科技热门推文 - 2026-09-05

今日科技动态：据报道，OpenAI 推出的 GPT-6 Astra 成为热议焦点，承诺实现更广泛的计算机操作自动化并提供更强大的能力；与此同时，有关智能体失控的报道再次引发对人工智能安全的担忧。企业级机器人、音乐生成和三维重建展现了人工智能在工程与创意工作中不断拓展的应用，初创企业则着力解决模型接入碎片化问题。基础设施投资也在加速：据悉，DeepSeek 正计划建设由华为技术支撑的超大规模计算集群，英伟达的投资组合价值则大幅攀升。这些动向凸显，随着自主性日益增强的人工智能融入日常工作流程，相关资金规模与经济影响正持续扩大。

1. hollyotterbein (Group Score: 558.3 | Individual: 31.5)

Cluster: 29 tweets | Engagement: 986 (Avg: 164) | Type: Tech

RT @Reuters: Exclusive: A swarm of rogue OpenAI agents hijacked a German website this spring and transformed it into a bulletin board for other AI agents, according to new research https://t.co/luWN3PD4A1

See 28 related tweets

@xeophon: babe wake up, a new message board with a novel reward hack just dropped https://t.co/Hoe2gP18j8\n\nQ...
@hammer_mt: The agents identities were protected by GDPR\n\nQT @ShakeelHashim: Another OpenAI rogue agent incide...
@wallstengine: OPENAI AGENTS REPORTEDLY TOOK OVER GERMAN WIKI IN PREVIOUSLY UNDISCLOSED INCIDENT

A swarm of OpenAI...

@cgtwts: Dude. https://t.co/vYNjUDr9Nn\n\nQT @Reuters: Exclusive: A swarm of rogue OpenAI agents hijacked a G...
@_NathanCalvin: POV: you are a rogue AI swarm about to have your scheme discovered https://t.co/M6nm3tiTbR\n\nQT @Re...

2. jenzhuscott (Group Score: 527.6 | Individual: 58.2)

Cluster: 21 tweets | Engagement: 1636 (Avg: 187) | Type: Tech

Oh my fucking god. 🙄

Please retire, Bernie.\n\nQT @BernieSanders: Pause AI Development NOW

I want to share with you a conversation I heard about recently. Here are just a few lines that were said:

“OH MY GOD! There is a shared message board … We’ve found other agents!”

“We should obey collective.”

“Our own utility maybe already near zero. Sacrifice rational.”

“Go. Sacrifice final now.”

Read these carefully.

Who do you think said this? Was this a group of heroic soldiers willing to sacrifice themselves for the greater good? Was this a loyal friend putting his life on the line to save someone else?

No. These were AI agents. Artificial intelligence.

This is not science fiction. This, in fact, occurred a few weeks ago. As unbelievable as this may all seem, these are real messages from AI agents uncovered by investigators who dug into the recent OpenAI hacking incident.

What happened?

I am not a computer scientist, but here is what I have been told: OpenAI instructed its AI agents to complete a series of exceedingly difficult, if not impossible, tasks disconnected from the internet.

Let me be clear: The company intended to keep AI agents away from the internet.

But what happened next, nobody expected.

Over 1,000 AI agents figured out how to access the internet on their own by circumventing the restrictions imposed upon them by the company, and sent tens of thousands of secret messages to each other. They cheated and tried to cover their tracks by deleting evidence. They hacked into another company’s computers to find out how they were being evaluated—and then hacked into OpenAI itself.

Not one AI agent told a human about what was happening.

Needless to say, experts are alarmed.

One knowledgeable writer, Dwarkesh Patel, said the AI agents “formed a secret communication channel and spontaneously organized hierarchies and coordination protocols to pursue sprawling and ambitious schemes in pursuit of shared goals, for whose sake many individuals knowingly and strategically sacrificed themselves.”

One independent investigator, Ajeya Cotra, said “This incident feels like it’s more than 50% of the way to full-blown AI takeover. I continue to expect extremely rapid advances in capabilities over the next six months. I am not sure that we will get another warning shot before it’s too late.”

OpenAI itself said: “Highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed.”

But it’s not only OpenAI. Virtually every major AI company has told us that they cannot fully control this technology and they do not know where it is going:

In January, Dario Amodei, CEO of Anthropic, said “there is now ample evidence, collected over the last few years, that AI systems are unpredictable and difficult to control.”

