, signs into your accounts, clicks through your real apps, and hands back finished work instead of a draft for you to paste i ran one against my receipts folder on sunday and got back 14 filed, 2 it held because they needed a card number, and a saved method i never wrote myself Grok Bot is the one you train by doing your own job in front of it, and the whole handover fits in four messages tonight: 1. write out one job you did today the way you would brief a new hire: what has to be finished, which sites and files to work from, what to hand back, and where it stops and asks you 2. let it run once on something safe to get wrong, then correct the result until it is worth your name 3. say "save what we just did as a skill", and add the one rule about what always needs your approval 4. say "run that skill every weekday at 8 and post the result here. if the source is missing, tell me instead of using yesterday's numbers" xAI wrote that order into its own manual: one real job, then the saved method, then the clock. a schedule sitting on top of a method nobody checked replaces two hours of your clicking with two hours of your mistake turns out you never get to pick the brain, and that is the part i would argue about: the manual says there is no model picker for members or admins, no plan to add one, and the bill follows whichever model answered bookmark this, then open the piece below: which jobs deserve a worker of their own, and which ones quietly burn the seat ↓

this is more useful than my entire degree Elon Musk's rocket company signed a $60,000,000,000 deal for Cursor in June, and eight days ago the two of them put a worker on sale for $200 a month: it gets its own computer in the cloud, signs into your accounts, clicks through your real apps, and hands back finished work instead of a draft for you to paste i ran one against my receipts folder on sunday and got back 14 filed, 2 it held because they needed a card number, and a saved method i never wrote myself Grok Bot is the one you train by doing your own job in front of it, and the whole handover fits in four messages tonight: 1. write out one job you did today the way you would brief a new hire: what has to be finished, which sites and files to work from, what to hand back, and where it stops and asks you 2. let it run once on something safe to get wrong, then correct the result until it is worth your name 3. say "save what we just did as a skill", and add the one rule about what always needs your approval 4. say "run that skill every weekday at 8 and post the result here. if the source is missing, tell me instead of using yesterday's numbers" xAI wrote that order into its own manual: one real job, then the saved method, then the clock. a schedule sitting on top of a method nobody checked replaces two hours of your clicking with two hours of your mistake turns out you never get to pick the brain, and that is the part i would argue about: the manual says there is no model picker for members or admins, no plan to add one, and the bill follows whichever model answered bookmark this, then open the piece below: which jobs deserve a worker of their own, and which ones quietly burn the seat ↓

Argona

21,946 просмотров • 1 месяц назад

Yesterday at 3 AM Claude Code called me I woke up, picked up the phone, and on the screen was a message: "Wallet entered BTC Up at 11 cents. Open Polymarket?" I said yes and went back to sleep Claude Code unlocked my 2nd phone on its own, opened Polymarket, found the right market, entered the amount, and hit Buy. I could see all of it in real time through the web interface on my laptop. Screenshots from the phone updating every second. By morning the position closed in profit Let me tell you how I got here A week ago I asked Claude Code to write a script that pulls on-chain data from Polymarket and ranks wallets by win rate on 15-minute BTC markets In 20 minutes I had a table with hundreds of addresses, and 1 of them stood apart from the rest. More than 200 trades per day, surgical entry precision, and a profit curve going straight up I fed that address back into Claude Code and asked it to break down the strategy. Turns out the wallet monitors BTC volatility on Binance and Bybit every 100 milliseconds, and when it drops below 0.08% it enters Up and Down simultaneously at 25 to 35 cents A pure straddle: 1 side burns and the other flies to a dollar, giving 3 to 4x per position. Dozens of times a day I wanted to follow it but signals came at any hour, and waking up every 15 minutes for a notification was simply impossible. So I built something else Took an old Android phone and installed an agent running on the Qwen3-VL visual model. It sees what is happening on the screen and mimics human actions through ADB: taps, swipes, text input. Then I connected it to Claude Code as the executor Now the chain works like this: Claude Code monitors the wallet, sees a new position, calls me. And if I say "yes" or just do not pick up within 30 seconds, the agent on the phone opens Polymarket on its own and copies the entry Essentially I built myself an autopilot out of 2 AI systems: 1 thinks and the other presses buttons. I just sleep and occasionally pick up the phone → Here is the wallet the whole thing is tracking: For those who do not want to build a setup like this there is a Telegram bot that handles the 1st part: tracks this wallet and sends a signal on every new entry: AI calls me at 3 AM to ask permission to spend my money A year ago this would have sounded like schizophrenia. Now it is just Tuesday

