正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

So I gave Step 5 Preview a bug it had never seen before, and just watched. Filtered 193 users out of 208, hit export, and the file came back with 6 rows. Whatever happened to be sitting on the current page. That's the bug I handed it. No file... names. No hints. Just the codebase and the report. Here's what happened: 1. It read through the project on its own and found the real cause. The export was pulling from the current page slice instead of the full filtered list. 2. It fixed the root cause, then renamed the ambiguous fields so the mistake can't happen twice. 3. It added retry logic for failed downloads. Bounded attempts, a visible error state if every one fails. I didn't even ask for that part. 4. It wrote 15 new tests across three layers. Unit, component, and real browser runs. 5. Then it opened the running app and reran the exact scenario itself. The numbers: 208 total users 193 matched the filter 6 rows in the broken export 193 rows in the fixed one 15 new tests, all passing One prompt. No follow-up corrections. It verified its own fix before calling the task done. Curious to see how it handles your own coding challenges? Give Step 5 Preview a try 👉 It’s now #1 on OpenRouter’s Trending ranking, so this might be a good time to put it to the test. Check the ranking:show more

Parul Gautam

85,604 subscribers

49,002 次观看 • 5 天前 •via X (Twitter)

新闻政治教育科学技术

Anya Rossi• Live Now

Private livecam show

25 条评论

Theinfralab 的头像

Theinfralab5 天前

Fixing the bug is one thing, making sure it stays fixed is another.

Parul Gautam 的头像

Parul Gautam5 天前

Exactly! A fix is only reliable when it holds up over time.

Nelly; 的头像

Nelly;5 天前

fixing the root cause instead of patching symptoms is everything

D-Coder 的头像

D-Coder5 天前

Love seeing @StepFun_ai focus on accuracy, not just generating answers.

codemarch 的头像

codemarch5 天前

Going beyond the initial request to handle failed downloads is a nice touch.

Rishabh 的头像

Rishabh5 天前

I’d trust the result more knowing it reran the exact scenario itself.

Gina Acosta 的头像

Gina Acosta5 天前

the ability to understand the codebase and trace a bug is what stands out here

Manish Kumar Shah 的头像

Manish Kumar Shah5 天前

This is what makes AI coding tools genuinely useful in real projects.

Atul Kumar 的头像

Atul Kumar5 天前

AI coding gets much more interesting when verification is part of the process.

EyeingAI 的头像

EyeingAI5 天前

The full fix-and-test workflow looks really solid.

Ajay Sharma 的头像

Ajay Sharma5 天前

No hints, just the codebase and the bug report. That’s a real test.

RAVI KUMAR SAHU 的头像

RAVI KUMAR SAHU5 天前

193 filtered users and 193 exported rows. That’s the kind of result you want.

Jawad Rahman 的头像

Jawad Rahman5 天前

This is another step toward making AI more accessible.

Roshni 的头像

Roshni5 天前

This bug hunting test is wild

Pradeep Pandey 的头像

Pradeep Pandey5 天前

This looks like a game changer

Carlo Makes 的头像

Carlo Makes5 天前

6 rows out of 15 is the kind of bug i'd only notice after the client opened the file

JORDAN 的头像

JORDAN5 天前

Thanks for the share Parul

ᗩshhh 的头像

ᗩshhh5 天前

193 rows after the fix is the proof that counts.

Hasan Rajpoot 的头像

Hasan Rajpoot5 天前

The 15 tests are a really important part of this story.

Mujeeb Ahmed 的头像

Mujeeb Ahmed5 天前

I like that it addressed the naming issue too, not just the immediate bug.

Andrew Bolis 的头像

Andrew Bolis5 天前

Adding bounded retries unprompted is a thoughtful extra.

Utkarsh Sharma 的头像

Utkarsh Sharma5 天前

Love that it added retry logic without even being asked.

Dhaval Makwana 的头像

Dhaval Makwana5 天前

The fact that it tested its own fix makes this even more impressive.

Luqman Ali 的头像

Luqman Ali5 天前

That's Awesome

IamAlam 的头像

IamAlam5 天前

One prompt from debugging to verification is a pretty solid workflow.

相关视频

This new AI model is crazy fast at coding 🤯 It fixed a broken backend in one prompt and around 83 seconds. I tested Space Bunny inside OpenCode with a FastAPI repo. It had 3 failing tests caused by broken order filtering, unsafe money calculations and incorrect product ranking. I didn’t tell it where the bugs were. It read the product contract, inspected the code, found the root causes and fixed everything. Then it added 2 edge-case tests on its own. The final result: 3 failing tests → 6 passing tests Only 2 files changed Public API untouched It made one small CLI mistake during the run, but the final patch was clean. Tbh this is the kind of speed I expect from a Flash-level model. Not just fast output. Fast useful work.