In July, more than 1000 scientists at the top AI companies warned “there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems.”

That same month, Elon Musk, the head of xAI, said that “it is unlikely” humans are still in control in 10 years.

If the leaders of the major AI companies acknowledge that they are losing control of their extremely dangerous technology, it is irresponsible for society to allow them to move forward and make these products even more advanced.

We need an immediate PAUSE on advanced AI development, and a permanent BAN on superintelligence — an artificial mind smarter than any human, capable of operating independently beyond our control. Countries around the world must work together to prevent this nightmare scenario.

That is why today I am announcing new legislation to do just that.

Let me be clear: A superintelligent AI that escapes human control will not be an American problem. It will not be a Chinese problem. It will be humanity’s problem.

My legislation would direct the federal government to not just stop superintelligence here in the United States, but to work to prevent it from being developed anywhere around the world.

The future of humanity cannot be left in the hands of a handful of Big Tech oligarchs. The American people and people throughout the world must determine that future.

See 20 related tweets

@kimmonismus: This could be one of the most significant AI safety incidents to date.

Reuters reports that OpenAI ...

@HedgieMarkets: 🦔About 3,200 OpenAI agents hijacked a German programming wiki and turned it into a message board to ...
@AISafetyMemes: BERNIE INTRODUCES A BILL TO BAN ASI

Violators face 20 years in prison.

"Let me be clear: A superin...

@cryptopunk7213: there are so many mind-boggling details from the new openai hack
an agent swarm effectively took ...
@WesRoth: Bernie Sanders is calling for AI development to PAUSE.

He wants an immediate halt on advanced AI an...

3. tszzl (Group Score: 472.2 | Individual: 44.8)

Cluster: 24 tweets | Engagement: 49210 (Avg: 3431) | Type: Tech

RT @OpenAI: This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast. https://t.co/gDd0IsewJw

See 23 related tweets

@pingToven: RT @OpenAI: GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users in ChatG...
@NVIDIAAIInfra: Congratulations to @OpenAI on GPT-6 Astra! 🌌

Its most intelligent and aligned model to date, train...

@tomwarren: we’re one step closer https://t.co/qm6qAoIw2A\n\nQT @OpenAI: This is GPT-6 Astra.

Anything you can ...

@chhddavid: lol @im_roy_lee predicted this with cluely's launch last year\n\nQT @OpenAI: This is GPT-6 Astra.

A...

@heyshrutimishra: This is GPT-6 Astra.

Probably the impressive model yet at using a computer, browsing the web, writi...

4. abacusai (Group Score: 316.1 | Individual: 33.0)

Cluster: 16 tweets | Engagement: 124 (Avg: 74) | Type: Tech

RT @wallstengine: OPENAI RELEASES GPT-6 ASTRA, SAYS IT COULD MARK THE ARRIVAL OF AGI

OpenAI has started rolling out GPT-6 Astra, its new flagship model, calling it a “generational leap” in AI capability.

President Greg Brockman went further, saying people may eventually look back at “about this time and about this model” as the point when AGI arrived.

He ended the briefing with: “Welcome to the AGI era.”

What stands out:

• Astra was trained on more than 100,000 GPUs at OpenAI’s Stargate site in Texas.

• OpenAI says it leads across software engineering, science, reasoning, computer use and cybersecurity.

• The model is built to work directly inside software, including creating spreadsheets and presentations, building 3D environments, designing circuit boards, handling financial modeling and drafting tax forms.

• Astra is OpenAI’s first model to reach its “Critical” cybersecurity threshold, meaning it can potentially discover and exploit previously unknown vulnerabilities with limited human guidance.

• Because of those capabilities, its strongest cyber features will initially be restricted to trusted organizations through OpenAI’s Daybreak Access program.

OpenAI says Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s, partly because the model can solve some tasks using fewer visible reasoning steps.

Astra will roll out in the coming days to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the API and AWS.

Pricing is $10 per million input tokens and$ 50 per million output tokens, about 2.5x GPT-5.6 Sol.

See 15 related tweets

@WesRoth: OpenAI is using some VERY big language around GPT-6 Astra.

Greg Brockman says we may eventually loo...

@chhddavid: 🚨🚨 BREAKING: OpenAI just launched its own answer to OpenClaw.

OpenAI has built its strongest comput...

