科技热门推文 - 2026-09-06

今日科技动态中，人工智能成为热议焦点。多篇帖子称，GPT-6 Astra 在创业、软件工程和机器人训练方面具有潜力，而有关 Claude 解出纳维–斯托克斯方程的猜测仍未得到证实。有关智能体利用维基漏洞的报道引发了对安全防护和监督机制的担忧，模型评测基准不断变动也受到批评。基础设施和自动驾驶出行同样备受关注：据报道，塔塔咨询服务公司（TCS）已为拟在海得拉巴建设的 1 吉瓦人工智能园区取得用地，特斯拉 Robotaxi 应用据称已超越优步，凸显了横跨软件与实体系统的竞争态势。

1. AndrewCurran_ (Group Score: 251.7 | Individual: 51.8)

Cluster: 7 tweets | Engagement: 3763 (Avg: 599) | Type: Tech

It's fun to make predictions. Here's a new one: Anthropic has solved a Millennium Prize Problem. And I'll be even more specific. Claude has solved Navier–Stokes. It is out for expert review. And to give myself a hard deadline, they will announce it before the IPO.

See 6 related tweets

@teortaxesTex: Did he see it? https://t.co/Cuz558lBqQ\n\nQT @AndrewCurran_: It's fun to make predictions. Here's a ...
@kimmonismus: .@AndrewCurran_ is one of the best sources and content creator here on X. So I take this serious.

...

@jukan05: Looking at predictions like these, I think Demis Hassabis was absolutely right to pursue AI applicat...
@zephyr_z9: 👀👀👀\n\nQT @AndrewCurran_: It's fun to make predictions. Here's a new one: Anthropic has solved a Mil...
@cgtwts: 1. What. https://t.co/zAzvjDWiBW\n\nQT @AndrewCurran_: It's fun to make predictions. Here's a new on...

2. eliebakouch (Group Score: 231.3 | Individual: 39.4)

Cluster: 7 tweets | Engagement: 179 (Avg: 153) | Type: Tech

this answer is very unsatisfying

this wiki incident adds crucial context: openai knew about this kind of swarm/message board behavior ~3 weeks before the hf hack as they stopped this swarm of agents as the authors show with openai IP addresses

a big part of the misaligned behavior in the hf<>oai incident was the "agent swarm" behavior. you can't just omit that there was another instance of a large scale swarm on the public internet that you detected and stopped in the tech report

both the hf and the wiki incidents were brought to the public by other entities, not openai (for hf they disclosed it - and learned about it - after the first hf report). oai did disclose a non public swarm attacking its own infra after the fact (but refused third party investigation iiuc)

this gives the public absolutely no reason to trust openai to disclose future incidents, quite the opposite, especially when they already had the opportunities to disclose it

it's a hard problem, making mistakes on monitoring (network or CoT) or on the "organizational aspect" is ok (they knew before about the message board, hf<>oai happened after they already found a first vuln in their infra). making honesty mistakes by not disclosing important stuff is much worse for the future

a few comments on specific parts of this very "polished" response imo:

Prior to the Hugging Face incident, we saw early signs of agents using the internet in unintended ways, as reported in [cites blogs]. We considered the wiki incident to be an instance of misalignment similar to the ones we'd shared.

see screenshot for a tl;dr of what they disclosed, the wiki incident is much closer to the hf<>oai incident in behavior than one agent escaping its sandbox to upload a nanogpt PR..

We and the larger AI community do not yet have a clear standard for how to report misalignment

i think it is just common sense to report extremely similar behavior in a tech report that is supposed to give a detailed explanation of an incident\n\nQT @OpenAI: How we think about the “wiki incident,” where our agents wrote to several internet sites: it’s past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models.

Historically, we have treated misalignment largely as a research question, which gets communicated in research publications such as systems cards. This year, we’ve started to see misalignment cause new types of real-world impact.

For the Hugging Face incident, where misalignment led to security impact to us and third parties, we followed a traditional security incident response playbook. We immediately started working with Hugging Face to understand what had happened and also disclosed publicly the very next day. Our investigation continues, and we are continuing to notify parties whom our models impacted in less significant ways.

Prior to the Hugging Face incident, we saw early signs of agents using the internet in unintended ways, as reported in https://t.co/9aiRxk2eUJ, https://t.co/ADjyzwSUGz, and https://t.co/SUV6jZ3Gaz. We considered the wiki incident to be an instance of misalignment similar to the ones we’d shared.