This new AI model is crazy fast at coding 🤯 It fixed a broken backend in one prompt and around 83 seconds. I tested Space Bunny inside OpenCode with a FastAPI repo. It had 3 failing tests caused by broken order filtering, unsafe money calculations and incorrect product ranking. I didn’t tell it where the bugs were. It read the product contract, inspected the code, found the root causes and fixed everything. Then it added 2 edge-case tests on its own. The final result: 3 failing tests → 6 passing tests Only 2 files changed Public API untouched It made one small CLI mistake during the run, but the final patch was clean. Tbh this is the kind of speed I expect from a Flash-level model. Not just fast output. Fast useful work.

Farhan

60,769 次观看 • 19 天前

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👇.

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 • 3 个月前

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.

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

39,764 次观看 • 4 个月前

Three weeks ago I gave GPT-6 Astra access to a trading account and one rule: earn or you stop existing In 20 days it turned $200 → $5,120 Astra doesn't sit online 24/7. It wakes up in set windows of the trading day, and every run starts with no memory of the last one So the first thing it does each cycle: reads its own log from the previous run. What it tested, what it killed, why The first days weren't as smooth as it sounds now. On day two it bled $80 on a trade that passed the backtest clean. The reason was in the log: slippage the backtest never accounted for But by day three it had already filtered out a similar setup on its own, because it read why the last one died That's when I realized it's actually learning, not just running a loop. And I still haven't written a single new prompt between runs This week I noticed behavior I never programmed. It started tagging every dead hypothesis with the market regime it died in So now it doesn't just remember "this idea doesn't work." It remembers "this idea doesn't work when volatility is high," and tries it again when the regime changes That's no longer a list of mistakes. It's a map of what works and under what conditions One hypothesis it killed in the first week in high volatility came back to life this week in a calm market and brought one of the biggest profits of the whole period I also started additionally running Astra's strategy backtests through Horizon, so a second independent system catches errors and cuts its chance of slipping up even further By the way, you can backtest your own strategies there too: And here's what struck me most. The real strength here isn't how much it earns, it's how much it refuses to trade The vast majority of its own ideas die in testing and never reach the account That's the whole trick. Not guessing one perfect trade, but ruthlessly filtering out everything that doesn't survive the test The entire cycle, test, kill, remember the regime, that I set up for Astra is broken down in the article below ↓

Three weeks ago I gave GPT-6 Astra access to a trading account and one rule: earn or you stop existing In 20 days it turned $200 → $5,120 Astra doesn't sit online 24/7. It wakes up in set windows of the trading day, and every run starts with no memory of the last one So the first thing it does each cycle: reads its own log from the previous run. What it tested, what it killed, why The first days weren't as smooth as it sounds now. On day two it bled $80 on a trade that passed the backtest clean. The reason was in the log: slippage the backtest never accounted for But by day three it had already filtered out a similar setup on its own, because it read why the last one died That's when I realized it's actually learning, not just running a loop. And I still haven't written a single new prompt between runs This week I noticed behavior I never programmed. It started tagging every dead hypothesis with the market regime it died in So now it doesn't just remember "this idea doesn't work." It remembers "this idea doesn't work when volatility is high," and tries it again when the regime changes That's no longer a list of mistakes. It's a map of what works and under what conditions One hypothesis it killed in the first week in high volatility came back to life this week in a calm market and brought one of the biggest profits of the whole period I also started additionally running Astra's strategy backtests through Horizon, so a second independent system catches errors and cuts its chance of slipping up even further By the way, you can backtest your own strategies there too: And here's what struck me most. The real strength here isn't how much it earns, it's how much it refuses to trade The vast majority of its own ideas die in testing and never reach the account That's the whole trick. Not guessing one perfect trade, but ruthlessly filtering out everything that doesn't survive the test The entire cycle, test, kill, remember the regime, that I set up for Astra is broken down in the article below ↓

qwinsi

74,829 次观看 • 25 天前

I put Space Bunny, the most-used model on OpenCode this week to the test on a broken launch dashboard. With one prompt and no hints, it: → Found two separate logic bugs → Fixed a broken 700% progress calculation → Repaired the task filters → Improved the result from 2/6 to 6/6 passing tests Fast, methodical, and reliable with tools, not just patching the UI, but verifying the fix. Space Bunny is currently free to try on OpenCode

I put Space Bunny, the most-used model on OpenCode this week to the test on a broken launch dashboard. With one prompt and no hints, it: → Found two separate logic bugs → Fixed a broken 700% progress calculation → Repaired the task filters → Improved the result from 2/6 to 6/6 passing tests Fast, methodical, and reliable with tools, not just patching the UI, but verifying the fix. Space Bunny is currently free to try on OpenCode