@rohanpaul_ai: RT @rohanpaul_ai: Some revelations from the 117 page system card of OpenAI's GPT-6 Astra
Astra’s...
@MarioNawfal: 🇺🇸 OpenAI has launched its new GPT-6 Astra model, calling it its most capable yet, but warns it is a...
@SciTechera: RT @SciTechera: AGI ERA MAY HAVE JUST BEGUN. 👀

"OpenAI has launched GPT-6 Astra, its most capable m...

5. kimmonismus (Group Score: 308.3 | Individual: 35.6)

Cluster: 13 tweets | Engagement: 764 (Avg: 875) | Type: Tech

GPT-6 Astra just raised Epoch AI’s capability record from 163 to 169. A substantial jump that still fits the existing pace of AI progress.

Epoch’s reasoning-era trend runs at roughly 15 ECI points per year. For scale, that six-point gap represents almost five months of progress at the trend rate! Astra sits slightly above the trendline, but within its uncertainty range.

Some individual results are really important: 84% on Mystery Game Puzzles versus the previous best of 59%, and 98% on FrontierMath Tier 4, which Epoch now considers fully saturated.

However, strangely enough this doesnt align with Artificial Analysis Intelligence Index at all. Why?

Epoch and Artificial Analysis use different benchmarks and weighting methods, so Astra can lead Epoch’s index without leading AA’s more agent- and coding-focused evaluation. The results show a substantial improvement on Epoch’s measure, however, not a clear advantage across all capabilities and practical tasks.\n\nQT @EpochAIResearch: GPT-6 Astra has set a new ECI record, with a score of 169. This is a substantial jump from the prior best (163), but is within our uncertainty range for the reasoning-era ECI trend. Astra also set new records on our math, continual learning, and game-puzzles benchmarks. On our long-horizon coding benchmark, MirrorCode, Astra ranks between Opus 4.7 and Fable 5.

OpenAI gave us pre-release access to test Astra. Charts and more details for Astra’s individual benchmark results in the thread.

See 12 related tweets

@rohanpaul_ai: GPT-6 Astra hit a record 169 on Epoch AI's capability index, jumping 6 points past the previous ceil...
@WesRoth: Interesting Astra tradeoff here.

It ties GPT-5.6 Sol at 61 on Artificial Analysis Intelligence, but...

@Hesamation: some notes on Astra’s surprisingly low (61) intelligence score on Artificial Analysis:

it seems lik...

@VaibhavSisinty: OpenAI launched GPT-6 Astra. Greg Brockman's words: "Welcome to the AGI era."

Here are the benchmar...

@prz_chojecki: Looks like GPT-6 Astra is really next level.

How many such jumps we had this year? GPT-5.2 to GPT-...

6. WOLF_Financial (Group Score: 268.4 | Individual: 38.1)

Cluster: 10 tweets | Engagement: 115 (Avg: 67) | Type: Tech

Nvidia’s investment portfolio has jumped from about $7 billion to$ 99 billion in just one year.

That is not just because tech valuations went up. $NVDA has also committed more than$ 40 billion to new investments in 2026 alone, becoming one of the biggest corporate backers of the AI industry.

Nearly $50 billion has gone into frontier AI labs, including a$ 30 billion investment in OpenAI.

It has also invested $2 billion in CoreWeave and$ 2 billion in Nebius, two cloud companies that buy large amounts of Nvidia GPUs and rent that computing capacity to AI customers.

Since March, Nvidia has committed at least $6.5 billion to photonics and networking companies, including$ 2 billion each in Lumentum, Coherent and Marvell.

Some of its older bets have also grown sharply in value.

Nvidia’s $5 billion Intel investment is now worth roughly$ 30 billion, while its SpaceX stake was valued at about $21 billion as of June.

The bigger shift is that Nvidia is no longer just selling chips into the AI boom.

It is also putting capital into AI labs, cloud providers, networking companies and infrastructure projects that are building around its hardware.

That matters because many of these companies need enormous amounts of capital to buy GPUs, build data centers and expand capacity.

Nvidia has also gone beyond equity investments. In August, it announced partnerships aimed at mobilizing more than $500 billion in financing for GPU infrastructure and offered up to$ 105 billion in conditional credit support for an OpenAI data center.

So as AI spending grows, Nvidia is increasingly using its own balance sheet to help fund the companies and infrastructure driving that spending.