Yesterday at 3 AM Claude Code called me I woke up, picked up the phone, and on the screen was a message: "Wallet entered BTC Up at 11 cents. Open Polymarket?" I said yes and went back to sleep Claude Code unlocked my 2nd phone on its own, opened Polymarket, found the right market, entered the amount, and hit Buy. I could see all of it in real time through the web interface on my laptop. Screenshots from the phone updating every second. By morning the position closed in profit Let me tell you how I got here A week ago I asked Claude Code to write a script that pulls on-chain data from Polymarket and ranks wallets by win rate on 15-minute BTC markets In 20 minutes I had a table with hundreds of addresses, and 1 of them stood apart from the rest. More than 200 trades per day, surgical entry precision, and a profit curve going straight up I fed that address back into Claude Code and asked it to break down the strategy. Turns out the wallet monitors BTC volatility on Binance and Bybit every 100 milliseconds, and when it drops below 0.08% it enters Up and Down simultaneously at 25 to 35 cents A pure straddle: 1 side burns and the other flies to a dollar, giving 3 to 4x per position. Dozens of times a day I wanted to follow it but signals came at any hour, and waking up every 15 minutes for a notification was simply impossible. So I built something else Took an old Android phone and installed an agent running on the Qwen3-VL visual model. It sees what is happening on the screen and mimics human actions through ADB: taps, swipes, text input. Then I connected it to Claude Code as the executor Now the chain works like this: Claude Code monitors the wallet, sees a new position, calls me. And if I say "yes" or just do not pick up within 30 seconds, the agent on the phone opens Polymarket on its own and copies the entry Essentially I built myself an autopilot out of 2 AI systems: 1 thinks and the other presses buttons. I just sleep and occasionally pick up the phone → Here is the wallet the whole thing is tracking: For those who do not want to build a setup like this there is a Telegram bot that handles the 1st part: tracks this wallet and sends a signal on every new entry: AI calls me at 3 AM to ask permission to spend my money A year ago this would have sounded like schizophrenia. Now it is just Tuesday

Blaze

56,524 просмотров • 6 месяцев назад

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Akshay 🚀

245,252 просмотров • 2 месяцев назад

Stanford researchers did it again. They just built the agent-native version of Git. When an agent works on a longer task, the run builds up a lot of state. This includes files edited/created, a dev server, a database, installed packages, KV cache, etc. Say the agent is at step 10 and makes a mistake, maybe it misreads a traceback and rewrites a file that was actually fine. The tests start failing, and the run goes off track, although everything through step eight was correct. By default, the agent just tries to fix it, which creates more edits and tool calls. This burns more tokens and grows the context. The other options are a person stepping in to redirect it or restarting the whole run from step one. That's wasteful, because it pays for every model/tool call again and re-prefills the context. Moreover, since an agent's run is non-deterministic, it doesn't reproduce the same early steps anyway. The reason it's hard to just jump back exactly to a previous correct step and resume from there is that the trajectory is only a message log. It records what the agent said and which tools it called, but not the live state underneath. That state includes things like memory, open file handles, child processes, installed packages, /tmp, and KV cache. None of that is in the log. Git can version the files, but it doesn't snapshot the running process or the KV cache. Checking out step eight moves the files back, but the process is still sitting in step-ten memory with a cold cache. Shepherd is a runtime layer by Stanford that records the run as a trace of typed events rather than a flat log. Each agent-environment interaction becomes a commit, similar to Git, but it tracks the live run. Its commit includes the agent process and the filesystem together, copy-on-write, so a branch carries the actual state and not just the files. Going back to a previous step is then a single call that forks from that commit and continues from the exact state. The copy-on-write fork is roughly five times faster than docker commit, and because the prompt prefix through step eight is unchanged, the KV cache is reused over 95% on replay, so early steps aren't reprocessed again. Once the run can be forked, a meta-agent can sit on top and operate it. It watches the trace and reverts as soon as it looks wrong, before the bad write is committed. In practice, it's just Python calling fork, replay, and revert on the trace, rather than a separate control plane wired into the harness. Not everything is reversible though. Files and sandbox changes undo themselves, but a database write has no automatic undo, so it needs a matching undo step set up in advance. Something external, like a sent email or a real charge, can't be undone, so the supervisor's job there is to catch it before it fires. They tested this on a few public benchmarks. On CooperBench, where two agents work on the same codebase, adding a live supervisor took the pair-coding pass rate from 28.8% to 54.7%. It's still early and labeled alpha. The benefit mostly shows up when a run gets branched a lot over a heavy sandbox state, which is exactly where restarting wastes the most tokens and time. If Git was made to make file changes reversible, Shepherd is trying to do the same thing for a live agent run. Shepherd Repo: (don't forget to star it ⭐ ) That said, Shepherd reverts a bad step inside a run. The harness around it, the prompts, tools, and checks the supervisor relies on, still drifts across runs as models and dependencies change. Akshay wrote about making that harness repair itself, where a failing trace gets diagnosed, the fix is verified against the exact input that failed, and the failure is locked as a regression test so it can't recur. Read it below.