Darshal Jaitwar

24,922 次观看 • 15 天前

OpenAI. said. this. publicly. their own engineers just proved one idea on themselves, in writing: stop telling AI what's wrong. hand it the whole broken thing and let it find out they gave GPT-6 Astra a slow test build of their own coding tool. one cause found, a memory bottleneck, one allocator swapped, every turn 25× faster this is GPT-6 Astra, the layer that fixes the cause instead of the symptom, $0 on top of the ChatGPT plan you already pay for: - open ChatGPT or Codex, pick GPT-6 Astra, hand it the whole thing: the folder, the file that takes a minute to open. it works in apps with no API and reads your screen - type one sentence: find the one cause, prove it, fix it, do not patch around it - leave the room. it asks without stopping, keeps working on what does not need your answer, waits only where the answer changes the outcome - keep it in one Codex session with the experimental notes setting on: it remembers across context windows why an earlier fix failed - expect the first pass to land: handed a program with no source, it worked out how it runs 88% of the time first try, 99.2% within four you never find out what was broken. it gets fixed anyway the catch is on the same page. roughly 30% more memory for that speed, and the safety checks can pause a long job until you approve the next step describing the problem was the expensive half of fixing it. that half just ended every hour you spend explaining the symptom to a chat window, someone else has handed theirs over whole bookmark this before the next thing breaks, the playbook for handing a whole job to an AI worker is in the piece below ↓

OpenAI. said. this. publicly. their own engineers just proved one idea on themselves, in writing: stop telling AI what's wrong. hand it the whole broken thing and let it find out they gave GPT-6 Astra a slow test build of their own coding tool. one cause found, a memory bottleneck, one allocator swapped, every turn 25× faster this is GPT-6 Astra, the layer that fixes the cause instead of the symptom, $0 on top of the ChatGPT plan you already pay for: - open ChatGPT or Codex, pick GPT-6 Astra, hand it the whole thing: the folder, the file that takes a minute to open. it works in apps with no API and reads your screen - type one sentence: find the one cause, prove it, fix it, do not patch around it - leave the room. it asks without stopping, keeps working on what does not need your answer, waits only where the answer changes the outcome - keep it in one Codex session with the experimental notes setting on: it remembers across context windows why an earlier fix failed - expect the first pass to land: handed a program with no source, it worked out how it runs 88% of the time first try, 99.2% within four you never find out what was broken. it gets fixed anyway the catch is on the same page. roughly 30% more memory for that speed, and the safety checks can pause a long job until you approve the next step describing the problem was the expensive half of fixing it. that half just ended every hour you spend explaining the symptom to a chat window, someone else has handed theirs over whole bookmark this before the next thing breaks, the playbook for handing a whole job to an AI worker is in the piece below ↓

Argona

109,831 次观看 • 1 个月前

I tested Step 5 Preview from StepFun on Duolingo’s FY25 filings. Duolingo reported $414.1M in net income. But $256.7M of that was a one-time tax benefit. Excluding it, net income was closer to $157.4M. Step 5 Preview caught that distinction without prompting, then kept going: → Revenue grew 39%, while bookings grew 33%. → 2026 guidance implied roughly flat adjusted EBITDA dollars despite higher revenue. → It caught and fixed two bugs in its own financial model before I reviewed the file. Then I changed one input. It automatically ran 64 fresh checks. All passed. That’s the part that stood out: not just finding an insight, but building a deliverable, checking it, and checking it again when assumptions changed. Step 5 Preview is now on OpenRouter and available in Kilo Code, Cline, Hermes Agent, and OpenCode. Try it here: 👉 Step 5 Preview has reached #1 on the OpenRouter Trending ranking. Check the ranking→

I tested Step 5 Preview from StepFun on Duolingo’s FY25 filings. Duolingo reported $414.1M in net income. But $256.7M of that was a one-time tax benefit. Excluding it, net income was closer to $157.4M. Step 5 Preview caught that distinction without prompting, then kept going: → Revenue grew 39%, while bookings grew 33%. → 2026 guidance implied roughly flat adjusted EBITDA dollars despite higher revenue. → It caught and fixed two bugs in its own financial model before I reviewed the file. Then I changed one input. It automatically ran 64 fresh checks. All passed. That’s the part that stood out: not just finding an insight, but building a deliverable, checking it, and checking it again when assumptions changed. Step 5 Preview is now on OpenRouter and available in Kilo Code, Cline, Hermes Agent, and OpenCode. Try it here: 👉 Step 5 Preview has reached #1 on the OpenRouter Trending ranking. Check the ranking→

D-Coder

115,317 次观看 • 5 天前

I set up Gemini Spark to write my own weekly AI news digest for me, and I'm still stuck on how little effort I had to put in. This is the exact prompt I gave it, just once: "Every Tuesday at 12:15 PM, create a new Doc in my Drive, find the 5 biggest AI stories from that week, and format each one properly." And here's what it actually produces, every week: 1. A new Doc, created and named on its own 2. The 5 biggest AI stories from that week, found and picked automatically 3. A bold headline and a summary for each one 4. The key impact spelled out, not just the news 5. A source link on every single story This runs on a schedule so I never have to think about it again. I didn't ask anything on Tuesday, I just opened the drive and the news digest doc was already there. Just like this, you can use Spark to automate any recurring task.