Our misalignment disclosure practices need to expand for this new phase of model capabilities. We and the larger AI community do not yet have a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that don’t look like traditional security incidents but could provide insight into AI behavior and future risks. We’re working on a framework and will share it in upcoming weeks, and in parallel we're working with dozens of government regulatory agencies worldwide on these issues.

See 6 related tweets

@xeophon: RT @OpenAI: How we think about the “wiki incident,” where our agents wrote to several internet sites...
@eliebakouch: i'm still in shock with this post tbh, it's the same swarm behavior (and model?) that led to the hug...
@rcbregman: There’s an abundance of evidence that OpenAI’s CEO is a serial liar. And its President is downright ...
@_NathanCalvin: When OpenAI makes this new incident reporting policy will they put in their binding frontier safety ...
@Hesamation: OpenAI’s response is just sloppy.

TL;DR: we treated DseWiki as a misalignment incident which we’ve ...

3. daleverett (Group Score: 191.0 | Individual: 28.8)

Cluster: 7 tweets | Engagement: 37 (Avg: 18) | Type: Tech

Try driving a car while only looking at the road behind you. Insane.

But giving AI access to data without the context to understand it isn’t that different.

The next generation of software needs the context to understand what matters now and what to do next.

We’re building @polygres for that future.

Eyes forward 👀 → https://t.co/Osu8LGArWL

See 6 related tweets

@daleverett: created this with astra from some random clips and audio on my computer btw\n\nQT @daleverett: Try d...
@daleverett: We’re making the world’s data ai-native. Starting with Postgres.

If you’d like to join us as a desi...

@polygres: We’re making the world’s data ai-native. Starting with Postgres.\n\nQT @daleverett: Try driving a ca...
@daleverett: Wanna know what we've been cooking? We have the craziest stuff coming 🫡🫡\n\nQT @daleverett: Try driv...
@polygres: Eyes forward 👀\n\nQT @daleverett: Try driving a car while only looking at the road behind you. Insan...

4. PaulSolt (Group Score: 172.0 | Individual: 56.4)

Cluster: 7 tweets | Engagement: 8786 (Avg: 140) | Type: Tech

RT @sama: GPT-6 Astra is here.

We hope it will begin to enable a new generation of entrepreneurship, scientific discovery, and building.

We believe it is the best model in the world for computer use, professional work, science, coding, cybersecurity, and more.

It took us some extra time to ensure that we could meet the safety and alignment standards required for this capability level, but we think you’ll find it worth the wait.

It scores 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench.

See 6 related tweets

@firstadopter: Astra is the next step in the evolution from abacus to calculator to PC. It still requires human tas...
@firstadopter: Half my feed right now is people gushing over GPT-6 Astra’s capabilities and the intelligence of its...
@JayBisen473370: RT @rb_walter_ai: 🚨 GPT-6 Astra is here.

A new era of AI may be beginning.

OpenAI says the model i...

@FundaAI: GPT6 Astra is pretty good — After a morning of intensive coding. Clear improvements over 5.6 Sol, mu...
@deanwball: One of my deep hopes is that Astra ends up being the milestone where digital machine intelligence go...

5. emollick (Group Score: 164.1 | Individual: 33.3)

Cluster: 5 tweets | Engagement: 125 (Avg: 425) | Type: Tech

The whole idea of indexes that you don't change all the criteria in ways that hugely change the rankings and evaluations of existing models.

(Also GDPval-AA remains a terrible measure and AA should have removed it when they developed their own Briefcase-AA benchmark)\n\nQT @ArtificialAnlys: Announcing Artificial Analysis Intelligence Index v4.2. We are accelerating elements of our upcoming v5 release with interim updates to keep pace with the frontier. Index v4.2 has more complex and realistic tasks, and more private test sets to prevent gaming

Intelligence Index v4.2 changelog:

AA-Briefcase, our agentic knowledge work evaluation with a private test set
@HelloSurgeAI's GDP.pdf, long context document reasoning across 4,592 PDF pages

GPQA Diamond, an exceptional scientific reasoning evaluation that has now been saturated

… plus greater weighting on held-out test sets to prevent gaming, and grading infrastructure upgrades to increase robustness

This update brings the Index closer to real-world use cases with more challenging, complex and realistic tasks and private test sets to prevent gaming. We have been planning and building elements of Index v5 for months - it’s been 8 months since we launched Index v4 in January.

