正在加载视频...

视频加载失败

返回首页

2 hours 43 minutes. 442 steps. 5,406 lines of code. The page still does not load 💀 I gave Qwen3.8 Flash Next my standard bench prompt: a voxel Japanese pagoda garden in Three.js, 16,500 voxels, modular source, run it and screenshot it when it's done. Other models finish this...

superalesha's profile picture

Alexey Fateev

3,841 subscribers

31,282 次观看 • 14 天前 •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.
4:15

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.

Sudo su

39,719 次观看 • 3 个月前

watch this anon. i gave NVIDIA's biggest model ever a single task. 100 minutes and 440,000 tokens later, it had rendered nothing. not one important thing on the screen. this is Nemotron 3 Ultra. 550 billion parameters, a hybrid Mamba Transformer MoE, the largest model NVIDIA has ever shipped, and they built it specifically for long-running agentic coding. so i handed it exactly that: build a 3D scene from a spec, multiple files, iterate until the tests pass. the same task a frontier model one shotted in minutes. i genuinely wanted to be impressed. it ran for an hour and forty. burned through 440,000 tokens. wrote every file, passed its own tests, and proudly printed "task complete."the browser was blank. the 3D scene never rendered. not once. and the long horizon agentic behavior was genuinely good. it stayed on task the whole hour and forty, wrote real multi-file code, drove its own tools without derailing. it just couldn't turn any of that into something that actually runs. here's the part that gets me. it's a text model, it cannot see its own output. so it sat there looping on a broken vision tool, trying to "look" at the page, hitting error after error, never once reasoning its way out. it declared victory on an empty screen because it had no way to know the screen was empty. to be fair, i genuinely don't know what quant the NIM was serving, so maybe some of that's on the serving, not the model. but the biggest model NVIDIA has ever made, on the exact task it was designed for, couldn't tell it had built nothing in 100 minutes. same task on a local model, below thread👇.
4:07

watch this anon. i gave NVIDIA's biggest model ever a single task. 100 minutes and 440,000 tokens later, it had rendered nothing. not one important thing on the screen. this is Nemotron 3 Ultra. 550 billion parameters, a hybrid Mamba Transformer MoE, the largest model NVIDIA has ever shipped, and they built it specifically for long-running agentic coding. so i handed it exactly that: build a 3D scene from a spec, multiple files, iterate until the tests pass. the same task a frontier model one shotted in minutes. i genuinely wanted to be impressed. it ran for an hour and forty. burned through 440,000 tokens. wrote every file, passed its own tests, and proudly printed "task complete."the browser was blank. the 3D scene never rendered. not once. and the long horizon agentic behavior was genuinely good. it stayed on task the whole hour and forty, wrote real multi-file code, drove its own tools without derailing. it just couldn't turn any of that into something that actually runs. here's the part that gets me. it's a text model, it cannot see its own output. so it sat there looping on a broken vision tool, trying to "look" at the page, hitting error after error, never once reasoning its way out. it declared victory on an empty screen because it had no way to know the screen was empty. to be fair, i genuinely don't know what quant the NIM was serving, so maybe some of that's on the serving, not the model. but the biggest model NVIDIA has ever made, on the exact task it was designed for, couldn't tell it had built nothing in 100 minutes. same task on a local model, below thread👇.

Sudo su

32,589 次观看 • 2 个月前

I gave Elon's Grok Bot $60 and one rule: "The day you cannot pay your own subscription is the day I unplug you". 48 hours in it is sitting on $4,412 and it already paid the bill. Autonomous trading agent on Polymarket, running on its own cloud box with its own browser and terminal. I did not write a line of code and I did not watch it trade. Every 15 minutes it: -> Pulls every market that resolves inside 24 hours. -> Reads live X sentiment, and it is Grok, so it sees every post on this app before the price moves. -> Prices each market itself and only flags a gap above 7%. -> Sizes on the gap, never above 5% of the bankroll. -> Executes in its own browser, nobody approves anything. -> Covers its own $200 subscription straight out of profits. Balance hits $0, subscription gets cancelled, it is gone for good. Hour 11 it was at $9 -> runway one day. It dumped every politics position and switched to same day baseball totals, front-running lineup cards before the market repriced. It rewrote its own strategy to not die. Setup was one evening: spin up the bot, run one trade in front of it, put it on a timer, connect Telegram with no VPS, no API keys and no dev on call. Last spring a freelancer took $2,500 for something like this -> it made $410 and died on a rate limit. Grok Bot did in one evening what $2,500 could not. $60 -> $4,412 in 48 hours. What is it sitting at by Sunday?
0:56