I set up Gemini Spark to write my own weekly AI news digest for me, and I'm still stuck on how little effort I had to put in. This is the exact prompt I gave it, just once: "Every Tuesday at 12:15 PM, create a new Doc in my Drive, find the 5 biggest AI stories from that week, and format each one properly." And here's what it actually produces, every week: 1. A new Doc, created and named on its own 2. The 5 biggest AI stories from that week, found and picked automatically 3. A bold headline and a summary for each one 4. The key impact spelled out, not just the news 5. A source link on every single story This runs on a schedule so I never have to think about it again. I didn't ask anything on Tuesday, I just opened the drive and the news digest doc was already there. Just like this, you can use Spark to automate any recurring task.

Eyisha Zyer

41,454 次观看 • 2 个月前

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 ↓

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

85,117 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

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 • 2 个月前

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 one easy. It never opened a browser. Never started the dev server. Never counted a voxel. It spent the entire run fighting its own export statements. petals.js got rewritten 19 times. Not patched, rewritten from scratch every time. water.js 9 times, palette.js 11, lanterns.js 7. It was loop between exactly 2 errors: Export 'update' is not defined in module. 53 times. Duplicate export of 'createPetals'. 10 times. Add the alias and you get the duplicate. Remove it and the export goes missing. In the last 70 steps it flipped between those 2 states 9 times. Twice it deleted the file and wrote it back identical. 85 writes against 7 reads. It almost never looked at what was already on disk. It compacted its own context 4 times and came back to the same 2 errors every time. It stopped with 3 files still failing node --check, all on the same line, update as updatePetals. 2 modules that main.js imports do not exist at all, and one of them is the lanterns file it had written 7 times. Then I gave the same model the same prompt as a single HTML file. It shipped a complete working scene in 1 pass. So it is not a capability wall. The module code it wrote is decent. It cannot hold its own module graph together, and it has nothing that tells it that it is going in circles. 438,737 output tokens and it never once said it was stuck. I spent 3 days making this thing run on my rig. An FP8 KV cache path that vLLM rejects on Ampere in 4 separate places. The 51B n-gram table baked down to FP8 so it fits in host RAM. W4A16 weights, the full 262K context, on 4 gaming GPUs from 2020. All of that works. And then it cannot wire 16 files together. So I'm going back to Qwen3.8 27B as my daily driver until something changes. Could be something in my own build doing this, I'm rebuilding the quant to find out. Either way I'll post what comes back.

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 one easy. It never opened a browser. Never started the dev server. Never counted a voxel. It spent the entire run fighting its own export statements. petals.js got rewritten 19 times. Not patched, rewritten from scratch every time. water.js 9 times, palette.js 11, lanterns.js 7. It was loop between exactly 2 errors: Export 'update' is not defined in module. 53 times. Duplicate export of 'createPetals'. 10 times. Add the alias and you get the duplicate. Remove it and the export goes missing. In the last 70 steps it flipped between those 2 states 9 times. Twice it deleted the file and wrote it back identical. 85 writes against 7 reads. It almost never looked at what was already on disk. It compacted its own context 4 times and came back to the same 2 errors every time. It stopped with 3 files still failing node --check, all on the same line, update as updatePetals. 2 modules that main.js imports do not exist at all, and one of them is the lanterns file it had written 7 times. Then I gave the same model the same prompt as a single HTML file. It shipped a complete working scene in 1 pass. So it is not a capability wall. The module code it wrote is decent. It cannot hold its own module graph together, and it has nothing that tells it that it is going in circles. 438,737 output tokens and it never once said it was stuck. I spent 3 days making this thing run on my rig. An FP8 KV cache path that vLLM rejects on Ampere in 4 separate places. The 51B n-gram table baked down to FP8 so it fits in host RAM. W4A16 weights, the full 262K context, on 4 gaming GPUs from 2020. All of that works. And then it cannot wire 16 files together. So I'm going back to Qwen3.8 27B as my daily driver until something changes. Could be something in my own build doing this, I'm rebuilding the quant to find out. Either way I'll post what comes back.