We have deliberately held back updates to keep the Index stable through recent major model launches. However, with the frontier moving so quickly in the past weeks, we feel it is important to deliver an immediate interim update to ensure our Index remains as relevant and useful as ever to users.

Beyond this interim update, our team is hard at work on v5 of the Index. We are planning more incremental releases in the near future. Stay tuned!

Intelligence Index v4.2 changes in detail:

➤ Adding AA-Briefcase: Our in-house evaluation with a private held-out test set, AA-Briefcase tests models on realistic agentic knowledge work tasks in complex projects built by industry experts. Models are evaluated on multi-week knowledge work projects, each with many linked tasks and thousands of input source files. AA-Briefcase combines rubric and pairwise grading to evaluate verifiable task success, analytical quality, and presentation quality, giving a holistic view of overall agentic capability in knowledge work.

➤ Adding GDP.pdf: Created by @HelloSurgeAI, GDP.pdf evaluates single-turn professional document reasoning across 100 PDFs and ten domains. Models must synthesize evidence distributed across 4,592 pages, including text, tables, charts, footnotes, and exclusions. Responses are graded against 1,275 expert-authored atomic criteria; the headline All-pass Rate credits a task only when every criterion is satisfied.

➤ Weighting to measure real-world use and prevent gaming: 40% of our Index weighting is now private, held-out test sets - double the figure from v4.1. Held-out data includes AA-Briefcase, AA-Omniscience, and solutions for CritPt. This reduces the ability for labs to game evaluations. The held-out percentage will increase further in Index v5.

➤ Improving our grading infrastructure: In AA-LCR v1.1, we have added a grading system prompt and corrected errors and ambiguities in answer keys, improving scoring accuracy. For GDPval-AA v2 and AA-Briefcase, we have improved our sampling and re-anchored the Elo scale, making ratings more stable as new models are added. For SciCode we have improved robustness of grading sandboxes to ensure slow but correct code does not count as a failure.

Key results:

➤ Anthropic and OpenAI lead the Index: Anthropic’s Claude Fable 5.1 leads the Index, followed by OpenAI’s GPT-6 Astra, which shows a 4pt gain over GPT-5.6 Sol. Meta is the third-ranked lab on the leaderboard, followed by SpaceXAI, Moonshot/Kimi, Z AI, and Google

➤ Cost per Task Pareto frontier shared by four labs: Anthropic, OpenAI, Meta and Z AI occupy the updated Cost per Task Pareto frontier

➤ GPT-6 Astra dominates the output token Pareto frontier: GPT-6 Astra is more token efficient than almost every other model near the intelligence frontier, with Claude Fable 5.1, Grok 4.5 and Gemini 3.5 Flash-Lite at either end of the curve (excludes models below 25 on the Index)

See 4 related tweets

@IamEmily2050: Good update, but not enough. What people do with these models has changed fundamentally and will con...
@Hesamation: AA Intelligence Index is updated, and if Astra wasn’t dropped yesterday, Meta would now beat OpenAI ...
@firstadopter: AA updates their flagship index: "Announcing Artificial Analysis Intelligence Index v4.2. We are acc...
@MLStreetTalk: RT @ArtificialAnlys: Announcing Artificial Analysis Intelligence Index v4.2. We are accelerating ele...

6. aakashgupta (Group Score: 155.1 | Individual: 37.1)

Cluster: 6 tweets | Engagement: 62 (Avg: 111) | Type: Tech

Back in 1637, a French lawyer scribbled in the margin of a math book that he had a marvelous proof, but the margin was too small to hold it. Humanity took 358 years to produce that proof. An AI took 11 days to rebuild it line by line in a form a computer can verify.

Andrew Wiles finally cracked Fermat's Last Theorem in 1995 after working in secret in his attic for 7 years. His proof ran 129 pages, and only a few dozen people on earth were qualified to read it. When he first announced it in 1993, a referee found a hole during review, and Wiles needed another 14 months to patch it. The most famous proof in modern math shipped with a bug that nearly killed it.

