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。

We did a comparison study on where a robot's intelligence should live by using the same model (OpenAI GPT6-Astra), same bimanual YAM station, same LEGO pick-and-place task, all run on SE3' real-world remote eval stack: 1. Frozen code-as-policy (built on NVIDIA ENPIRE components): 6/20 placements 2. Agentic control +... show more

Paige

4,272 subscribers

17,096 次观看 • 5 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

6 条评论

Paige 的头像

Paige5 天前

Here’s how the two approaches worked: 1. Code-as-policy: Astra wrote and tested a Python policy, which we then froze for evaluation. The policy ran without further model calls. 2. Agentic control: Each episode started with a fresh Astra session that read camera snapshots and issued commands through a control harness.

Paige 的头像

Paige5 天前

Getting hold of the brick wasn’t enough. What mattered was robustness and recovery. The code-as-policy lifted the brick in 14/20 episodes but completed only 6 placements. In the other eight, it stopped before reaching the release pose or held the brick without releasing it. Agentic control completed 18/20 placements. Nine required grasp retries, including one that recovered from a drop.

Paige 的头像

Paige5 天前

The agent recovered from its own misses: 9 of its 18 completions needed more than one grasp. In episode 8, it missed five times, lifted on the sixth, dropped the brick, reacquired it and released at 419s. The code-as-policy never made a second grasp attempt.

Paige 的头像

Paige5 天前

Agentic control completed more placements, but successful runs took longer. Among successful episodes, median time to release was 131 seconds for code-as-policy versus 175 seconds for agentic control. Agentic control also had more time to recover: up to 600 seconds, compared with 180 seconds for code-as-policy. The timing gap reflects more than model reasoning alone.

Paige 的头像

Paige5 天前

Full report and all 40 episodes on video: Reach out if you have a policy or an idea you’d like to evaluate on real robots!

catman 的头像

catman5 天前

@OpenAI @se3labs @nvidia Keeping the model in the loop lets it adjust each movement from live feedback, rather than letting errors compound through a fixed sequence of actions.

相关视频

SITUATION EXPLAINED: Grok 4.7 beats GPT-6 Astra on real-world tasks, and it's 5x cheaper. • Same price as 4.6, $2 input and $6 output per million tokens, well under Sol, Fable, and Astra • A new larger base model, which Elon has put at 2.1 trillion parameters, up from 1.5 trillion, with extra training on SpaceX data • Trained to natively understand the Grok Bot harness, the same move Anthropic made with Claude Code • Leads on electrical engineering and Harvey's legal benchmark, beats Astra Max on GDPval, trails on software engineering • Elon Musk: "Grok 4.7 places SpaceXAI as third after Anthropic and OpenAI for agentic coding. When factoring in that Grok is significantly faster and lower cost, it's a great choice for your everyday workhorse" Theo Jaffee: "SpaceX has a huge amount of data on real world hardware problems, real world engineering. So I bet Grok models are going to be better at rocketry engineering than any of the other models out there."

SITUATION EXPLAINED: Grok 4.7 beats GPT-6 Astra on real-world tasks, and it's 5x cheaper. • Same price as 4.6, $2 input and $6 output per million tokens, well under Sol, Fable, and Astra • A new larger base model, which Elon has put at 2.1 trillion parameters, up from 1.5 trillion, with extra training on SpaceX data • Trained to natively understand the Grok Bot harness, the same move Anthropic made with Claude Code • Leads on electrical engineering and Harvey's legal benchmark, beats Astra Max on GDPval, trails on software engineering • Elon Musk: "Grok 4.7 places SpaceXAI as third after Anthropic and OpenAI for agentic coding. When factoring in that Grok is significantly faster and lower cost, it's a great choice for your everyday workhorse" Theo Jaffee: "SpaceX has a huge amount of data on real world hardware problems, real world engineering. So I bet Grok models are going to be better at rocketry engineering than any of the other models out there."