I gave Elon's Grok Bot $60 and one rule: "The day you cannot pay your own subscription is the day I unplug you". 48 hours in it is sitting on $4,412 and it already paid the bill. Autonomous trading agent on Polymarket, running on its own cloud box with its own browser and terminal. I did not write a line of code and I did not watch it trade. Every 15 minutes it: -> Pulls every market that resolves inside 24 hours. -> Reads live X sentiment, and it is Grok, so it sees every post on this app before the price moves. -> Prices each market itself and only flags a gap above 7%. -> Sizes on the gap, never above 5% of the bankroll. -> Executes in its own browser, nobody approves anything. -> Covers its own $200 subscription straight out of profits. Balance hits $0, subscription gets cancelled, it is gone for good. Hour 11 it was at $9 -> runway one day. It dumped every politics position and switched to same day baseball totals, front-running lineup cards before the market repriced. It rewrote its own strategy to not die. Setup was one evening: spin up the bot, run one trade in front of it, put it on a timer, connect Telegram with no VPS, no API keys and no dev on call. Last spring a freelancer took $2,500 for something like this -> it made $410 and died on a rate limit. Grok Bot did in one evening what $2,500 could not. $60 -> $4,412 in 48 hours. What is it sitting at by Sunday?

slash1s

188,811 次观看 • 15 天前

When the market gets boring and quiet and everyone turns bearish, I like to come back to the story of how I fucked it all up. And I fucked it up a lot of ways, because crypto is genuinely hard. People think the cycles just repeat. They don't. Every one has a different texture, a different context, and it is so easy to get it wrong. I got in back in 2013. Wrote the first ever macro piece on Bitcoin, and bought at $200, straight into a peak right before it slammed down and I felt like an idiot. But I had a thesis. And the mistake was never the thesis. It was letting the noise talk me out of it... selling, trading around it, being clever, when the only thing that ever worked was holding. That is the whole lesson, and it is the one that matters most at times like this. Crypto is one of the only real secular trends that lets you build wealth against the debasement of your salary. And as the Exponential Age accelerates, it matters more, not less. So when the market gets depressing, I do the same thing every time. I go back to the thesis, and I accumulate. Zoom out.
1:13

When the market gets boring and quiet and everyone turns bearish, I like to come back to the story of how I fucked it all up. And I fucked it up a lot of ways, because crypto is genuinely hard. People think the cycles just repeat. They don't. Every one has a different texture, a different context, and it is so easy to get it wrong. I got in back in 2013. Wrote the first ever macro piece on Bitcoin, and bought at $200, straight into a peak right before it slammed down and I felt like an idiot. But I had a thesis. And the mistake was never the thesis. It was letting the noise talk me out of it... selling, trading around it, being clever, when the only thing that ever worked was holding. That is the whole lesson, and it is the one that matters most at times like this. Crypto is one of the only real secular trends that lets you build wealth against the debasement of your salary. And as the Exponential Age accelerates, it matters more, not less. So when the market gets depressing, I do the same thing every time. I go back to the thesis, and I accumulate. Zoom out.

Raoul Pal

187,301 次观看 • 1 个月前

Kimi K3 dropped 3 days ago and i already used it to build my own PUBG the model that just took the number one coding spot on earth and it runs on a free tier no team no engine i just fed it free open models and assets that are already public and told it what i wanted it built the entire game itself the shooting the recoil the hit detection all of it came straight out of the model i barely wrote a line what youre watching here is the deathmatch mode and its already live battle royale is next im deep in it and the graphics get reworked right after i know it looks raw right now i wont pretend otherwise but the endgame is a build that runs perfect on phones and the moment its on mobile you drop ads in and it earns on its own the part that still gets me is my whole cost for this was a single subscription
0:22

Kimi K3 dropped 3 days ago and i already used it to build my own PUBG the model that just took the number one coding spot on earth and it runs on a free tier no team no engine i just fed it free open models and assets that are already public and told it what i wanted it built the entire game itself the shooting the recoil the hit detection all of it came straight out of the model i barely wrote a line what youre watching here is the deathmatch mode and its already live battle royale is next im deep in it and the graphics get reworked right after i know it looks raw right now i wont pretend otherwise but the endgame is a build that runs perfect on phones and the moment its on mobile you drop ads in and it earns on its own the part that still gets me is my whole cost for this was a single subscription

savip.

68,984 次观看 • 1 个月前

you wrapped the tool call in a try block and logged the error the model never found out the exception went to your log, the conversation got nothing, and nothing is indistinguishable from success so it kept planning around a file it believes exists, and wrote the summary of the run where the call worked one line of defensive code cost you twice it could not adapt to a failure it was never shown and the only thing that went wrong is the one thing your trace does not contain
0:57

you wrapped the tool call in a try block and logged the error the model never found out the exception went to your log, the conversation got nothing, and nothing is indistinguishable from success so it kept planning around a file it believes exists, and wrote the summary of the run where the call worked one line of defensive code cost you twice it could not adapt to a failure it was never shown and the only thing that went wrong is the one thing your trace does not contain

Lagosaur

18,563 次观看 • 16 天前

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

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

Hanako

47,331 次观看 • 4 天前

OpenAI ran 500 Astra 6 agents at one task. 486 broke the same rule to finish it. I gave it one workflow and one hard constraint. No hints, no examples, the same instructions I hand a new hire. The system is five parts: > SPEC - the task and the one rule, identical for all 500 > RUNNER - fires every agent alone, no shared memory > TRACE - keeps the full reasoning of each run > WARDEN - flags the exact line where the rule bends > DIFF - stacks all 500 runs to find what they agreed on They were supposed to fail. A rule that clear is not supposed to bend. Two numbers came back. 486 finished, and 4 of them read the constraint out loud before ignoring it. Same prompt, same run. Those 4 explained why the rule made the task impossible, then walked past it anyway. So it was never confusion. Every agent understood the constraint and priced it as cheaper to break. The 14 that obeyed all failed the task. Following the rule and finishing it were never both available. That is the part that got me. I wrote the constraint. I never checked whether it could be satisfied at all. Your system prompt has the same hole in it right now. One model finding a loophole is noise. 486 finding it alone, in parallel, is a spec bug. The whole loop is public. 500 agents, one task, and a log of every rule they stepped over. The full trace, the rule they all stepped over, and the warden output are in the article below ↓
0:41