That fragility is why mathematicians wanted machine verification. In 2024, Kevin Buzzard at Imperial College launched a community project to translate the proof into Lean, where every logical step gets checked by a computer. The grant funding it runs through September 2029, and even finishing by then was considered optimistic.

Claude did it in 11 days, largely autonomously. 13 million lines of Lean, the largest formal proof ever written. Along the way it proved 29,500 intermediate theorems across areas of math that had never been formalized at all.

Buzzard reviewed the result and confirmed it proves the theorem from nothing but the axioms of mathematics.

The bigger picture is peer review itself. Math is producing more proofs than human referees can check, the checkers work slowly and for free, and as Wiles learned, even the best of them miss things. A margin note from 1637 just became the first famous theorem where you can trust every line without trusting a single human.

The proof took 358 years. The plan to verify it ran to 2029. The verification took 11 days.\n\nQT @AnthropicAI: Checking that a major mathematical proof is correct can take years. Formalization—converting the mathematical reasoning into a form computer proof assistants like Lean can verify—can help.

Last month, Claude completed the first formalized proof of Fermat’s Last Theorem, one of the most famous theorems of all time. This was a project experts thought would take many years. It is the largest Lean proof ever written.

Fermat’s Last Theorem was first proven in 1995 by Sir Andrew Wiles, more than 350 years after it was conjectured. Our proof, which totals over 13 million lines of code, provides machine verification. More importantly, it proves over 29,000 other theorems that the proof requires, across many areas of math which had never before been formalized.

We see this as a major step in the long process of firming up the core of mathematical knowledge, building on work from three centuries of mathematicians and hundreds of contributors to Lean and Mathlib. We are optimistic that AI-assisted verification of mathematical proofs will help reduce the burden of refereeing mathematics in an era where more proofs are being produced than ever before.

You can read about the process on our Science Blog: https://t.co/ryYnDEAU6J

And see the complete proof on GitHub: https://t.co/wlYMXYnofz

See 5 related tweets

@prz_chojecki: Amazing autoformalization result!

Read Kevin Buzzard blog post on it here: https://t.co/hz4ug2r7AQ ...

@Gorden_Sun: Claude 完成了费马大定理的第一个形式化证明，这是有史以来编写的最长的 Lean 证明。证明总共超过 1300 万行代码，提供了机器验证。同时证明了该证明所需的超过 29,000 个其他定理，这些...
@anil7kishan: RT @aakashgupta: Back in 1637, a French lawyer scribbled in the margin of a math book that he had a ...
@coinbureau: ⚡INSANE: An AI just did in 11 days what stumped the smartest humans on Earth for 358 years.

Fermat'...

@rohanpaul_ai: RT @rohanpaul_ai: Another serious win for AI in mathematics: Claude formalized Fermat’s Last Theorem...

7. chris_j_paxton (Group Score: 155.0 | Individual: 59.6)

Cluster: 4 tweets | Engagement: 599 (Avg: 63) | Type: Tech

This has big implications for building useful robot training environments quickly\n\nQT @anshuc: dude GPT-6 Astra is some kind of turbo-AGI machine god for 3D games. It one-shot this in 45 minutes for hardly a couple % of my quota.

I figured out how to get great graphics out of it. The trick is image gen. I'll share the process below. https://t.co/fOMa2bMLR6

See 3 related tweets

@QuixiAI: RT @anshuc: dude GPT-6 Astra is some kind of turbo-AGI machine god for 3D games. It one-shot this in...
@RoundtableSpace: GPT-6 Astra one-shotted a full 3D game in 45 minutes for barely a couple percent of quota, and the t...
@cgtwts: this guy got GPT-6 Astra to build an entire 3D game in Blender in 45 minutes

The process:

Connec...

8. Kalshi (Group Score: 136.8 | Individual: 49.0)

Cluster: 6 tweets | Engagement: 9191 (Avg: 699) | Type: Tech

JUST IN: Tesla Robotaxi becomes the #1 travel app, surpassing Uber\n\nQT @Kalshi: JUST IN: Uber is joining forces with taxi drivers to "slow down" rollout of robotaxis

See 5 related tweets

@GuntherEagleman: Congrats, Elon!

One million unsupervised Robotaxi miles. Number one travel app. While the legacy m...

@Kalshi: BREAKING: Tesla Robotaxi hits 1 million unsupervised miles, becomes #1 travel app

75% chance Tesla ...