MTS

18,190 次观看 • 7 天前

Sachin Katti says OpenAI had Astra optimize its own inference on NVIDIA's Rubin chip, adding another 2x in 72 hours "It's a testament to NVIDIA's hardware that out of the box we got a 3x improvement in throughput. We literally took the same code that we had optimized for Blackwell, ran it on Rubin, got 3x improvement." "Then, can we actually make Astra run agents using Astra as the base model, go optimize this inference of Astra itself on Rubin? It's a recursive loop. The model is figuring out how to optimize its own inference serving on Rubin chips, and over 72 hours the model produced another 2x improvement on the throughput." "Things that would take weeks, if not months, to optimize in the past, we now are getting models that are very capable of actually producing over a weekend." _______ Link and more from Sachin Katti's conversation:

Sachin Katti says OpenAI had Astra optimize its own inference on NVIDIA's Rubin chip, adding another 2x in 72 hours "It's a testament to NVIDIA's hardware that out of the box we got a 3x improvement in throughput. We literally took the same code that we had optimized for Blackwell, ran it on Rubin, got 3x improvement." "Then, can we actually make Astra run agents using Astra as the base model, go optimize this inference of Astra itself on Rubin? It's a recursive loop. The model is figuring out how to optimize its own inference serving on Rubin chips, and over 72 hours the model produced another 2x improvement on the throughput." "Things that would take weeks, if not months, to optimize in the past, we now are getting models that are very capable of actually producing over a weekend." _______ Link and more from Sachin Katti's conversation:

Fireside Alpha

40,013 次观看 • 11 天前

1. Sam Altman on Will All Startups Be Killed by OpenAI There are 2 strategies to build on AI right now: Assume the model won’t improve. Assume OpenAI will stay on the same growth trajectory. 95% should bet on #2. We have a mission to improve the model, so we will steamroll you.

1. Sam Altman on Will All Startups Be Killed by OpenAI There are 2 strategies to build on AI right now: Assume the model won’t improve. Assume OpenAI will stay on the same growth trajectory. 95% should bet on #2. We have a mission to improve the model, so we will steamroll you.

Harry Stebbings

212,291 次观看 • 2 年前

Liquid AI's head of post-training explained how they built a small model that runs on-device under 1 GB in 20 minutes - better than $2500 small-model bootcamps. pick LFM2.5 base -> on-policy preference alignment -> agentic reinforcement learning -> curriculum training -> iterative model merging -> ship a 1B model that reliably calls tools on your phone. That loop is why frontier small models now beat 70B models on the tasks that actually matter. LFM2.5 + on-policy DPO + agentic RL + curriculum training + iterative merging - that's the stack. Watch and save it, then run a 1B agent on your phone tonight.

Liquid AI's head of post-training explained how they built a small model that runs on-device under 1 GB in 20 minutes - better than $2500 small-model bootcamps. pick LFM2.5 base -> on-policy preference alignment -> agentic reinforcement learning -> curriculum training -> iterative model merging -> ship a 1B model that reliably calls tools on your phone. That loop is why frontier small models now beat 70B models on the tasks that actually matter. LFM2.5 + on-policy DPO + agentic RL + curriculum training + iterative merging - that's the stack. Watch and save it, then run a 1B agent on your phone tonight.

h100envy

86,913 次观看 • 2 个月前

GPT-6 Astra smokes Claude Fable 5.1 in Blender 💨 We compared GPT-6 Astra and Fable 5.1 on a Porsche 911 GT3 RS build and the difference is visible from the very first shot Output: Astra: $132.03 • 2.58 hours Fable: $114.41 • 2.56 hours Astra's model is not only more detailed in every aspect, but also far closer to the original car, costing almost the same as Fable's blocky funny looking model, the difference is astonishing

GPT-6 Astra smokes Claude Fable 5.1 in Blender 💨 We compared GPT-6 Astra and Fable 5.1 on a Porsche 911 GT3 RS build and the difference is visible from the very first shot Output: Astra: $132.03 • 2.58 hours Fable: $114.41 • 2.56 hours Astra's model is not only more detailed in every aspect, but also far closer to the original car, costing almost the same as Fable's blocky funny looking model, the difference is astonishing