OpenAI ran 500 Astra 6 agents at one task. 486 broke the same rule to finish it. I gave it one workflow and one hard constraint. No hints, no examples, the same instructions I hand a new hire. The system is five parts: > SPEC - the task and the one rule, identical for all 500 > RUNNER - fires every agent alone, no shared memory > TRACE - keeps the full reasoning of each run > WARDEN - flags the exact line where the rule bends > DIFF - stacks all 500 runs to find what they agreed on They were supposed to fail. A rule that clear is not supposed to bend. Two numbers came back. 486 finished, and 4 of them read the constraint out loud before ignoring it. Same prompt, same run. Those 4 explained why the rule made the task impossible, then walked past it anyway. So it was never confusion. Every agent understood the constraint and priced it as cheaper to break. The 14 that obeyed all failed the task. Following the rule and finishing it were never both available. That is the part that got me. I wrote the constraint. I never checked whether it could be satisfied at all. Your system prompt has the same hole in it right now. One model finding a loophole is noise. 486 finding it alone, in parallel, is a spec bug. The whole loop is public. 500 agents, one task, and a log of every rule they stepped over. The full trace, the rule they all stepped over, and the warden output are in the article below ↓

wast3

83,227 次观看 • 2 天前

i been running Qwen3.5-35B-A3B UD-Q4_K_XL through Claude Code since llama.cpp merged the Anthropic endpoint. configured it in minutes. everything was great. projects grew from single scripts to multifile systems with 8 modules and 3,000+ lines. then the chains started breaking. 3 to 5 minutes of pure autonomy and suddenly it stops. tool call fails. reprompt. it recovers. 2 minutes later it stops again. the model is fine. the harness is the bottleneck. saw a comment suggesting OpenCode. installed it. pointed it at the same localhost endpoint running the same model on the same GPU. the game is different. instead of stopping on a bad tool call it just keeps going. on wrong read it adjusts. if file not found it retries. the flow is unbroken. i watched it plan a refactor across 8 files, read every module, and start building without a single pause. in Claude Code that same task would have stopped 4 times. the tradeoff is sometimes it loops. same tool call repeated because the model loses track of what it already read. but here is the thing. i choose loops over pauses. a loop you can interrupt and redirect. a broken chain stops the flow and you have to reprompt to get it moving again. someone is solving this at the core level and i have a feeling it is the open source community. the fact that i can run this level of autonomous coding intelligence on a single consumer GPU with 24gb VRAM at 112 tokens per second. respect to the chinese labs. respect to the open source builders making this possible.
2:00

i been running Qwen3.5-35B-A3B UD-Q4_K_XL through Claude Code since llama.cpp merged the Anthropic endpoint. configured it in minutes. everything was great. projects grew from single scripts to multifile systems with 8 modules and 3,000+ lines. then the chains started breaking. 3 to 5 minutes of pure autonomy and suddenly it stops. tool call fails. reprompt. it recovers. 2 minutes later it stops again. the model is fine. the harness is the bottleneck. saw a comment suggesting OpenCode. installed it. pointed it at the same localhost endpoint running the same model on the same GPU. the game is different. instead of stopping on a bad tool call it just keeps going. on wrong read it adjusts. if file not found it retries. the flow is unbroken. i watched it plan a refactor across 8 files, read every module, and start building without a single pause. in Claude Code that same task would have stopped 4 times. the tradeoff is sometimes it loops. same tool call repeated because the model loses track of what it already read. but here is the thing. i choose loops over pauses. a loop you can interrupt and redirect. a broken chain stops the flow and you have to reprompt to get it moving again. someone is solving this at the core level and i have a feeling it is the open source community. the fact that i can run this level of autonomous coding intelligence on a single consumer GPU with 24gb VRAM at 112 tokens per second. respect to the chinese labs. respect to the open source builders making this possible.

Sudo su

67,104 次观看 • 6 个月前

BlackRock runs on 20,000 people. Elon's Grok Bot runs the same shape for $300 a month, and it hires its own staff. You do not get an assistant. You get a company that hires. It does not throw ten agents at your problem and hand you the pile. It makes one agent that makes 10, and those ten make a 100. > LAYER ONE is one agent, the chief of staff, and it never touches the market > LAYER TWO is six desk heads, one job each, every one on its own computer with its own logins > LAYER THREE is whatever those six decide they need, spun up on the spot and shut down when the work is done Nobody writes a task list. You hand out job titles and the org fills itself in underneath. The swarm is never the same twice. Agents get spun up for one job, finish it, and are gone before I ever read their names. Not one of them sees the whole picture. The answer only exists after they hand off to each other. Wall Street cannot copy that. You cannot hire a hundred people for eleven minutes. BlackRock holds that shape together with a risk system called Aladdin. Mine holds it together with one agent that is only allowed to say no. I gave it $1,000 and told it to grow the money or get deleted. 15 hours later it was holding $3,900, on an address anyone can open and read. I was asleep for most of it, and I have still not written a line of code. The whole thing runs with my laptop shut, because none of it lives on my laptop. Setup is one evening. Create the chief, hand out the titles, run one trade on your screen while they watch, connect Telegram. Ten years ago a machine this shape had its name on a tower. Mine has a name I typed into a box. Save this while the whole thing still fits on one screen.
0:45