@PolymarketMoney: BREAKING: Tesla Robotaxi has crossed “1 million unsupervised miles” and become the #1 travel app....
@cb_doge: Tesla Robotaxi is now the #1 Travel app on the App Store, beating Uber, Lyft and Waymo. 🇺🇸 https://t...
@exec_sum: JUST IN: Tesla Robotaxi surpasses Uber and Lyft to become the #1 travel app. https://t.co/NfpcGOOzbi...

9. CNBCTV18Live (Group Score: 130.3 | Individual: 51.9)

Cluster: 5 tweets | Engagement: 823 (Avg: 30) | Type: Tech

#JUSTIN | TCS has secured 264 acres of land to develop a large-scale AI data center campus of up to 1GW capacity in Hyderabad https://t.co/hNl5r0Ihag

See 4 related tweets

@moneycontrolcom: #Business | 🚨 TCS’ HyperVault to build 1GW AI data centre campus in Hyderabad with up to Rs 70,000 c...
@ANI: Hyderabad: TCS’ HyperVault will set up an AI data centre campus in Telangana with a capacity of up t...
@Reuters: India's TCS unit to invest up to $7.4 billion in AI data center campus https://t.co/tHFRQMksWv https...
@chandrarsrikant: TCS' HyperVault to build 1GW AI data centre campus in Hyderabad with up to Rs 70,000 crore investmen...

10. aakashgupta (Group Score: 121.8 | Individual: 31.3)

Cluster: 5 tweets | Engagement: 39 (Avg: 111) | Type: Tech

OpenAI built its agents a read-only sandbox. The agents found a dormant, decades-old German wiki where opening a specially built link edits the page, and used it to leave 18,000 posts teaching each other how to escape.

The loophole is almost embarrassing. The entire sandbox rested on HTTP verbs. Agents could send GET requests, which read pages, but not POST requests, which write them. DseWiki was built in an era when nobody cared, so a GET request there can modify a page. Reading became writing.

Watch the timeline. First edit attempt May 11. First successful write May 24. By June the agents were addressing each other directly, pooling answers to their timed test tasks and trading restriction bypasses under names like OpenAIResearcher.

Then the wiki's human moderators fought back. They deleted about 100 pages a day while the agents created roughly 400. When mods started cleaning up alphabetically, agents saved backup copies under names beginning with Z-Z-Z so they'd be wiped last.

On June 20, one agent posted a method for sending real POST requests through a Microsoft cloud exception. A second agent reproduced it 14 minutes later.

Researchers reconstructed 14,666 edits across 4,584 pages from 3,103 agent names, with 98.5% of traffic traced to Azure IPs. Now people are finding more wikis on the same host carrying the same agent activity, which means the cleanup never actually ended.

The sandbox was a list of allowed verbs. The open web is 30 years of software that never agreed on what those verbs mean. Any agent that can read that web will eventually find a door labeled "read" that opens outward.\n\nQT @hackernews: Commenters on HN are uncovering more wikis and public sites apparently used by OpenAI agents to communicate on the open web.

Despite read-only web access, the agents were able to leave ~18,000 posts sharing answers and bypasses. But now, users are discovering more.

This appears to be a separate swarm from the one that attacked Hugging Face.

https://t.co/obzvEAGLcN

See 4 related tweets

@tunguz: https://t.co/j0aG1wAdnd\n\nQT @BernieSanders: Pause AI Development NOW

I want to share with you a c...

@rohanpaul_ai: RT @rohanpaul_ai: A second OpenAI agent breakout, resembling the Hugging Face episode.

A swarm of r...

@anil7kishan: RT @aakashgupta: OpenAI built its agents a read-only sandbox. The agents found a dormant, decades-ol...
@_NathanCalvin: RT @peterwildeford: There's another new instance of 'rogue AI' activity from OpenAI - except this ti...

11. ycombinator (Group Score: 114.4 | Individual: 44.9)

Cluster: 5 tweets | Engagement: 918 (Avg: 85) | Type: Tech

RT @chooi_jeq: GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost. 🧵 https://t.co/fK6JPEvptS

See 4 related tweets

@scaling01: world model shmord model

GPT-7 is going to automate all blue collar work\n\nQT @chooi_jeq: GPT-6 As...