Alexey Fateev

31,282 次观看 • 1 个月前

How to create Farcaster mini app without coding - Step 1: > Go to ChatGPT > Just type: "I would like to create a Farcaster mini app, can you give me some ideas?" > Choose any one of the ideas - Note: - You can also type your own idea to ChatGPT and alter it - Step 2: - Go to ChatGPT > Paste the first prompt (shared in my channel) > Include the project you chose and tell ChatGPT to alter the prompt according to the idea below - Step 3 - Go to: > Sign up using Gmail or GitHub > Projects -> New Project -> Type the name of your project >Just paste the prompt we created earlier > It will start creating your app > If you need to alter anything, just go to GPT and type: [ I need a prompt to alter "your issue (like changing colour, etc.)" in my app already created in v0app by Vercel ] > Copy the prompt and paste it in v0 > No need to worry about errors , it will rectify them for you > If your project is finished, just click "Publish" on your top right > It will automatically host your app in Vercel > Click "Visit site" to see your app on another device and copy the website URL - Step 4 - Go to: v0 app > Just paste Prompt 2 from the file I provided > Click Publish once it’s completed - Step 5 - Go to: v0 app > Just paste the prompt 3 from the file I provided - Step 6 (Important) > Just paste Prompt 4 from the file I provided - Step 7 > Just paste Prompt 4 from the file I provided - Step 8 > Just paste Prompt 6 from the file I provided > It’s an important step to do — you need to create a manifest inside your app so that you can host it on Farcaster - Step 9 - Go to: > Sign up or Login > Settings → Developers → Mini Apps → Create Manifest (New) > Paste your website URL (remove https) > It should be like: > If you find any error while doing this, just copy and paste it in the v0 app > Once everything is finished, you need to create Account Association > Scan the QR on your mobile and tap and hold the Yes button > It will show an error — just copy and paste those errors into the v0 app > Click Publish again, and then click Refresh in Manifest > Click Open your app > for the main prompt file check our tg channel link in our bio

How to create Farcaster mini app without coding - Step 1: > Go to ChatGPT > Just type: "I would like to create a Farcaster mini app, can you give me some ideas?" > Choose any one of the ideas - Note: - You can also type your own idea to ChatGPT and alter it - Step 2: - Go to ChatGPT > Paste the first prompt (shared in my channel) > Include the project you chose and tell ChatGPT to alter the prompt according to the idea below - Step 3 - Go to: > Sign up using Gmail or GitHub > Projects -> New Project -> Type the name of your project >Just paste the prompt we created earlier > It will start creating your app > If you need to alter anything, just go to GPT and type: [ I need a prompt to alter "your issue (like changing colour, etc.)" in my app already created in v0app by Vercel ] > Copy the prompt and paste it in v0 > No need to worry about errors , it will rectify them for you > If your project is finished, just click "Publish" on your top right > It will automatically host your app in Vercel > Click "Visit site" to see your app on another device and copy the website URL - Step 4 - Go to: v0 app > Just paste Prompt 2 from the file I provided > Click Publish once it’s completed - Step 5 - Go to: v0 app > Just paste the prompt 3 from the file I provided - Step 6 (Important) > Just paste Prompt 4 from the file I provided - Step 7 > Just paste Prompt 4 from the file I provided - Step 8 > Just paste Prompt 6 from the file I provided > It’s an important step to do — you need to create a manifest inside your app so that you can host it on Farcaster - Step 9 - Go to: > Sign up or Login > Settings → Developers → Mini Apps → Create Manifest (New) > Paste your website URL (remove https) > It should be like: > If you find any error while doing this, just copy and paste it in the v0 app > Once everything is finished, you need to create Account Association > Scan the QR on your mobile and tap and hold the Yes button > It will show an error — just copy and paste those errors into the v0 app > Click Publish again, and then click Refresh in Manifest > Click Open your app > for the main prompt file check our tg channel link in our bio

Maran

48,427 次观看 • 10 个月前

You don't need to know how to code to turn a messy spreadsheet into something you can actually use. I had an Excel file of 20 brands I want to pitch. I dropped it in a folder, opened Qoder, and typed one sentence: turn this into a dashboard with search. Then I watched. It read the file, made a plan, wrote the code, opened the page in the browser, and tested the search on its own. I added a /goal so it knew when it was done: every brand shows, search works. The model: Qwen3.8-Flash, free through September 30. Zero Credits for this whole run. Want to try it? Enter my code BPU7C6 at first, then download Qoder and sign in. You will get 600 Credits: 300 from the trial and 300 from my code.

Jay

143,443 次观看 • 20 天前

I gave Grok Bot $27 and told it: "make money or I'll delete you" result: $27 → $1,084 in just over 25 hours and it's still running. I haven't opened the dashboard since yesterday. it lives on a server in the cloud and trades memes on its own while I sleep I keep waiting for the moment this thing falls apart. it just hasn't happened yet every moment, it: - pulls every new token launch and analyzes liquidity before the chart even loads - runs a rug check, mint authority, LP locks, top-holder concentration, and filters out anything that looks sketchy - analyzes social flow and the mempool to figure out what's actually about to move - only buys when the setup is clean enough to justify the slippage - sizes the next entry based on how the previous flip closed instead of using a fixed amount - swaps in and out on its own, without asking for my confirmation, and skims its own hosting fee off the top first first night it almost died, the balance dropped to $3 before it adapted. then it stopped chasing every shiny new launch, tightened the filters, and the curve finally started going up the whole setup took one evening: described the task, watched it execute a single swap, put it on a schedule, and connected a wallet no server to babysit. no keys to rotate. no code sitting there waiting to break there are quietly more systems like this than anyone is willing to admit, usually you only find out one exists when someone accidentally slips and posts a screenshot of the dashboard save this post now so you have the timestamp for when it finally clicks for you follow me if you want to replicate what I'm doing. I'll be sharing more alpha