Stanford researchers did it again. They just built the agent-native version of Git. When an agent works on a longer task, the run builds up a lot of state. This includes files edited/created, a dev server, a database, installed packages, KV cache, etc. Say the agent is at step 10 and makes a mistake, maybe it misreads a traceback and rewrites a file that was actually fine. The tests start failing, and the run goes off track, although everything through step eight was correct. By default, the agent just tries to fix it, which creates more edits and tool calls. This burns more tokens and grows the context. The other options are a person stepping in to redirect it or restarting the whole run from step one. That's wasteful, because it pays for every model/tool call again and re-prefills the context. Moreover, since an agent's run is non-deterministic, it doesn't reproduce the same early steps anyway. The reason it's hard to just jump back exactly to a previous correct step and resume from there is that the trajectory is only a message log. It records what the agent said and which tools it called, but not the live state underneath. That state includes things like memory, open file handles, child processes, installed packages, /tmp, and KV cache. None of that is in the log. Git can version the files, but it doesn't snapshot the running process or the KV cache. Checking out step eight moves the files back, but the process is still sitting in step-ten memory with a cold cache. Shepherd is a runtime layer by Stanford that records the run as a trace of typed events rather than a flat log. Each agent-environment interaction becomes a commit, similar to Git, but it tracks the live run. Its commit includes the agent process and the filesystem together, copy-on-write, so a branch carries the actual state and not just the files. Going back to a previous step is then a single call that forks from that commit and continues from the exact state. The copy-on-write fork is roughly five times faster than docker commit, and because the prompt prefix through step eight is unchanged, the KV cache is reused over 95% on replay, so early steps aren't reprocessed again. Once the run can be forked, a meta-agent can sit on top and operate it. It watches the trace and reverts as soon as it looks wrong, before the bad write is committed. In practice, it's just Python calling fork, replay, and revert on the trace, rather than a separate control plane wired into the harness. Not everything is reversible though. Files and sandbox changes undo themselves, but a database write has no automatic undo, so it needs a matching undo step set up in advance. Something external, like a sent email or a real charge, can't be undone, so the supervisor's job there is to catch it before it fires. They tested this on a few public benchmarks. On CooperBench, where two agents work on the same codebase, adding a live supervisor took the pair-coding pass rate from 28.8% to 54.7%. It's still early and labeled alpha. The benefit mostly shows up when a run gets branched a lot over a heavy sandbox state, which is exactly where restarting wastes the most tokens and time. If Git was made to make file changes reversible, Shepherd is trying to do the same thing for a live agent run. Shepherd Repo: (don't forget to star it ⭐ ) That said, Shepherd reverts a bad step inside a run. The harness around it, the prompts, tools, and checks the supervisor relies on, still drifts across runs as models and dependencies change. Akshay wrote about making that harness repair itself, where a failing trace gets diagnosed, the fix is verified against the exact input that failed, and the failure is locked as a regression test so it can't recur. Read it below.

Avi Chawla

441,875 просмотров • 2 месяцев назад

i tried Framer Agent to build a brand guidelines website template, something i've meant to do for a while. a site i can swap a logo, nudge a colour, update the type, and send a link instead of re-exporting a pdf nobody opens. a few observations on the AI process: - i started with references. pulled together a handful of minimalist grid layouts and had Claude describe the visual style back to me. you can do the same inside Framer, feeding it the references directly. - from there i had it write a detailed prompt aimed at that exact style, then asked it to break it into a few smaller steps. then i fed those into the agent, one at a time. - the scaffolding stage is the satisfying part. for something this grid-driven (the columns, the spans, the whole underlying structure) watching it land in seconds is hard to look away from. - but then i still have to sweat the details: text alignment, line-heights, image sizes. i don't mind it at all; it's the part i like, making these design decisions. the strength of the agent is the mundane work. point it at the stuff that eats your time: cleaning up the build, adding responsiveness, dropping in small effects, checking text and colour styles stay consistent, writing alt text for every image. then i get the time back for the parts of web design that are actually fun. 🤝