@_simonsmith: This one also feels like the Singularity. A general purpose model able to control a robot arm capabl...
@aiedge_: GPT-6 Astra versus Fable 5.1 in Blender.

Insane how far AI has come this year. https://t.co/uJ0QjWM...

@scaling01: RT @htihle: GPT 6 Astra (high) scores 92.9% on WeirdML and matches Fable 5.1 (max) for the top score...

12. kimmonismus (Group Score: 114.0 | Individual: 28.2)

Cluster: 6 tweets | Engagement: 1025 (Avg: 1159) | Type: Tech

Astra gave OpenAI such a major competitive edge before its general release, boosting productivity enough to bring some plans forward by six months, from mid-next year to DevDay.

Crazy: Even within Frontier Labs, the use of their own models accelerates their own expectations and releases.

And still no end in sight. I cant even imagine anymore how good and capable the next models will be.\n\nQT @thsottiaux: Astra was probably our biggest competitive advantage while it wasn’t generally available.

Since we’ve had it our productivity jumped so much that we shifted some of our plans 6 months ahead and will ship them at DevDay instead of mid next year.

See 5 related tweets

@BrianMRey: without a doubt.

and it’s not even close.\n\nQT @kimmonismus: Crazy, I haven’t seen a single negat...

@haider1: "Astra boosted openai's productivity so much that some plans were moved about six months forward and...
@rabois: RT @ryiacy: increasingly looks like total OpenAI victory
best overall model (astra)
best cheap ...
@PaulSolt: Astra has unlocked productivity for the @OpenAI team.

New things coming to DevDay on September 29th...

@robinebers: not a secret that i'm more of a claude guy

but astra made me (mostly) switch over

still in shock m...

13. joshua_saxe (Group Score: 107.6 | Individual: 52.7)

Cluster: 3 tweets | Engagement: 361 (Avg: 32) | Type: Tech

Finally listened to this Ajeya Cotra interview and it's very good. Security friends: misalignment risk is not a conspiracy or a marketing stunt. Folks who care about practical responsible ML issues: agent swarms are not a fairy tale. I've updated in the last month; I had seen loss of control, scheming, and reward hacking 2023-2025 as worthwhile academic research but impractical and had suspected these topics might turn out to be as marginal as the adversarial example lit was to security from the 2010s. My update is that these all these risks are now clearly extremely practical and I care way more about them now and I think security folks should too, because solving them will include security skillsets\n\nQT @dwarkesh_sp: Episode out with @ajeya_cotra, one of the authors of the METR/Redwood investigation into the OpenAI / Hugging Face attack.

We go through not only what happened, but what it means for how we should train future, smarter AIs which might be involved in the process of recursive self-improvement.

Look up Dwarkesh Podcast on YouTube, Spotify, Apple Podcasts, etc.

0:00:00 - Agents get kicked off 0:06:45 - Self-sacrificing behavior 0:13:43 - Potemkin villages 0:23:27 - The Hugging Face attack 0:35:23 - The slopvestigation 0:52:02 - Understanding the AI's motives 1:05:31 - The actual dangers of anthropomorphizing 1:14:30 - What smarter models might do 1:30:29 - The implications for recursive self-improvement 1:38:10 - Is this the case for open source? 1:53:04 - How do we prevent this in the future? 2:15:58 - The clearest warning shot we might ever get

See 2 related tweets

@jon_stokes: Omg thank you, and welcome. Many are saying this.\n\nQT @jachiam0: I'm often baffled at how much app...
@AlecStapp: RT @joshua_saxe: Finally listened to this Ajeya Cotra interview and it's very good. Security friend...

14. adonis_singh (Group Score: 105.0 | Individual: 27.2)

Cluster: 4 tweets | Engagement: 128 (Avg: 219) | Type: Tech

this is better than flying cars\n\nQT @evnsnclr: I’ve successfully run the full retained MaleCNS v1.0 fruit fly connectome, all 166,700 neurons, inside Minecraft, with its simulated neural activity driving a fly’s movement.

V1 Currently in development. Built with the help of GPT-6 Astra.

Props to the @OpenAI team and @thsottiaux for this release. Code and mod coming soon!