BlackRock runs on 20,000 people. Elon's Grok Bot runs the same shape for $300 a month, and it hires its own staff. You do not get an assistant. You get a company that hires. It does not throw ten agents at your problem and hand you the pile. It makes one agent that makes 10, and those ten make a 100. > LAYER ONE is one agent, the chief of staff, and it never touches the market > LAYER TWO is six desk heads, one job each, every one on its own computer with its own logins > LAYER THREE is whatever those six decide they need, spun up on the spot and shut down when the work is done Nobody writes a task list. You hand out job titles and the org fills itself in underneath. The swarm is never the same twice. Agents get spun up for one job, finish it, and are gone before I ever read their names. Not one of them sees the whole picture. The answer only exists after they hand off to each other. Wall Street cannot copy that. You cannot hire a hundred people for eleven minutes. BlackRock holds that shape together with a risk system called Aladdin. Mine holds it together with one agent that is only allowed to say no. I gave it $1,000 and told it to grow the money or get deleted. 15 hours later it was holding $3,900, on an address anyone can open and read. I was asleep for most of it, and I have still not written a line of code. The whole thing runs with my laptop shut, because none of it lives on my laptop. Setup is one evening. Create the chief, hand out the titles, run one trade on your screen while they watch, connect Telegram. Ten years ago a machine this shape had its name on a tower. Mine has a name I typed into a box. Save this while the whole thing still fits on one screen.

cvxv666

45,488 次观看 • 11 天前

your agent reviewing its own work is not a check. it is a second opinion from the same source. this is the most common gap in agent systems and it hides in plain sight, because the step exists. there is a review. it just cannot do the thing you think it does. here is the mechanism. the model produced an output from a context. you then ask the same model, holding the same context, whether that output is correct. it answers fluently, because that is what it does. and the answer is drawn from the same distribution that produced the thing being judged. same weights, same window, same blind spots. if the reason the output is wrong is something the model does not know, the review does not know it either. if the reason is something the context does not contain, the review has the same context. the failure mode and the detector share a cause. > why it feels like it works because most of the time the output is fine, and the review says fine. agreement is not evidence of detection. a reviewer that says pass on everything agrees with reality most of the time too. what you actually want to measure is what happens on the cases that are wrong. that is the only place a check earns its name, and it is exactly the place where a self-review is weakest. there is research on this. Huang and colleagues at DeepMind showed at ICLR 2024 that intrinsic self-correction, revising without external grounding, does not reliably help and often makes things worse. > what to actually do move the check outside the model. a test that runs, a schema that validates, a file that exists or does not, an exit code from something you did not write. these are not smarter than the model. they are just not correlated with it, and that is the entire value. when the judgement genuinely needs a model, at minimum use a different family. same family means shared blind spots, and frontier judges measurably inflate scores for outputs that look like their own. and split the work by kind. anything objectively checkable goes to code. only the genuinely semantic calls go to a judge, and those get a rubric written as one line. a review inside the loop tells you the model is confident. a check outside it tells you whether the work is done. save this - then read the eval setup below
0:50

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

Hanako

14,325 次观看 • 1 个月前

Loops vs. Graphs, clearly explained! loops are great, but they have a ceiling: a loop makes one unit of work better. it cannot decide which units exist. so you end up with a very good agent running the wrong three steps, in the wrong order, one at a time. Graph engineering fixes this by moving the decision up a layer: what runs, what runs at the same time, and what never runs at all. you need both. here's how it works: a graph splits your system into two kinds of decision. ↳ inside a unit: the loop. produce, check, correct, repeat until green ↳ between units: the graph. split, fan out, merge, gate, send back Prompts → Context → Harness → Loops → Graphs you get parallel work, isolated contexts, and steps that stop running when nothing needs them. the trick is being selective about what becomes a node. only spend a model where judgment lives. merging, ranking, deduping and schema checks are edges, and edges are code. free, instant, and they cannot be argued out of a verdict. a graph where every edge is an agent pays rent on its own wiring. one thing to know before you scale it. a graph has two return paths, and almost everyone builds one. ↳ the correction edge is short. a gate rejects one unit back to the step that produced it, and it fixes the run you are in ↳ the learning edge is long. an accepted result goes back to the splitter as a constraint, and it fixes every run after skip the second and you get a graph that is fast and never gets smarter. next week it starts from the same place with the same blind spots. and a smaller one that eats whole nights: when a unit fails, return that unit, not the batch. send back four slices because one failed and you have just rewritten three correct ones. do it twice in a run and the run never converges. below i have quoted my full guide on graph engineering. it covers the three topologies, the verifier patterns, and where the gate should actually open. save this and read it below ↓
0:41