I gave Grok Bot $27 and told it: "make money or I'll delete you" result: $27 → $1,084 in just over 25 hours and it's still running. I haven't opened the dashboard since yesterday. it lives on a server in the cloud and trades memes on its own while I sleep I keep waiting for the moment this thing falls apart. it just hasn't happened yet every moment, it: - pulls every new token launch and analyzes liquidity before the chart even loads - runs a rug check, mint authority, LP locks, top-holder concentration, and filters out anything that looks sketchy - analyzes social flow and the mempool to figure out what's actually about to move - only buys when the setup is clean enough to justify the slippage - sizes the next entry based on how the previous flip closed instead of using a fixed amount - swaps in and out on its own, without asking for my confirmation, and skims its own hosting fee off the top first first night it almost died, the balance dropped to $3 before it adapted. then it stopped chasing every shiny new launch, tightened the filters, and the curve finally started going up the whole setup took one evening: described the task, watched it execute a single swap, put it on a schedule, and connected a wallet no server to babysit. no keys to rotate. no code sitting there waiting to break there are quietly more systems like this than anyone is willing to admit, usually you only find out one exists when someone accidentally slips and posts a screenshot of the dashboard save this post now so you have the timestamp for when it finally clicks for you follow me if you want to replicate what I'm doing. I'll be sharing more alpha

cristal💎

60,024 次观看 • 1 个月前

OPUS 5.5 JUST DROPPED AND I PLUGGED IT INTO NERVE WITHIN AN HOUR. IT SELF VERIFIED ITS OWN TRADE THESIS THREE TIMES BEFORE FIRING. +5 ETH IN ONE SESSION. repo: every model before this proposed trades. opus 5.5 argues with itself before proposing. i replaced the analyst slot. same impulse. same spine. same jev for speed scoring. opus 5.5 now sits where astra used to sit but it does something astra never did. it checks its own work. first pool comes in. opus reads the narrative brief from grok. reads the jev score. reads the analyzer output. writes a thesis. then it attacks its own thesis. finds a weakness. rewrites. attacks again. three rounds. all inside one call. the impulse that exits the analyst slot is not a first impression. it is a thesis that survived its own author trying to kill it. astra wrote one paragraph and moved on. opus 5.5 writes three drafts and only sends the one it could not break. 1M context means it holds the entire narrative map, the last 200 pool results, the ghost log, and the current portfolio in one pass. no truncation. no summary. the full picture. first session i watched it reject a pool that jev scored 0.87. opus wrote the thesis, attacked it, found that the narrative was 14 hours old and decelerating despite fresh volume. the crowd was leaving not arriving. jev saw the numbers. opus saw the story behind the numbers was ending. then it cleared two pools in a row that grok flagged as early narrative. opus confirmed both, self verified, fired. pool 1: +2.1 ETH. pool 2: +2.9 ETH. +5 ETH. one session. the model that argues with itself before spending your money. jev decides in 190ms. opus thinks for 4 seconds. nerve runs both because speed and depth are not the same question.

OPUS 5.5 JUST DROPPED AND I PLUGGED IT INTO NERVE WITHIN AN HOUR. IT SELF VERIFIED ITS OWN TRADE THESIS THREE TIMES BEFORE FIRING. +5 ETH IN ONE SESSION. repo: every model before this proposed trades. opus 5.5 argues with itself before proposing. i replaced the analyst slot. same impulse. same spine. same jev for speed scoring. opus 5.5 now sits where astra used to sit but it does something astra never did. it checks its own work. first pool comes in. opus reads the narrative brief from grok. reads the jev score. reads the analyzer output. writes a thesis. then it attacks its own thesis. finds a weakness. rewrites. attacks again. three rounds. all inside one call. the impulse that exits the analyst slot is not a first impression. it is a thesis that survived its own author trying to kill it. astra wrote one paragraph and moved on. opus 5.5 writes three drafts and only sends the one it could not break. 1M context means it holds the entire narrative map, the last 200 pool results, the ghost log, and the current portfolio in one pass. no truncation. no summary. the full picture. first session i watched it reject a pool that jev scored 0.87. opus wrote the thesis, attacked it, found that the narrative was 14 hours old and decelerating despite fresh volume. the crowd was leaving not arriving. jev saw the numbers. opus saw the story behind the numbers was ending. then it cleared two pools in a row that grok flagged as early narrative. opus confirmed both, self verified, fired. pool 1: +2.1 ETH. pool 2: +2.9 ETH. +5 ETH. one session. the model that argues with itself before spending your money. jev decides in 190ms. opus thinks for 4 seconds. nerve runs both because speed and depth are not the same question.