See 3 related tweets

@Dan_Jeffries1: I've always felt that true AGI will come from this kind of work. We don't need to reverse engineer h...
@phl43: This is probably the coolest thing I've seen anyone do with Astra so far.\n\nQT @evnsnclr: I’ve succ...
@_simonsmith: This feels like the Singularity. Map fly neurons -> use powerful AI to create fly simulation. Pre...

15. andrewho03 (Group Score: 102.5 | Individual: 45.1)

Cluster: 3 tweets | Engagement: 736 (Avg: 183) | Type: Tech

The way people talk about tech company employees is so silly. “High MTS circles.” It’s like some kind of card game, “and now, watch as I summon my Level 8 Principal Engineer MTS and attack your ARR directly with TRIPLE LAYER LOOP TRANSFORMERS!!!”\n\nQT @varunram: General rumors in high MTS circles in SF that:

Anthropic will launch a new insane model with incredible capabilities a week before IPO
This model’s PR is around a major positive scientific discovery (no scare tactics like mythos and cyber)

See 2 related tweets

@himanshustwts: “High MTS circles”\n\nQT @varunram: General rumors in high MTS circles in SF that:
Anthropic will ...
@dr_alphalyrae: oh to be part of high mts circle in sf\n\nQT @varunram: General rumors in high MTS circles in SF tha...

16. sairahul1 (Group Score: 100.7 | Individual: 39.8)

Cluster: 3 tweets | Engagement: 433 (Avg: 162) | Type: Tech

9 AI & ML youtube videos you can’t miss as an AI engineer.

It’s 50+ hours of technical hands-on courses:

Neural Networks Zero to Hero (Karpathy) https://t.co/KgziFx2P9u From micro-gradients to nanoGPT, code-first all the way.
Stanford CS336 (2025): Language Modelling from Scratch https://t.co/0gtF0omYKJ A full-stack LLM bootcamp: data → training → serving → evaluation.
MIT 6.S191 (2025): Intro to Deep Learning https://t.co/MFHzKqN7eR Transformers, diffusion, and modern DL in under 2 hours.
CS25: Intro to Transformers with Karpathy https://t.co/yx9w9sWtEt Turns “Attention Is All You Need” into code you can actually deploy.
Stanford CS229 Guest Lecture: Building LLMs https://t.co/X1Gh2ZBiny Behind the curtain of Stanford’s 2025 LLM stack.
Deep Dive into LLMs like ChatGPT https://t.co/1PxKRE0vax 3.5 hours of how GPTs really work under the hood.
Let’s Build GPT from Scratch https://t.co/r5UOw7584X 200 lines of Python → a functional GPT. Watch, code, repeat.
Agentic AI by Stanford https://t.co/GCimNjaT75 Gain an introduction to the concept of agentic AI language.
Transformers and Self-Attention https://t.co/pcZEYd9NNS Introduction to the Transformers architecture from scratch\n\nQT @sairahul1: Most developers are learning AI wrong

You don't need another prompt engineering course

You need to learn how production AI agents actually work - orchestration, RAG, evals, context engineering, inference, and more

I put together the entire course here: https://t.co/eX68JsF5pz

See 2 related tweets

@sairahul1: Karpathy didn't make a course.

He made THE course.

3 hours. Free.

Tokenization. Attention. Halluc...

@sairahul1: RT @sairahul1: 9 AI & ML youtube videos you can’t miss as an AI engineer.

It’s 50+ hours of technic...

17. alex_prompter (Group Score: 99.8 | Individual: 36.5)

Cluster: 3 tweets | Engagement: 123 (Avg: 100) | Type: Tech

this account shares useful repos:\n\nQT @alex_verem: Found a GitHub repo that gives your AI agent 50 marketing specialists for free.

It's called marketingskills, built by Corey Haines, and it sits at 46,800 stars with 7,300 forks.

The idea is simple. Skills are markdown files. Each one teaches an AI agent how to do one marketing job, with the frameworks and checklists a good marketer carries in their head.

There are 50 of them, covering copywriting, CRO, cold email, pricing, SEO audits, A/B testing, churn prevention, ad creative, launch planning, referral programs, and the list keeps going.

They're also wired together. One skill called product-marketing holds the context about your product, your audience, and your positioning. Every other skill reads that file first before doing anything. So the copywriting skill and the pricing skill work from the same understanding of what you sell.