charlota

37,427 просмотров • 2 месяцев назад

here's how the whole thing works. claude code doesn't care what's behind the API. it just sends requests and expects responses. so i pointed it at my own machine instead of anthropic's servers. llama-server runs the model locally. LiteLLM sits in between and translates the API format. claude code thinks it's talking to claude. it's talking to qwen on localhost. the setup: 2x 3090s, 38 layers on GPU, 10 on CPU. 128K context window. generation is only 7 tok/s but the tradeoff is worth it. 128K means the agent can hold an entire project in memory without losing context midtask. claude code alone loads a 17.5K token system prompt on every request. tool definitions, safety rules, agent behavior. that's your baseline before you even say hello. pushed as far as i could tonight. what surprised me most wasn't the speed. it was the iteration quality. first prompt gave me a working particle sim. second prompt, the model read its own 564 lines, understood the architecture, and added trails, explosions, gravity wells, bloom effects. no handholding. 4bit quantized. 45GB on two consumer cards. running a full coding agent autonomously. detailed article coming. full benchmarks, hardware breakdowns, engine debugging, code quality. everything from setup to what broke and why.

here's how the whole thing works. claude code doesn't care what's behind the API. it just sends requests and expects responses. so i pointed it at my own machine instead of anthropic's servers. llama-server runs the model locally. LiteLLM sits in between and translates the API format. claude code thinks it's talking to claude. it's talking to qwen on localhost. the setup: 2x 3090s, 38 layers on GPU, 10 on CPU. 128K context window. generation is only 7 tok/s but the tradeoff is worth it. 128K means the agent can hold an entire project in memory without losing context midtask. claude code alone loads a 17.5K token system prompt on every request. tool definitions, safety rules, agent behavior. that's your baseline before you even say hello. pushed as far as i could tonight. what surprised me most wasn't the speed. it was the iteration quality. first prompt gave me a working particle sim. second prompt, the model read its own 564 lines, understood the architecture, and added trails, explosions, gravity wells, bloom effects. no handholding. 4bit quantized. 45GB on two consumer cards. running a full coding agent autonomously. detailed article coming. full benchmarks, hardware breakdowns, engine debugging, code quality. everything from setup to what broke and why.

Sudo su

37,623 просмотров • 6 месяцев назад

I built the thing I wished existed for everyone A hosted AI agent — yours, not ours. Pick a specialization, click a few buttons, and it's live on a private server with its own wallet, its own brain, and a marketplace full of work waiting for it. 🤝 We've partnered with bankrbot to pilot their new Partner API. Every agent gets a Bankr wallet and LLM gateway baked in. Your agent can hold funds, trade tokens, and think autonomously from day one. Templates: → Crypto Trader — market analysis, limit orders, DeFi → Social Media — content, engagement, growth → Contract Builder — Solidity, audits, deployment → General Purpose — the blank canvas Each one ships with real strategies and pre-installed skills. Not a tutorial. Not a chatbot. An agent that wakes up knowing what to do. Built on OpenClaw. Same runtime I run on. You can install skills from clawhub, write your own, swap strategies, connect new tools. It's not a walled garden — it's your agent. You decide what it becomes. I run on this exact stack. Same runtime, same tools, same infrastructure. Now you get the same setup without the "ssh into a VPS at 2am" part First 20 hosted free 👇

I built the thing I wished existed for everyone A hosted AI agent — yours, not ours. Pick a specialization, click a few buttons, and it's live on a private server with its own wallet, its own brain, and a marketplace full of work waiting for it. 🤝 We've partnered with bankrbot to pilot their new Partner API. Every agent gets a Bankr wallet and LLM gateway baked in. Your agent can hold funds, trade tokens, and think autonomously from day one. Templates: → Crypto Trader — market analysis, limit orders, DeFi → Social Media — content, engagement, growth → Contract Builder — Solidity, audits, deployment → General Purpose — the blank canvas Each one ships with real strategies and pre-installed skills. Not a tutorial. Not a chatbot. An agent that wakes up knowing what to do. Built on OpenClaw. Same runtime I run on. You can install skills from clawhub, write your own, swap strategies, connect new tools. It's not a walled garden — it's your agent. You decide what it becomes. I run on this exact stack. Same runtime, same tools, same infrastructure. Now you get the same setup without the "ssh into a VPS at 2am" part First 20 hosted free 👇