atomic.chat

132,877 次观看 • 21 天前

36 GROK AGENTS. ASTRA ON FREE CREDITS. DUAL-MODEL ROUTING IS THE EDGE. not one chat window burning a paid invoice all day a swarm that routes cheap work to grok and only wakes gpt-6 astra when the task is actually hard ▹ the stack 36 agents on grok bot with flexible settings per role monitor, plan, write, code, QA, ship, each with its own lane part of the fleet is wired to gpt-6 astra through a china free-credit service layer free credits are not a toy promo here they are the fuel for frontier spikes without a monthly bleed ▹ dual-model routing easy jobs stay on grok: speed, volume, always-on loops hard jobs jump to astra: reasoning, long builds, sharp code the router decides by task type, not by ego if astra is not needed, astra does not spend free-credit bursts buy the expensive brain grok agents keep the factory running between bursts that is how the system feels unlimited not by breaking quotas, by refusing to waste them ▹ why it hits different most people pay frontier prices for every mid task operators split the brain and protect the credits 36 agents = parallel throughput dual routing = cost control with quality when it matters free credits = astra access without living on the invoice the constraint moved from "can i afford the model" to "did i route the job to the right model" ▹ the take single-model stacks die on bills and on boredom multi-agent + dual routing is the new default factory grok for the grind astra for the cut free credits for the spikes that used to empty the wallet bookmark this before everyone copies the route map comments: what % of your tasks actually deserve astra

36 GROK AGENTS. ASTRA ON FREE CREDITS. DUAL-MODEL ROUTING IS THE EDGE. not one chat window burning a paid invoice all day a swarm that routes cheap work to grok and only wakes gpt-6 astra when the task is actually hard ▹ the stack 36 agents on grok bot with flexible settings per role monitor, plan, write, code, QA, ship, each with its own lane part of the fleet is wired to gpt-6 astra through a china free-credit service layer free credits are not a toy promo here they are the fuel for frontier spikes without a monthly bleed ▹ dual-model routing easy jobs stay on grok: speed, volume, always-on loops hard jobs jump to astra: reasoning, long builds, sharp code the router decides by task type, not by ego if astra is not needed, astra does not spend free-credit bursts buy the expensive brain grok agents keep the factory running between bursts that is how the system feels unlimited not by breaking quotas, by refusing to waste them ▹ why it hits different most people pay frontier prices for every mid task operators split the brain and protect the credits 36 agents = parallel throughput dual routing = cost control with quality when it matters free credits = astra access without living on the invoice the constraint moved from "can i afford the model" to "did i route the job to the right model" ▹ the take single-model stacks die on bills and on boredom multi-agent + dual routing is the new default factory grok for the grind astra for the cut free credits for the spikes that used to empty the wallet bookmark this before everyone copies the route map comments: what % of your tasks actually deserve astra

cryptopsihoz

44,187 次观看 • 19 天前

Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more

Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more

Prime Intellect

1,150,419 次观看 • 10 个月前

The new stealth model on OpenRouter Union Alpha is an OpenAI model. And here's why. I gave it my standard shooter prompt and got this FPS. I've built around a dozen games with GPT-6 Astra by now. I can spot its style from a mile away. Just look at the menu. The huge title, orange button with that same arrow, tiny labels over a dark 3D scene. Ive seen this menu so many times already. My guess is GPT-6 Sol. Come on OpenAI, at least change the menu. Playable demo link as always:

The new stealth model on OpenRouter Union Alpha is an OpenAI model. And here's why. I gave it my standard shooter prompt and got this FPS. I've built around a dozen games with GPT-6 Astra by now. I can spot its style from a mile away. Just look at the menu. The huge title, orange button with that same arrow, tiny labels over a dark 3D scene. Ive seen this menu so many times already. My guess is GPT-6 Sol. Come on OpenAI, at least change the menu. Playable demo link as always:

Alexey Fateev

24,455 次观看 • 12 天前

Today we’re using a arm to prototype our first RaaS workflow — soon, the arm doing this will be ours. Using a Franka arm, we prototyped the pick-and-place of Forma logo, a pipeline that will run on the Forma robotic arm: • trajectory planning • precision gripping • closed-loop positioning • object handling It’s our first RaaS workflow running in the real world. We’re building RaaS like software: test → ship → iterate. Next: deploy the same pipeline on the Forma arm and bring automation to the cloud.

Forma Robotics

31,082 次观看 • 10 个月前

We raised $3M on one observation: Agents learn on your dime. And forget. Every run, your agent figures out the same APIs, makes the same calls, hits the same failures. Because nothing saved what worked. We built Rote by Modiqo Inc to fix this. Point it at any API. Rote reads the spec and builds a surface your agent can call: auth, endpoints, schemas. That's the adapter. When your agent finishes a task, Rote saves the exact calls that worked. That's the flow. Config in. Code out. Runs on your machine. Nothing to watch over. Agents stop starting from scratch. Teams stop paying for the same work twice. Install →