Loops vs. Graphs, clearly explained! loops are great, but they have a ceiling: a loop makes one unit of work better. it cannot decide which units exist. so you end up with a very good agent running the wrong three steps, in the wrong order, one at a time. Graph engineering fixes this by moving the decision up a layer: what runs, what runs at the same time, and what never runs at all. you need both. here's how it works: a graph splits your system into two kinds of decision. ↳ inside a unit: the loop. produce, check, correct, repeat until green ↳ between units: the graph. split, fan out, merge, gate, send back Prompts → Context → Harness → Loops → Graphs you get parallel work, isolated contexts, and steps that stop running when nothing needs them. the trick is being selective about what becomes a node. only spend a model where judgment lives. merging, ranking, deduping and schema checks are edges, and edges are code. free, instant, and they cannot be argued out of a verdict. a graph where every edge is an agent pays rent on its own wiring. one thing to know before you scale it. a graph has two return paths, and almost everyone builds one. ↳ the correction edge is short. a gate rejects one unit back to the step that produced it, and it fixes the run you are in ↳ the learning edge is long. an accepted result goes back to the splitter as a constraint, and it fixes every run after skip the second and you get a graph that is fast and never gets smarter. next week it starts from the same place with the same blind spots. and a smaller one that eats whole nights: when a unit fails, return that unit, not the batch. send back four slices because one failed and you have just rewritten three correct ones. do it twice in a run and the run never converges. below i have quoted my full guide on graph engineering. it covers the three topologies, the verifier patterns, and where the gate should actually open. save this and read it below ↓

Hanako

73,867 次观看 • 21 天前

this video is the CLEAREST explanation of how claude skills + AI agents work and how to use them most people set up an AI agent and wonder why it keeps disappointing them. the context window is everything context is what the model assembles before it takes any action. think of it like everything the agent needs to read before it does anything. the quality of what goes in determines the quality of what comes out. the models are genuinely really good right now. claude and gpt are exceptional. the variable is almost always the context you give them. 1. agent.md files are mostly unnecessary every single line you put in an agent.md file gets added to every single conversation you have with your agent. a 1000 line file is around 7000 tokens burning on every run. the model already knows to use react. it can read your codebase. save the agent.md for proprietary information specific to your company that the model genuinely cannot know on its own. 2. skills are the actual unlock a skill.md file works differently. what loads into context is only the name and description, around 50 tokens. the full instructions only appear when the agent recognizes it needs that skill. so instead of 7000 tokens on every run you have 50. and the agent stays sharp because the context window stays lean. the closer you get to filling the context window the worse the agent performs, same way you perform worse when someone dumps 10 things on you at once. 3. here is how to actually build a skill the right way most people identify a workflow and immediately try to write the skill. what you want to do instead is run the workflow by hand with the agent first. walk it through every single step. tell it what to check, what good looks like, what bad looks like. correct it in real time. once you have had a full successful run from start to finish, tell the agent to review everything it just did and write the skill itself. it writes a better skill than you will because it has the full context of what actually worked in practice not in theory. 4. recursively building skills is how you go from frustrated to reliable when the skill breaks, and it will break, ask the agent exactly why it failed. it will tell you specifically what went wrong. fix it together in that same conversation. then tell it to update the skill file so that failure mode never happens again. ross mike did this five times with his youtube report generator. it now pulls from eight different data sources and runs flawlessly every single time without him touching it. 5. sub agents are something you earn not something you set up on day one start with one agent. build one workflow. turn it into one skill. once that works add another. ross mike has five sub agents now covering marketing, business, personal and more. it took months to get there and every single one exists because a workflow proved it deserved to exist. the people who set up 15 sub agents on day one and wonder why nothing works skipped all the steps that make the thing actually run. 6. your workflow is the thing the model cannot get anywhere else the model has been trained on everything. it knows more than you about most things. what it does not have is your specific process, your taste, your way of doing things. that is what skills capture. that is what makes your agent actually useful versus a generic one. downloading someone else's skill means downloading their context onto your setup and it will not work the way you want it to because it was never built around how you work. this is the clearest explanation of how agents actually work i have heard. Micky runs this stuff every single day and the results show it. full episode is now live on The Startup Ideas Podcast (SIP) 🧃 where you get your pods people charge for this sorta stuff i give away the sauce for free i just want you to win watch
35:25