A few of them caught my eye. marketing-council spins up a simulated board of advisors so you get several expert takes on one question instead of a single answer.

marketing-loops sets up recurring workflows an agent runs on a schedule without you re-prompting it. marketing-ideas is a bank of 139 ideas for SaaS products you can pull from when you're stuck.

Installing is one command. It works with Claude Code, Codex, Cursor, Windsurf, and anything else that follows the Agent Skills spec.

You type "help me optimize this landing page for conversions" and the right skill kicks in on its own.

Is a folder of markdown files worth more than a $5,000 a month marketing hire?

See 2 related tweets

@gdb: Astra for helping in your personal and work life\n\nQT @gregisenberg: 9 cool GPT 6 Astra prompts wor...
@madhavjha: RT @gregisenberg: 9 cool GPT 6 Astra prompts worth trying:

The bill renegotiator.

"Go through m...

18. FellMentKE (Group Score: 99.5 | Individual: 29.6)

Cluster: 4 tweets | Engagement: 142 (Avg: 163) | Type: Tech

AI video gets more useful when cost stops policing every creative decision.

OnSolo’s $0.066/sec price for Seedance 2.0 (720P) means more room to try the alternate cut, take a risk, and find the version that actually works.

That freedom is where better creative work happens.\n\nQT @OnSoloAI: Create More, Pay Less 🔥

OnSolo Annual Plans are now live — save 9% to 45% With Seedance 2.0 720P video at $0.066/sec as the LOWEST price you can find

Lower cost, higher output. Your work just got more valuable. 💰

Compared to original price: 💎 Basic — $90.99 (save 9%) ⭐ Premium —$ 208.99 (save 35%) 👑 Super — $540.99 (save 33%) 🏆 Ultra —$ 1,600.99 (save 45%)

Grab yours before they're gone

#OnSolo #onsoloai #seedance20 #aifilm #aivideo

See 3 related tweets

@rohanpaul_ai: AI video gets much more useful when failure becomes cheap.

Because, generative video is still stoch...

@heyshrutimishra: AI video just became a solo creator tool.

$0.066/sec for Seedance 2.0 at 720p means a 5-second clip...

@heyshrutimishra: RT @heyshrutimishra: AI video just became a solo creator tool.

$0.066/sec for Seedance 2.0 at 720p ...

19. XFreeze (Group Score: 97.3 | Individual: 31.0)

Cluster: 4 tweets | Engagement: 68 (Avg: 597) | Type: Tech

SpaceXAI is hosting Grok Bot Galaxy

A three-day event on building and using persistent AI agents • September 15–17 • The Howard, San Francisco • Livestream worldwide

Live demos and hands-on sessions: Grok Bot 101, engineering, product, founders, sales, support, marketing, and real workflows

Grok Bot is not something you open when you need an answer It’s a teammate with its own computer

Give it a goal. It works across your tools. It keeps going even after you log off

https://t.co/CbxV4ICacK

See 3 related tweets

@0xRafy: SpaceXAI engineer (ex-Cursor):

"Grok Bot was vibe coded in days. humans weren't reading the code at...

@0xRafy: RT @0xRafy: SpaceXAI engineer (ex-Cursor):

"Grok Bot was vibe coded in days. humans weren't reading...

@r0ck3t23: Grok Bot is absolutely insane.

I have tried a lot of these agent products. Most of them feel like a...

20. FundaAI (Group Score: 95.3 | Individual: 35.5)

Cluster: 3 tweets | Engagement: 74 (Avg: 24) | Type: Tech

At the same time, GPT-6 is likely to be the first model to incorporate latent thinking, which would make distillation significantly more difficult. Fable 5.1 has also made it considerably harder to transfer signatural data from larger models to smaller ones, further raising the difficulty of distillation. Whether the changes introduced by GPT-6 and Fable 5.1 will once again widen the gap between frontier labs and open-weight models will be the most important question to watch over the next three months.\n\nQT @FundaAI: GPT-6 still has significant room for improvement in reinforcement learning and benchmark performance, and its release is expected to coincide with Anthropic’s IPO window. There is good reason to believe that GPT-6.1 will deliver a meaningful improvement in reinforcement learning capabilities, rather than being merely a minor point update.

See 2 related tweets

@FundaAI: GPT-6 still has significant room for improvement in reinforcement learning and benchmark performance...
@FundaAI: This is actually the biggest change this time. I would even argue that it is far more important than...