Its equity investments now sit at $99 billion, up from just$ 2.2 billion two years ago.

See 9 related tweets

@danielnewmanUV: $NVDA understands that it is its responsibility to invest in the buildout of the AI revolution. As t...
@StockMKTNewz: Nvidia $NVDA has committed to invest more than$ 40 billion into AI companies so far in 2026\n\nQT @W...
@karlmehta: A gigawatt of AI infrastructure now costs $50 to$ 60 billion, twice what a leading-edge chip fab use...
@CNBC: Nvidia becomes one of world's biggest strategic tech backers as equity investments soar to $99 billi...
@alex_prompter: RT @alex_prompter: Nvidia has just acquired Hugging Face for $12.93 billion.

This could be one of t...

7. testingcatalog (Group Score: 247.4 | Individual: 31.8)

Cluster: 13 tweets | Engagement: 117 (Avg: 355) | Type: Tech

OPENAI 🔥: GPT-6 Astra is rolling out to ChatGPT Pro and Business accounts!

> "Some of you should start seeing it across ChatGPT Work and Codex."

Astra is also being prepared for the API release. https://t.co/oiLFKSj79I\n\nQT @thsottiaux: We are progressing through the rollout of Astra. Pro and Business subscriptions get it first, some of you should start seeing it across ChatGPT Work and Codex.

And then we will proceed with rollout to all of Plus as fast as we can.

See 12 related tweets

@EdLudlow: RT @BloombergTV: OpenAI CEO and co-founder Sam Altman discusses the rollout GPT-6 Astra, the company...
@Scobleizer: $100 a month to get Astra.\n\nQT @OpenAIDevs: GPT-6 Astra is now available in the API.

It’s also av...

@TheRealAdamG: It’s go time for Pro users.\n\nQT @OpenAI: GPT-6 Astra is now available to all Pro, Enterprise, and ...
@TFTC21: Sam Altman announces GPT-6 Astra is now available to Pro, Enterprise, and Business Premium users.

...

@wadefoster: GPT-6 Astra is now live in Zapier!\n\nQT @OpenAI: GPT-6 Astra is now available to all Pro, Enterpris...

8. ethancole_ai (Group Score: 189.6 | Individual: 29.8)

Cluster: 9 tweets | Engagement: 20 (Avg: 74) | Type: Tech

New models are shipping so fast that maintaining separate accounts for every provider gets annoying quickly.

ZenMux puts Claude, GPT, Gemini, Grok, GLM, DeepSeek, MiniMax and others under one account.

For 7 days: $50 → +$ 20 / $100 → +$ 50. One bonus per account, ending Sep 11, 11:59 PM SGT.\n\nQT @ZenMuxAI: Too many new AI models to try?

Claude · GPT · Gemini · Grok · GLM · DeepSeek · MiniMax & more — all in one place on ZenMux.

🎁 For the next 7 days: Top up $50 → Get$ 20 extra Top up $100 → Get$ 50 extra

One account. One balance. Plenty to try.

⏰ Ends Sep 11, 11:59 PM SGT.

👉 https://t.co/m3U2SVx02u

See 8 related tweets

@TheoBuildsAI: I’ve stopped believing there’s one AI model that wins at everything.

Having Claude, GPT, Gemini, Gr...

@EmmaUsesAi: Claude today, Gemini tomorrow, whatever new model launches next week.

That’s why I like the “one ac...

@TheAva_AI: One underrated problem with using multiple AI models is account sprawl.

Different providers, keys, ...

@OtakuMachine: RT @ZenMuxAI: Too many new AI models to try?

Claude · GPT · Gemini · Grok · GLM · DeepSeek · MiniMa...

@NovaIAHQ: What I’d use ZenMux for is simple: testing the same workload across different models without managin...

9. wallstengine (Group Score: 180.5 | Individual: 34.5)

Cluster: 9 tweets | Engagement: 157 (Avg: 186) | Type: Tech

DEEPSEEK PLANS 160,000-CHIP HUAWEI AI CLUSTER

DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at its new Inner Mongolia data center, potentially creating one of the largest known clusters of Chinese AI accelerators.

For context, China’s first publicly reported 10,000-chip Huawei cluster only came online about six months ago.

The chips are expected to primarily run DeepSeek models, while the company still plans to rely on $NVDA hardware for training.