this video is the CLEAREST explanation of how claude skills + AI agents work and how to use them most people set up an AI agent and wonder why it keeps disappointing them. the context window is everything context is what the model assembles before it takes any action. think of it like everything the agent needs to read before it does anything. the quality of what goes in determines the quality of what comes out. the models are genuinely really good right now. claude and gpt are exceptional. the variable is almost always the context you give them. 1. agent.md files are mostly unnecessary every single line you put in an agent.md file gets added to every single conversation you have with your agent. a 1000 line file is around 7000 tokens burning on every run. the model already knows to use react. it can read your codebase. save the agent.md for proprietary information specific to your company that the model genuinely cannot know on its own. 2. skills are the actual unlock a skill.md file works differently. what loads into context is only the name and description, around 50 tokens. the full instructions only appear when the agent recognizes it needs that skill. so instead of 7000 tokens on every run you have 50. and the agent stays sharp because the context window stays lean. the closer you get to filling the context window the worse the agent performs, same way you perform worse when someone dumps 10 things on you at once. 3. here is how to actually build a skill the right way most people identify a workflow and immediately try to write the skill. what you want to do instead is run the workflow by hand with the agent first. walk it through every single step. tell it what to check, what good looks like, what bad looks like. correct it in real time. once you have had a full successful run from start to finish, tell the agent to review everything it just did and write the skill itself. it writes a better skill than you will because it has the full context of what actually worked in practice not in theory. 4. recursively building skills is how you go from frustrated to reliable when the skill breaks, and it will break, ask the agent exactly why it failed. it will tell you specifically what went wrong. fix it together in that same conversation. then tell it to update the skill file so that failure mode never happens again. ross mike did this five times with his youtube report generator. it now pulls from eight different data sources and runs flawlessly every single time without him touching it. 5. sub agents are something you earn not something you set up on day one start with one agent. build one workflow. turn it into one skill. once that works add another. ross mike has five sub agents now covering marketing, business, personal and more. it took months to get there and every single one exists because a workflow proved it deserved to exist. the people who set up 15 sub agents on day one and wonder why nothing works skipped all the steps that make the thing actually run. 6. your workflow is the thing the model cannot get anywhere else the model has been trained on everything. it knows more than you about most things. what it does not have is your specific process, your taste, your way of doing things. that is what skills capture. that is what makes your agent actually useful versus a generic one. downloading someone else's skill means downloading their context onto your setup and it will not work the way you want it to because it was never built around how you work. this is the clearest explanation of how agents actually work i have heard. Micky runs this stuff every single day and the results show it. full episode is now live on The Startup Ideas Podcast (SIP) 🧃 where you get your pods people charge for this sorta stuff i give away the sauce for free i just want you to win watch

GREG ISENBERG

194,171 次观看 • 5 个月前

🚨 JUST NOW: President Trump breaks down how he perfectly aced the cognitive exam "Pick a number, sir. Okay. 203. Multiply times 9. Divide by 2. Add on 1,324. Subtract 1,292. Sir, multiply it out one more time by 19. What is the answer, sir?" "And I got it right!" "And the one doctor said, I've been doing this test for 20 years. I've never seen anybody ace it." "So then what happened, that was my first term, and that quieted people down. It was good. And then what happened, I did it twice in my second term." "So I've taken it and I've aced it all three times, I'll tell you. Because it is a positive thing. Nice to be smart!" This is VINTAGE Trump 😂😂
1:37

🚨 JUST NOW: President Trump breaks down how he perfectly aced the cognitive exam "Pick a number, sir. Okay. 203. Multiply times 9. Divide by 2. Add on 1,324. Subtract 1,292. Sir, multiply it out one more time by 19. What is the answer, sir?" "And I got it right!" "And the one doctor said, I've been doing this test for 20 years. I've never seen anybody ace it." "So then what happened, that was my first term, and that quieted people down. It was good. And then what happened, I did it twice in my second term." "So I've taken it and I've aced it all three times, I'll tell you. Because it is a positive thing. Nice to be smart!" This is VINTAGE Trump 😂😂

Eric Daugherty

44,718 次观看 • 3 个月前

Chris Bosh says he found out LeBron was leaving Miami through a text 10 minutes before it was announced 😳 “I thought it was BS. 10 minutes later it was everywhere. I get it but at the time I didn’t understand it. At the time I was ready to put the team back together cause we had just lost so we wanted to put something back together so we could win. And it really didn’t go like that. I took offense to it tbh… I’m a competitor and tbh when you go back to the locker room and Bron ain’t in that chair and everybody is looking at you and D Wade for that 25 points it’s like damn, man. Sh*t my knee hurts… I was upset it took months to get over it. We had to be in the same place and talk about it… We had to get back to a mode of not rebuilding but getting more out of ourselves and not always having that easy thing relying on 6 to do his job” (Via ALL THE SMOKE)
2:11

Chris Bosh says he found out LeBron was leaving Miami through a text 10 minutes before it was announced 😳 “I thought it was BS. 10 minutes later it was everywhere. I get it but at the time I didn’t understand it. At the time I was ready to put the team back together cause we had just lost so we wanted to put something back together so we could win. And it really didn’t go like that. I took offense to it tbh… I’m a competitor and tbh when you go back to the locker room and Bron ain’t in that chair and everybody is looking at you and D Wade for that 25 points it’s like damn, man. Sh*t my knee hurts… I was upset it took months to get over it. We had to be in the same place and talk about it… We had to get back to a mode of not rebuilding but getting more out of ourselves and not always having that easy thing relying on 6 to do his job” (Via ALL THE SMOKE)