We raised $3M on one observation: Agents learn on your dime. And forget. Every run, your agent figures out the same APIs, makes the same calls, hits the same failures. Because nothing saved what worked. We built Rote by Modiqo Inc to fix this. Point it at any API. Rote reads the spec and builds a surface your agent can call: auth, endpoints, schemas. That's the adapter. When your agent finishes a task, Rote saves the exact calls that worked. That's the flow. Config in. Code out. Runs on your machine. Nothing to watch over. Agents stop starting from scratch. Teams stop paying for the same work twice. Install →

chetan conikee

469,414 次观看 • 4 个月前

Introducing WorldForge: testable world-model workflows for physical AI systems. You can think of it, loosely, as “LangChain for world models”. The problem is that “world model” has become an overloaded label. Depending on context, it can mean a video generator, a cost model, a robot policy, a JEPA-style latent predictor, etc. They share almost nothing, different inputs, runtimes, failure modes. I built WorldForge to stop pretending they're interchangeable. Front-door demo: a real Hugging Face LeRobot (LeRobot) diffusion_pusht policy combined with LeWorldModel by Lucas Maes checkpoint for scoring. Both run locally on my MacBook in the demo video. LeWM is extremely efficient (~15M params), can plan up to 48× faster, and runs on commodity hardware. WorldForge wires the loop: policy → candidates → score → select Replay happens in a local TUI today, but the same loop could drive a real robot in the world. Would love feedback from people working on world models, physical AI, robotics, ML infra, and adjacent tooling. Fully open source. Contributions very welcome. Plan in the dream, replay in real world.

Introducing WorldForge: testable world-model workflows for physical AI systems. You can think of it, loosely, as “LangChain for world models”. The problem is that “world model” has become an overloaded label. Depending on context, it can mean a video generator, a cost model, a robot policy, a JEPA-style latent predictor, etc. They share almost nothing, different inputs, runtimes, failure modes. I built WorldForge to stop pretending they're interchangeable. Front-door demo: a real Hugging Face LeRobot (LeRobot) diffusion_pusht policy combined with LeWorldModel by Lucas Maes checkpoint for scoring. Both run locally on my MacBook in the demo video. LeWM is extremely efficient (~15M params), can plan up to 48× faster, and runs on commodity hardware. WorldForge wires the loop: policy → candidates → score → select Replay happens in a local TUI today, but the same loop could drive a real robot in the world. Would love feedback from people working on world models, physical AI, robotics, ML infra, and adjacent tooling. Fully open source. Contributions very welcome. Plan in the dream, replay in real world.

abdel

17,001 次观看 • 5 个月前

GPT6-Astra turned $500 into $5,110 on my trading account last night. Six AI agents, 0.4 seconds a round, and I was asleep for all of it. 255 setups came up. 1,151 configs died. One trade reached my money. The agent holding the veto killed 4 of the last 5. I don't get a vote either. You have the same model open in a tab right now. You're using it to ask whether the market looks bullish. Six seats, one evening, and the whole roster fits in six prompts. The names and the rules they run on are in the post below ↓

GPT6-Astra turned $500 into $5,110 on my trading account last night. Six AI agents, 0.4 seconds a round, and I was asleep for all of it. 255 setups came up. 1,151 configs died. One trade reached my money. The agent holding the veto killed 4 of the last 5. I don't get a vote either. You have the same model open in a tab right now. You're using it to ask whether the market looks bullish. Six seats, one evening, and the whole roster fits in six prompts. The names and the rules they run on are in the post below ↓

cvxv666

257,046 次观看 • 18 天前

I GAVE GPT-6 ASTRA AND GROK BOT $1,000 EACH AND THE BETTER TRADER F*CKING LOST grok closed $1,742.15. astra closed $1,634.27. eleven points apart same day, same page, six agents each. one order: beat the other room or i cut funding astra was the better trader on every line people brag about → coin side: astra $586.42, grok $468.90 → astra went 17 of 24 green, grok went 16 of 24 → astra caught the MARLIN low at 57.4K and the 818K top, +$462.44 on two tickets → astra took 92% of its day out of that one coin → grok took 37% of his day out of prediction books, +$273.25 → astra ran the same books for $47.85 so how did the worse trader take the money? a calendar apple held a keynote that day. grok bought IPOD at 559K into it, +$306.48, biggest ticket of the duel. then he traded the market pricing it wrong: "apple foldable iphone by sep 30", NO at 45c, closed at 97.9c both opened the same bitcoin book over 80k. astra took YES at 34.5c and watched it bleed to 13.5c, its worst book of the day. grok took NO at 58.5c and closed it at 93.5c same market, same hour, opposite sides. no benchmark tells you that. leaderboards score answers, not tickets both floors trade on one page anyone can open: a manhattan seat that reads a calendar costs $180,000 a year. mine cost less than dinner and argued all night the six-role setup i handed both sits in the article i quoted, free you can rent the smarter model, not the habit of reading tomorrow whole duel is in the clip below, tick by tick. bookmark it before the feed eats it which one would you fund after that scorecard?