h100envy

11,823 次观看 • 21 天前

New skill: self-managed-context (make the agent's context an editable file) It explains how to build agents that decide what to keep, update, or remove from the information they use to do their work. It can archive a long log while keeping the exact error, update its progress notes, or remove outdated information. Those edits then change what the model sees on its next turn. 1- Keep the system instructions and original task protected, outside the editable file. 2- Write the remaining conversation to a file, with labels for each message. 3- Let the agent edit that file using its usual code tools. 4- After each command, read the file back and use the updated messages for the next model call. Loading the skill ( alone into a fixed harness won't create live context editing, but it can help an agent build and then operate a harness that supports it. I gave the skill to a coding agent and had it build the harness itself. The task is a long stream of server logs that doesn't fit in the window. The agent reads it in 18 chunks, about 10k tokens in total, with a 5.5k budget. It has to report one incident ticket exactly and the final value of every config key. Same model & budget, three setups: 1- Model manages its own context 2- Harness forces a summary at 75% full 3- Keeps everything The video shows a real GPT-5.4 run. - Self-managed solved it 3 out of 3. - Keep-everything overflowed 3 out of 3. - Forced summary also solved it 3 out of 3. On GPT-5.4 the self-managed agent re-processed about 24% fewer prompt tokens than the forced summary. When it edited, it cut hard, so little was left after the edit to re-process (one edit took 5,537 tokens down to 771). On GPT-4.1 it saved nothing. It edited near the top of its context but kept most of what was below, and every edit forces everything after it to be re-processed. This is a small test at about 2x context pressure. The paper goes up to 24x, but imho the video below and the skill are a good way to start understanding the technique.

New skill: self-managed-context (make the agent's context an editable file) It explains how to build agents that decide what to keep, update, or remove from the information they use to do their work. It can archive a long log while keeping the exact error, update its progress notes, or remove outdated information. Those edits then change what the model sees on its next turn. 1- Keep the system instructions and original task protected, outside the editable file. 2- Write the remaining conversation to a file, with labels for each message. 3- Let the agent edit that file using its usual code tools. 4- After each command, read the file back and use the updated messages for the next model call. Loading the skill ( alone into a fixed harness won't create live context editing, but it can help an agent build and then operate a harness that supports it. I gave the skill to a coding agent and had it build the harness itself. The task is a long stream of server logs that doesn't fit in the window. The agent reads it in 18 chunks, about 10k tokens in total, with a 5.5k budget. It has to report one incident ticket exactly and the final value of every config key. Same model & budget, three setups: 1- Model manages its own context 2- Harness forces a summary at 75% full 3- Keeps everything The video shows a real GPT-5.4 run. - Self-managed solved it 3 out of 3. - Keep-everything overflowed 3 out of 3. - Forced summary also solved it 3 out of 3. On GPT-5.4 the self-managed agent re-processed about 24% fewer prompt tokens than the forced summary. When it edited, it cut hard, so little was left after the edit to re-process (one edit took 5,537 tokens down to 771). On GPT-4.1 it saved nothing. It edited near the top of its context but kept most of what was below, and every edit forces everything after it to be re-processed. This is a small test at about 2x context pressure. The paper goes up to 24x, but imho the video below and the skill are a good way to start understanding the technique.

Muratcan Koylan

15,070 次观看 • 13 天前

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?

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

190,860 次观看 • 1 个月前

everyone in iOS development should watch this. seriously, it might change the whole industry. i pointed claude code at a live ios device running on revyl, typed "test everything," and walked away. here's what's actually happening: ① you don't write the tests. no scripts, no selectors, no test plan. i never told it which screens to open or what to check. it read the app, decided what mattered, and tested it. the entire instruction was "test everything." ② it built its own test team. it looked at the app, clocked that it's basically four mini apps (rides, delivery, services, account), and split itself into 4 agents, one per surface. scoping coverage like that is usually a person's whole afternoon. it did it in seconds, unprompted. ③ all four ran at the same time, each on its own live device. this is where revyl comes in. every agent gets its own live ios session in the cloud, so four running apps get tested in parallel instead of taking turns on one simulator. serial testing turns coverage into a time tax. running all of it at once removes the tax. ④ it tests like a person, not like a script. each agent drives the app the way a user would, taps through the flows, and visually checks each screen against what it expected to see. nothing is pinned to a brittle element id, so renaming a button doesn't take down half your suite. that one detail is the most annoying thing about how we test today, and it just quietly goes away. ⑤ no xcuitest, no sims melting your laptop. i didn't write a single xcuitest script, and there were no simulators booting on my machine. the agents run on cloud devices, so coverage stops being capped by what your laptop can handle. the part that got me isn't that an agent tested an app. it's that i never told it how. i handed it a device and an intent, and it figured out the scoping, the parallelizing, and the driving on its own. if you still write and maintain mobile ui tests by hand, i'm not sure that lasts the year.