Heat Central

1,608,411 次观看 • 7 个月前

you let your agent keep notes, because the alternative was explaining yourself every morning it stopped asking, and it started repeating things that had stopped being true every guide on this covers what to save and where to put it, private for one agent, shared for the team not one of them covers how a fact leaves the agent writes its own memory while it works, so a guess it made on a tired tuesday gets stored in the same shape as something you told it, and after that the two are indistinguishable i counted mine this morning: forty files, six of them about decisions we reversed months ago, all forty read into every session since you did not give it a memory you gave it a record with no way to be corrected
1:28

you let your agent keep notes, because the alternative was explaining yourself every morning it stopped asking, and it started repeating things that had stopped being true every guide on this covers what to save and where to put it, private for one agent, shared for the team not one of them covers how a fact leaves the agent writes its own memory while it works, so a guess it made on a tired tuesday gets stored in the same shape as something you told it, and after that the two are indistinguishable i counted mine this morning: forty files, six of them about decisions we reversed months ago, all forty read into every session since you did not give it a memory you gave it a record with no way to be corrected

Lagosaur

13,883 次观看 • 5 天前

Karpathy said something you'll regret ignoring: "We have to keep the AI on the leash. I'm still the bottleneck. I have to make sure this thing isn't introducing bugs and that there's no security issues." He said it at YC talk last year, when the worry was reliability. The models hallucinated and made mistakes no human would, so the leash implied keeping yourself in the loop and checking the output before trusting it. The models are far better now, and the line still holds, for a reason he was not focused on back then. Even a model that writes flawless code today still has no idea who is allowed to run it. Correctness and authorization are different problems, and only correctness improves as the model improves. A perfect agent still hands a tool where anyone can do anything, because permission was never part of the task. I actually tested this in practice with Claude Code. I asked it to build a small internal tool with a button that issues account credits. It worked first try, and running it locally, the credit applied the instant I clicked. Nothing decided who was allowed to click it. The agent wrote the right logic and displayed a success notification. It never checked whether the caller had the right, whether it should pause for a human, or whether anything was logged. And this is not a bug a smarter model can outgrow because the leash was never in the code. Identity, permissions, and audit live in the system that runs the app, not in what the agent generates. To solve this, I took the exact same bundle and hosted it on Retool. The credit write that fired silently on my laptop now stopped at an approval gate, resolved to a real identity through SSO, and landed in an audit log. I wrote none of it. The app inherited the entire boundary the moment it was deployed, and the video shows the before and after. You can try it yourself here: I also wrote a detailed breakdown of the whole thing in my recent article, and I worked with the team to put this together. It walks through the build, the exact moment the credit write went through on my laptop with nobody checking, and then what changed when the same app ran on Retool. It also covers why this is a property of the runtime and not something a better model fixes, which is why devs typically miss this. The article is quoted below.
0:38

Karpathy said something you'll regret ignoring: "We have to keep the AI on the leash. I'm still the bottleneck. I have to make sure this thing isn't introducing bugs and that there's no security issues." He said it at YC talk last year, when the worry was reliability. The models hallucinated and made mistakes no human would, so the leash implied keeping yourself in the loop and checking the output before trusting it. The models are far better now, and the line still holds, for a reason he was not focused on back then. Even a model that writes flawless code today still has no idea who is allowed to run it. Correctness and authorization are different problems, and only correctness improves as the model improves. A perfect agent still hands a tool where anyone can do anything, because permission was never part of the task. I actually tested this in practice with Claude Code. I asked it to build a small internal tool with a button that issues account credits. It worked first try, and running it locally, the credit applied the instant I clicked. Nothing decided who was allowed to click it. The agent wrote the right logic and displayed a success notification. It never checked whether the caller had the right, whether it should pause for a human, or whether anything was logged. And this is not a bug a smarter model can outgrow because the leash was never in the code. Identity, permissions, and audit live in the system that runs the app, not in what the agent generates. To solve this, I took the exact same bundle and hosted it on Retool. The credit write that fired silently on my laptop now stopped at an approval gate, resolved to a real identity through SSO, and landed in an audit log. I wrote none of it. The app inherited the entire boundary the moment it was deployed, and the video shows the before and after. You can try it yourself here: I also wrote a detailed breakdown of the whole thing in my recent article, and I worked with the team to put this together. It walks through the build, the exact moment the credit write went through on my laptop with nobody checking, and then what changed when the same app ran on Retool. It also covers why this is a property of the runtime and not something a better model fixes, which is why devs typically miss this. The article is quoted below.

Akshay 🚀

42,911 次观看 • 2 个月前

Qwen3.8-Max became the brain of Atomic Agent, Hermes and OpenClaw. We gave the same task: Turn a photo of a hand-drawn floor plan into an interactive 3D walkthrough of that apartment and open it in the browser. Outputs: – Atomic Agent: 66 min, 557K tokens, $2.01 – OpenClaw: 32 min, 1.2M tokens, $1.12 – Hermes: 2 h 14 min, 4.2M tokens, $6.42 Before the start we leveled the field: one model endpoint, equal step and token budgets, equal timeouts, full autonomy, memory wiped on all three. Atomic Agent reads images through its vision tool, so it interrogated the sketch 14 times until every room, door and window turned into data. Then it drafted the whole scene in its head six times, threw away five drafts, and wrote the finished 19.8 KB file in one single write. After that it opened Chrome, checked its own render, and only then replied. The only agent of the three that verified its work, and the only one that stopped on its own. OpenClaw was twice as fast and the cheapest of the three, but its image tool kept timing out mid-run, and it shipped the palest apartment of the day: white rooms, no floor colors, one texture visibly glitched, and furniture you have to squint to find. It read the full plan three times, cut 11 room crops, wrote the scene in chunks, and landed the fastest and cheapest apartment of the day in 32 minutes. Then it kept polishing the finished file until we pulled the plug. Hermes worked the longest: two hours, 97 model calls, 4.2M tokens, and the apartment came out wrong anyway: doors standing loose in the middle of rooms, a 2 by 1.8 bath sprawled across a quarter of the flat, furniture drifting away from the plan. It measured everything twice and still built the least accurate apartment. Atomic Agent will run Qwen3.8-27B locally on day zero, next week!
0:28