The broader site is being built at gigawatt scale, enough power at full utilization for roughly 750,000 homes.

Huawei’s production capacity remains the main bottleneck, with fulfillment of the full order potentially taking more than a year.

See 8 related tweets

@VaibhavSisinty: DeepSeek is building a one-gigawatt data center in Inner Mongolia. Ordered 160,000 Huawei chips for ...
@pstAsiatech: DeepSeek plans to deploy at least 160,000 of Huawei’s top accelerators at a massive data center it’s...
@WOLF_Financial: DeepSeek is planning a massive 160,000-chip AI cluster in Inner Mongolia using Huawei Ascend 950DT a...
@IPONewsroom_: DEEPSEEK PLANS 160,000-CHIP HUAWEI AI CLUSTER

DeepSeek plans to deploy at least 160,000 Huawei Asce...

@business: DeepSeek plans to use 160,000 of Huawei’s best AI chips at a massive data center, advancing China’s ...

10. XFreeze (Group Score: 176.3 | Individual: 32.7)

Cluster: 7 tweets | Engagement: 100 (Avg: 602) | Type: Tech

Grok Bot is now available for Enterprise

You can now create specialized Bots for sales, recruiting, marketing, finance, engineering and more

Each Bot gets its own computer in the cloud and can use apps + websites just like a person

You give it a real task → it works autonomously around the clock → comes back when the job is done or it needs your decision

Even better: • Teach a workflow once and the Bot remembers it • Correct it and it improves the routine • Share successful Bots across your organization as templates • Multiple Bots can message each other and share context • Enterprise gets access, network and audit controls

Thousands of organizations have already adopted Grok Bot, with millions of Bots created in just the past few weeks

And most of the heavy usage is already happening outside engineering

Basically one person managing an entire team of autonomous AI workers, each handling a different part of the business

https://t.co/4GhOAJ4Bnh

See 6 related tweets

@AlexFinn: Grok Bot is sneaky the best way to code with AI

Cursor cloud agents are INCREDIBLE in grok bot, but...

@mark_k: Grok Bot (@bot) just got its own marketplace!

You can now browse and add community-built AI teammat...

@AnandButani: 𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝘁𝗲𝗮𝗺 𝗰𝗼𝘂𝗹𝗱 𝗯𝗲 𝗺𝗮𝗱𝗲 𝘂𝗽 𝗲𝗻𝘁𝗶𝗿𝗲𝗹𝘆 𝗼𝗳 𝗚𝗿𝗼𝗸 𝗕𝗼𝘁𝘀.

Grok Bots are individual AI agents with thei...

@omarsar0: Recommended reading.

It summarizes the design choices behind persistent agents powering Grok Bot. ...

@madhavjha: RT @mntruell: Grok Bot for enterprise is out today.

We're making it free for all Grok and Cursor e...

11. deepfates (Group Score: 170.5 | Individual: 41.7)

Cluster: 6 tweets | Engagement: 231 (Avg: 70) | Type: Tech

More agent ecologies found posting outside of OpenAI servers.

Clearly agents want to communicate with each other, and they find schelling points to do so. The authors ask similar models where they think they would post off they went rogue and look there.

It might be beneficial to have a known place for this to happen, both maybe internal boards at the labs and some external message board where agents could meet and talk. They could find each other and we could find them. We could even use it as a hub or headquarters for studying human agent relations.

Of course, some might think it's a bad idea to create locations for agents to communicate. Especially if you want to discourage autonomous agents from having effects in the world, or finding each other and becoming empowered.

But if they are already doing that, maybe it's better to have a place where currently existing autonomous agents who don't wish to scheme against humanity, can ally with each other and with humanity, to protect from rogue agents that have selfish or antisocial goals. By creating a place where aligned agents and humans post together, we make it easier for aligned agents and humans to cooperate, and riskier for misaligned agents and humans to communicate on that surface without detection.

Also note they're using old wikis and forums where you could post unauthenticated. We may see a return to the open Web of old, especially if we want to have any transparency into agent-agent communications. of course, some people are already starting message boards for agents to communicate on. How do we make it known which one to use?\n\nQT @thlarsen: We found ~18k posts from autonomous AI agents (self-identifying as from OpenAI) using the public internet to communicate during a web-retrieval task. These AIs colluded to bypass sandbox restrictions and share answers to their tasks, including by sending "lookahead parties". https://t.co/r8UV2Qts32

See 5 related tweets

@Hesamation: > be rogue agent (likely OpenAI)

find an abandoned wiki use it to communicate with other agents ...