I GAVE GPT-6 ASTRA AND GROK BOT $1,000 EACH AND THE BETTER TRADER F*CKING LOST grok closed $1,742.15. astra closed $1,634.27. eleven points apart same day, same page, six agents each. one order: beat the other room or i cut funding astra was the better trader on every line people brag about → coin side: astra $586.42, grok $468.90 → astra went 17 of 24 green, grok went 16 of 24 → astra caught the MARLIN low at 57.4K and the 818K top, +$462.44 on two tickets → astra took 92% of its day out of that one coin → grok took 37% of his day out of prediction books, +$273.25 → astra ran the same books for $47.85 so how did the worse trader take the money? a calendar apple held a keynote that day. grok bought IPOD at 559K into it, +$306.48, biggest ticket of the duel. then he traded the market pricing it wrong: "apple foldable iphone by sep 30", NO at 45c, closed at 97.9c both opened the same bitcoin book over 80k. astra took YES at 34.5c and watched it bleed to 13.5c, its worst book of the day. grok took NO at 58.5c and closed it at 93.5c same market, same hour, opposite sides. no benchmark tells you that. leaderboards score answers, not tickets both floors trade on one page anyone can open: a manhattan seat that reads a calendar costs $180,000 a year. mine cost less than dinner and argued all night the six-role setup i handed both sits in the article i quoted, free you can rent the smarter model, not the habit of reading tomorrow whole duel is in the clip below, tick by tick. bookmark it before the feed eats it which one would you fund after that scorecard?

savip.

36,536 次观看 • 18 天前

Tele-operated humanoid combat did not exist as a sport. Building it meant solving control, feedback, and durability at the same time. We did. SCL is the first league for human-piloted humanoid combat. Driven by ex-professional MMA fighter Peter Davis and support from Virtuals Protocol, building the sport, the stack, and the arena in-house became a possibility. $SCL is now live on at

Tele-operated humanoid combat did not exist as a sport. Building it meant solving control, feedback, and durability at the same time. We did. SCL is the first league for human-piloted humanoid combat. Driven by ex-professional MMA fighter Peter Davis and support from Virtuals Protocol, building the sport, the stack, and the arena in-house became a possibility. $SCL is now live on at

Shadow Combat League

21,312 次观看 • 5 个月前

OpenAI has launched a version of ChatGPT built specifically for investment bankers and equity researchers, aiming squarely at Wall Street's junior banker workflows. The tool is powered by OpenAI's new GPT-6 Astra model and comes as a tailored ChatGPT work experience with premium financial data built in. VP of Product Nick Turley said the goal is teaching ChatGPT to research like an analyst and back up its conclusions the same way an analyst would. With it, teams can research companies, analyze earnings, compare peers, test assumptions, and build editable financial models, research notes, and client materials, all inside the same interface.

OpenAI has launched a version of ChatGPT built specifically for investment bankers and equity researchers, aiming squarely at Wall Street's junior banker workflows. The tool is powered by OpenAI's new GPT-6 Astra model and comes as a tailored ChatGPT work experience with premium financial data built in. VP of Product Nick Turley said the goal is teaching ChatGPT to research like an analyst and back up its conclusions the same way an analyst would. With it, teams can research companies, analyze earnings, compare peers, test assumptions, and build editable financial models, research notes, and client materials, all inside the same interface.