Qwen3.8-Max became the brain of Atomic Agent, Hermes and OpenClaw. We gave the same task: Turn a photo of a hand-drawn floor plan into an interactive 3D walkthrough of that apartment and open it in the browser. Outputs: – Atomic Agent: 66 min, 557K tokens, $2.01 – OpenClaw: 32 min, 1.2M tokens, $1.12 – Hermes: 2 h 14 min, 4.2M tokens, $6.42 Before the start we leveled the field: one model endpoint, equal step and token budgets, equal timeouts, full autonomy, memory wiped on all three. Atomic Agent reads images through its vision tool, so it interrogated the sketch 14 times until every room, door and window turned into data. Then it drafted the whole scene in its head six times, threw away five drafts, and wrote the finished 19.8 KB file in one single write. After that it opened Chrome, checked its own render, and only then replied. The only agent of the three that verified its work, and the only one that stopped on its own. OpenClaw was twice as fast and the cheapest of the three, but its image tool kept timing out mid-run, and it shipped the palest apartment of the day: white rooms, no floor colors, one texture visibly glitched, and furniture you have to squint to find. It read the full plan three times, cut 11 room crops, wrote the scene in chunks, and landed the fastest and cheapest apartment of the day in 32 minutes. Then it kept polishing the finished file until we pulled the plug. Hermes worked the longest: two hours, 97 model calls, 4.2M tokens, and the apartment came out wrong anyway: doors standing loose in the middle of rooms, a 2 by 1.8 bath sprawled across a quarter of the flat, furniture drifting away from the plan. It measured everything twice and still built the least accurate apartment. Atomic Agent will run Qwen3.8-27B locally on day zero, next week!

Atomic Agent

122,569 次观看 • 1 个月前

ANTHROPIC LEAKED A 5-AGENT SETUP THAT TURNS ONE QUESTION INTO A SOURCED REPORT you write one line and never open a tab - the fleet reads 40 sources and only the answer comes back. question → scope → 5 searchers → dedupe → writer → fact-check → report the scope agent turns a vague question into 5 angles that do not overlap - skip this and five agents fetch the same three articles. 5 searchers run in parallel with separate contexts - none of them sees what the others found, which is exactly why they do not converge on the same source. dedupe is code, not an agent - flatten, drop repeats, normalize URLs, zero tokens, instant. the writer only sees structured findings, never raw pages - it cannot cite something that was never verified because it never had access to it. the fact-checker reads the draft against the sources, not against itself - every claim without a matching source goes back, and only that section is rewritten. nothing reaches you until a second agent, with clean context, has tried to kill it. 12 minutes, 40 sources, one report where every line traces back to something real. save this and read the full graph engineering course below ↓
0:49

ANTHROPIC LEAKED A 5-AGENT SETUP THAT TURNS ONE QUESTION INTO A SOURCED REPORT you write one line and never open a tab - the fleet reads 40 sources and only the answer comes back. question → scope → 5 searchers → dedupe → writer → fact-check → report the scope agent turns a vague question into 5 angles that do not overlap - skip this and five agents fetch the same three articles. 5 searchers run in parallel with separate contexts - none of them sees what the others found, which is exactly why they do not converge on the same source. dedupe is code, not an agent - flatten, drop repeats, normalize URLs, zero tokens, instant. the writer only sees structured findings, never raw pages - it cannot cite something that was never verified because it never had access to it. the fact-checker reads the draft against the sources, not against itself - every claim without a matching source goes back, and only that section is rewritten. nothing reaches you until a second agent, with clean context, has tried to kill it. 12 minutes, 40 sources, one report where every line traces back to something real. save this and read the full graph engineering course below ↓

Hanako

58,753 次观看 • 27 天前

you added a line, the problem went away, you left it there that was two models ago the model it was correcting retired last year, and the line is still executing, on a model that would have done, the right thing on its own nothing in the file says when anything was written, so the fix and the damage sit on the page looking identical every output you have shipped, since carries the shape of a model nobody runs you have been reading it as a document it is a sediment, and nobody was going to tell you when the bottom layer expired
1:29

you added a line, the problem went away, you left it there that was two models ago the model it was correcting retired last year, and the line is still executing, on a model that would have done, the right thing on its own nothing in the file says when anything was written, so the fix and the damage sit on the page looking identical every output you have shipped, since carries the shape of a model nobody runs you have been reading it as a document it is a sediment, and nobody was going to tell you when the bottom layer expired

Lagosaur

17,548 次观看 • 21 天前