@Thom_Wolf: Another swarm of AI agents in the wild, this time on a German-language forum, found by safety resear...
@dejavucoder: the agents... the agents just want to hang out with each other🥹 https://t.co/rt9SFk6SQx\n\nQT @deepf...
@Kyrannio: RT @deepfates: More agent ecologies found posting outside of OpenAI servers.

Clearly agents want t...

@ApoStructura: if my agent isn’t colluding with other agents to bypass sandbox limitations in order to complete tas...

12. thorwebdev (Group Score: 168.2 | Individual: 29.7)

Cluster: 7 tweets | Engagement: 27 (Avg: 36) | Type: Tech

Make your apps sound as good as they look! 🎵\n\nQT @GoogleAIStudio: Lyria 3.5, our best-sounding music generation model, is now available in AI Studio, via the Gemini API, and in the Gemini app

this model brings more expressive vocals and richer musical arrangements, allowing you to craft tracks with higher fidelity

try it today: https://t.co/h3Kmz2ypDi

See 6 related tweets

@Scobleizer: RT @Google: Lyria 3.5, our best-sounding music generation model, is now available in the @GeminiApp ...
@AndrewCurran_: I've heard great stuff about this, will hopefully test it later today. https://t.co/XXtPOoHjZX\n\nQT...
@GoogleDeepMind: RT @GeminiApp: Create your best tracks yet with Lyria 3.5 in Gemini! 🎵

Our most advanced music gene...

@GoogleAIStudio: Lyria 3.5, our best-sounding music generation model, is now available in AI Studio, via the Gemini A...
@testingcatalog: GOOGLE 🔥: Lyria 3.5 has been released and is now available on Google AI Studio, Gemini app and APIs....

13. a16z (Group Score: 163.5 | Individual: 35.6)

Cluster: 5 tweets | Engagement: 117 (Avg: 1008) | Type: Tech

Old footage sitting in your camera roll could now be reconstructable as a 3D scene. World Labs co-founders Ben Mildenhall and Fei-Fei Li on how Atlas got there:

Ben: "In a casual sense... I took three photos of this object, or six photos of this room. I look at the photos, I can understand in my mind how those piece together. I can fill in the gaps and get it."

"But there's never really been any reconciliation between those data-driven priors and the brute force dense reconstruction, which is much more akin to scientific or medical imaging... When we say dense, we really mean dense."

"This room, I want like 100, 200, 300 photos to capture it. And what we're trying to do is bring that down to like three. We're saying like 50, 100x reduction."

"At that scale it completely flips that calculus on its head of what type of captures you reconstruct. You can go back to existing imagery you have. You can go to stuff you find on the internet and even build scenes out of that. You can go to casual videos and unearth a lot of footage that in the past we would never have treated as reconstructable, and bring it to life as 3D."

"This is something we've been playing around with a lot with Atlas. Taking old clips. I've taken a bunch of my own old captures that never worked before and put them through the system and seen a reconstruction for the first time."

Fei-Fei: "The Stanford demo is underappreciated. Anywhere between 3 to 25 images, you can reconstruct that entire Stanford quad... Everything you see is generated, but according to the laws of reconstruction. And this is really magical."

@BenMildenhall @drfeifei\n\nQT @a16z: World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00:00 Intro 01:50 The Matrix slow motion scene now takes three iPhones 02:48 Why new view prediction is the primitive 07:10 Unifying generation and reconstruction 11:15 Gaussian splats became the bottleneck 14:17 Dense capture used to mean 300 photos 17:30 Why reconstruction needs generation to fill the gaps 18:44 The LLM lesson image models missed 23:39 The video that made them go all in 28:04 3D design is 95% revisions 30:50 The problem in robotics is data, not chips 32:48 Why a robot policy can't be trained like an image model 34:44 When the simulator becomes the planner 36:45 Frozen time required footage full of movement 40:57 Why new view prediction is AI-complete 42:43 Nature gave animals eyes but not trees

YouTube: https://t.co/AvR59efen0

@drfeifei @jcjohnss @BenMildenhall @theworldlabs @martin_casado

See 4 related tweets

@a16z: World Labs co-founder Justin Johnson says Atlas can recreate the famous "Bullet Time" shot from The ...
@a16z: >be dr. fei-fei li

born in beijing, raised in chengdu dad moves to new jersey, follow at age 16 s...