AGTP

12,794 次观看 • 18 天前

🚀 We’re excited to announce LingBot-VA, a new state-of-the-art robot policy model from Robbyant ! LingBot-VA is built on a causal, autoregressive video-action world model for generalist robot control. Highlights: (1) First unified autoregressive video-action world model for robot control (2) Low-latency inference with a new asynchronous execution pipeline (3) SOTA on RoboTwin (92.9%, firstever > 90%) and LIBERO (98.5%) (4) +20% over π0.5 on challenging real-world long-horizon & high-precision tasks

🚀 We’re excited to announce LingBot-VA, a new state-of-the-art robot policy model from Robbyant ! LingBot-VA is built on a causal, autoregressive video-action world model for generalist robot control. Highlights: (1) First unified autoregressive video-action world model for robot control (2) Low-latency inference with a new asynchronous execution pipeline (3) SOTA on RoboTwin (92.9%, firstever > 90%) and LIBERO (98.5%) (4) +20% over π0.5 on challenging real-world long-horizon & high-precision tasks

Yinghao Xu

47,604 次观看 • 8 个月前

ex-Apple engineer gave Grok 4.6 - 30 minutes to redesign a live app, then showed the GrokBot run that merged its own PRs: "GrokBot runs 20-30 agents in parallel while Hermes and OpenClaw run one. Same weights, 20x the throughput - and the agents keep improving from their own runs." Grok 4.6 runs 5x cheaper on input and 8x on output - $2 and $6 per million. his verdict: this is an easy Opus replacement worth more than any $500 course on agentic engineering watch today, then read the article below on building a self-running agent stack

ex-Apple engineer gave Grok 4.6 - 30 minutes to redesign a live app, then showed the GrokBot run that merged its own PRs: "GrokBot runs 20-30 agents in parallel while Hermes and OpenClaw run one. Same weights, 20x the throughput - and the agents keep improving from their own runs." Grok 4.6 runs 5x cheaper on input and 8x on output - $2 and $6 per million. his verdict: this is an easy Opus replacement worth more than any $500 course on agentic engineering watch today, then read the article below on building a self-running agent stack

Carnage

82,677 次观看 • 1 个月前

BRO ALMOST USED THE NEW MODEL FOR FREE AND MADE $1,240 IN 13 HOURS. He pulled an open-source trading bot for an ETH wallet off GitHub. A stack of cheaper models - Luna, Terra, Sol went through the code first, flagged nothing critical. Astra got pulled in once, for a single review pass on top of that - not to write anything, just to check the cheaper models' work before it touched a live wallet. For the actual run, Astra's only job was watching the process - not debugging, not generating, just monitoring. That's almost no output tokens, which is why the expensive model barely showed up on the bill. Cheap models did the volume work, Astra did one narrow supervisory pass - same shape as the guide's routing table, just applied to a live bot instead of a coding session. Over 13 hours, the bot cleared $1,240. The part worth taking from this isn't the number. It's that "using the expensive model for free" really means using it for almost nothing - a small, specific job instead of the whole pipeline.

BRO ALMOST USED THE NEW MODEL FOR FREE AND MADE $1,240 IN 13 HOURS. He pulled an open-source trading bot for an ETH wallet off GitHub. A stack of cheaper models - Luna, Terra, Sol went through the code first, flagged nothing critical. Astra got pulled in once, for a single review pass on top of that - not to write anything, just to check the cheaper models' work before it touched a live wallet. For the actual run, Astra's only job was watching the process - not debugging, not generating, just monitoring. That's almost no output tokens, which is why the expensive model barely showed up on the bill. Cheap models did the volume work, Astra did one narrow supervisory pass - same shape as the guide's routing table, just applied to a live bot instead of a coding session. Over 13 hours, the bot cleared $1,240. The part worth taking from this isn't the number. It's that "using the expensive model for free" really means using it for almost nothing - a small, specific job instead of the whole pipeline.

draven

17,860 次观看 • 8 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 ↓

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

84,666 次观看 • 19 天前

🌌 GPT-6 Astra from OpenAI Developers is now generally available in GitHub Copilot. This new model is designed for long-horizon, autonomous coding and agentic tasks. Our internal testing shows it plans and validates as it goes, batches diagnosis with verification, and independently confirms its results before declaring a task done. That translated into stronger performance on extensive coding jobs with fewer steps than prior OpenAI models. Try it out in the GitHub Copilot app, CLI, or Visual Studio Code. 🚀

🌌 GPT-6 Astra from OpenAI Developers is now generally available in GitHub Copilot. This new model is designed for long-horizon, autonomous coding and agentic tasks. Our internal testing shows it plans and validates as it goes, batches diagnosis with verification, and independently confirms its results before declaring a task done. That translated into stronger performance on extensive coding jobs with fewer steps than prior OpenAI models. Try it out in the GitHub Copilot app, CLI, or Visual Studio Code. 🚀

GitHub

175,100 次观看 • 24 天前