everyone in iOS development should watch this. seriously, it might change the whole industry. i pointed claude code at a live ios device running on revyl, typed "test everything," and walked away. here's what's actually happening: ① you don't write the tests. no scripts, no selectors, no test plan. i never told it which screens to open or what to check. it read the app, decided what mattered, and tested it. the entire instruction was "test everything." ② it built its own test team. it looked at the app, clocked that it's basically four mini apps (rides, delivery, services, account), and split itself into 4 agents, one per surface. scoping coverage like that is usually a person's whole afternoon. it did it in seconds, unprompted. ③ all four ran at the same time, each on its own live device. this is where revyl comes in. every agent gets its own live ios session in the cloud, so four running apps get tested in parallel instead of taking turns on one simulator. serial testing turns coverage into a time tax. running all of it at once removes the tax. ④ it tests like a person, not like a script. each agent drives the app the way a user would, taps through the flows, and visually checks each screen against what it expected to see. nothing is pinned to a brittle element id, so renaming a button doesn't take down half your suite. that one detail is the most annoying thing about how we test today, and it just quietly goes away. ⑤ no xcuitest, no sims melting your laptop. i didn't write a single xcuitest script, and there were no simulators booting on my machine. the agents run on cloud devices, so coverage stops being capped by what your laptop can handle. the part that got me isn't that an agent tested an app. it's that i never told it how. i handed it a device and an intent, and it figured out the scoping, the parallelizing, and the driving on its own. if you still write and maintain mobile ui tests by hand, i'm not sure that lasts the year.

Landseer Enga

23,963 次观看 • 4 个月前

THIS GUY CONNECTED HIS AI AGENTS TO HIS OBSIDIAN AND BUILT A BRAIN THAT LEARNS ON ITS OWN. HERE'S HOW TO BUILD IT Obsidian is just markdown files sitting in a folder. That turns out to be the perfect memory for an AI agent, because an agent can read and write those files directly. He wired his agents into the vault so they pull context from it, do the work, and write what they learned back. The notes aren't the point. The loop is, and it gets sharper every cycle How to build it: 1. Point an agent at your vault. The fastest way, no plugins, no API keys: open a terminal and run npx obsidian-mcp /path/to/your/vault. That exposes your Obsidian folder to Claude as a tool it can read, search, and write to. Add it to your Claude Code or Cowork config and restart 2. Confirm it can see the brain. Ask it: "list the notes in my vault and summarize what's in them." If it reads them back, the connection is live. Now it starts every task with everything the vault already holds instead of from zero 3. Give each agent one job and a write-back rule. Tell it: "research this, then save what you found as a new note in /brain with links to related notes." One agent researches, one summarizes, one plans. Each writes its output back into the vault 4. Close the loop. Add one line to every agent's instructions: "read /brain before starting, write your result back when done." Now each task leaves the vault richer, and the next run reads that before it works. It compounds instead of resetting 5. You only steer. Review what the brain produces, point it at the next thing. The agents handle the reading, writing, and connecting The edge isn't better notes. It's a brain that feeds itself, so the work gets sharper every cycle instead of starting over Bookmark this

THIS GUY CONNECTED HIS AI AGENTS TO HIS OBSIDIAN AND BUILT A BRAIN THAT LEARNS ON ITS OWN. HERE'S HOW TO BUILD IT Obsidian is just markdown files sitting in a folder. That turns out to be the perfect memory for an AI agent, because an agent can read and write those files directly. He wired his agents into the vault so they pull context from it, do the work, and write what they learned back. The notes aren't the point. The loop is, and it gets sharper every cycle How to build it: 1. Point an agent at your vault. The fastest way, no plugins, no API keys: open a terminal and run npx obsidian-mcp /path/to/your/vault. That exposes your Obsidian folder to Claude as a tool it can read, search, and write to. Add it to your Claude Code or Cowork config and restart 2. Confirm it can see the brain. Ask it: "list the notes in my vault and summarize what's in them." If it reads them back, the connection is live. Now it starts every task with everything the vault already holds instead of from zero 3. Give each agent one job and a write-back rule. Tell it: "research this, then save what you found as a new note in /brain with links to related notes." One agent researches, one summarizes, one plans. Each writes its output back into the vault 4. Close the loop. Add one line to every agent's instructions: "read /brain before starting, write your result back when done." Now each task leaves the vault richer, and the next run reads that before it works. It compounds instead of resetting 5. You only steer. Review what the brain produces, point it at the next thing. The agents handle the reading, writing, and connecting The edge isn't better notes. It's a brain that feeds itself, so the work gets sharper every cycle instead of starting over Bookmark this

Yarchi

58,643 次观看 • 4 个月前