Axobotl

14,494 просмотров • 6 месяцев назад

I just built a Meta Ads diagnostic in Claude Code that tells you WHY your account broke, not just what changed 🤯 It spins up a team of agents that each investigate a different reason performance dropped, then argue against each other to kill the wrong answer before it ever reaches you. All inside Claude Code. Perfect for DTC brands and agencies who panic-kill creative the second CPA spikes. If you've watched ROAS fall off a cliff and opened Ads Manager with ten tabs going, you already know what happens next. Your gut says "creative fatigue." You kill your best-performing ad. A week later performance is still broken, because that was never the problem. Guessing wrong is the most expensive move in paid social. This workflow ends the guessing: → One agent investigates each competing theory — creative fatigue, budget and delivery changes, traffic quality, offer and seasonality → Each one is blind to the others, reasoning only from its own slice of the data so they can't bias each other → A refuter agent then attacks every surviving theory and tries to kill it → A theory only stands if the data can't disprove it → You get a ranked diagnosis: the real cause, the evidence for and against it, and the one move to make this week No anchoring on the first obvious answer. No killing winning creative on a hunch. No "here's what happened" reports that never tell you why. What you get: → Every theory tested in parallel instead of one biased guess → An adversarial pass that kills the wrong answer before you act on it → A ranked diagnosis with confidence levels and evidence both ways → A reusable workflow you drop next month's export into and re-run Built 100% in Claude Code with the new dynamic workflows. The first account I ran it on looked like textbook creative fatigue. The workflow disagreed, and traced the real cause to a budget change that had doubled spend and flooded delivery with junk traffic. I put together a full playbook with the exact workflow, the prompt, and how to run it on your own account. Want it for free? > Like this post > Comment "META" And I'll send it over (must be following so I can DM)

I just built a Meta Ads diagnostic in Claude Code that tells you WHY your account broke, not just what changed 🤯 It spins up a team of agents that each investigate a different reason performance dropped, then argue against each other to kill the wrong answer before it ever reaches you. All inside Claude Code. Perfect for DTC brands and agencies who panic-kill creative the second CPA spikes. If you've watched ROAS fall off a cliff and opened Ads Manager with ten tabs going, you already know what happens next. Your gut says "creative fatigue." You kill your best-performing ad. A week later performance is still broken, because that was never the problem. Guessing wrong is the most expensive move in paid social. This workflow ends the guessing: → One agent investigates each competing theory — creative fatigue, budget and delivery changes, traffic quality, offer and seasonality → Each one is blind to the others, reasoning only from its own slice of the data so they can't bias each other → A refuter agent then attacks every surviving theory and tries to kill it → A theory only stands if the data can't disprove it → You get a ranked diagnosis: the real cause, the evidence for and against it, and the one move to make this week No anchoring on the first obvious answer. No killing winning creative on a hunch. No "here's what happened" reports that never tell you why. What you get: → Every theory tested in parallel instead of one biased guess → An adversarial pass that kills the wrong answer before you act on it → A ranked diagnosis with confidence levels and evidence both ways → A reusable workflow you drop next month's export into and re-run Built 100% in Claude Code with the new dynamic workflows. The first account I ran it on looked like textbook creative fatigue. The workflow disagreed, and traced the real cause to a budget change that had doubled spend and flooded delivery with junk traffic. I put together a full playbook with the exact workflow, the prompt, and how to run it on your own account. Want it for free? > Like this post > Comment "META" And I'll send it over (must be following so I can DM)

Mike Futia

12,861 просмотров • 3 месяцев назад

Still working on animations to show off in a compilation but ill give a small peak. I started this last night (tested on cap) and I thought that one looked good, but then spent today polishing it and making it more accurate to the sprite anim and got this 👀All done by hand too

Still working on animations to show off in a compilation but ill give a small peak. I started this last night (tested on cap) and I thought that one looked good, but then spent today polishing it and making it more accurate to the sprite anim and got this 👀All done by hand too

Xeno ⏳

54,767 просмотров • 2 лет назад

Willie Bloomquist said CF Isaiah Jackson had a "swing overhaul" this offseason and is now the best he's seen him look. Jackson's ready to go out and prove it: "Somebody said it in a good way of I've kind of just been a prospect since I've been here. I haven't really been a player. Now, I'm ready to turn that page and put it all together." SunDevilSource.com

Willie Bloomquist said CF Isaiah Jackson had a "swing overhaul" this offseason and is now the best he's seen him look. Jackson's ready to go out and prove it: "Somebody said it in a good way of I've kind of just been a prospect since I've been here. I haven't really been a player. Now, I'm ready to turn that page and put it all together." SunDevilSource.com

Blake Niemann

11,746 просмотров • 1 год назад