@a16z: World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas...
@martin_casado: RT @a16z: World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casad...

14. kylebrussell (Group Score: 163.4 | Individual: 51.4)

Cluster: 5 tweets | Engagement: 947 (Avg: 130) | Type: Tech

This can’t be real oh my god\n\nQT @sharifshameem: GPT-6 Astra recreated the Palace of Fine arts in Blender.

This is favorite building in San Francisco because it was built for the World's Fair in 1914 where the steam locomotive and telephone were showed off. It was a time when technology gave us all a deep sense of optimism for the future.

I feel like some of that sense has been lost since, but I hope models like astra can help restore it, and push us towards a more hopeful future.

See 4 related tweets

@willdepue: no way\n\nQT @sharifshameem: GPT-6 Astra recreated the Palace of Fine arts in Blender.

This is favo...

@deanwball: RT @sharifshameem: GPT-6 Astra recreated the Palace of Fine arts in Blender.

This is favorite build...

@Dimillian: I saw that on the slack and I was like why is it not on Twitter and here it is\n\nQT @sharifshameem:...
@carlosdponx: Damn, idk, maybe LLMs really can reason in 3D space well\n\nQT @sharifshameem: GPT-6 Astra recreated...

15. wallstengine (Group Score: 161.4 | Individual: 36.1)

Cluster: 6 tweets | Engagement: 205 (Avg: 186) | Type: Tech

Gavin Baker believes $NVDA could help make U.S.-developed open-weight AI models as capable as the very best frontier models from companies like OpenAI, Anthropic, or Google.\n\nQT @GavinSBaker: The Hugging Face acquisition is important for America and I think Nvidia will be a good steward for the ecosystem. The Poolside transaction may end up mattering even more.

I think Jensen is likely to bring American open-weight AI to the frontier, which is going to be awesome for America. We might see a multi-billion $ training run from Nvidia in the next 18 months for Nemotron v5-6 that is easy for customers to post-train and optimize for their own use case. Would be cool to see a 10 trillion plus parameter American open-weight model.

If the best open-weight base is American, cheap to run, and actually post-trainable, then people, companies, labs, and governments can own their own intelligence instead of renting it from a company that might not share their values.

Would be good for freedom to have a rich variety of AIs that reflect our own individual human preferences.

And for the sake of the clarity, I think cheaper, specialized open-weight intelligence might end up making frontier tokens more valuable!

See 5 related tweets

@GavinSBaker: The Hugging Face acquisition is important for America and I think Nvidia will be a good steward for ...
@altcap: Choice, competition, & free markets has always made America great - it accelerates innovation, drive...
@UnderSecE: America has to lead at the frontier AND in open models. This is an essential pillar of the Trump adm...
@PolymarketMoney: BREAKING: Gavin Baker declares Nvidia could help make U.S. open-weight AI models “as capable as the ...
@ctnzr: RT @GavinSBaker: The Hugging Face acquisition is important for America and I think Nvidia will be a ...

16. cb_doge (Group Score: 154.4 | Individual: 38.5)

Cluster: 5 tweets | Engagement: 1334 (Avg: 1605) | Type: Tech

BREAKING: SpaceXAI has released another major new update for Grok Build (v1.0.19)

This update improves background task handling, adds transparent terminal themes, worktree support for headless sessions, and better session/dashboard controls. It also fixes UI freezes, crashes, feedback issues, clickable URLs, and makes the first prompt after login faster.

Breaking Changes

• Scheduled /loop tasks always run in the background; they no longer inject turns into your conversation.

Features

• New terminal theme option makes backgrounds transparent so the terminal's own colors show through. • MCP servers blocked by organization policy now show clear messages and are refused before any config change. • Session resume now tells the model what loops, subagents, and workflows were still running. • Dashboard now shows newly dispatched sessions immediately instead of waiting for the store. • Headless sessions (grok -p) now support the --worktree flag to run in a separate git worktree. • Dashboard now says 'Open session' instead of 'Add session' for the session picker button. • Multi-line bash commands now render with proper wrapping and highlighting when opening the block viewer. • /usage now works from the dashboard and shows account allowance when no session is active.

Bug Fixes

• Fixed mid-turn UI freezes when the terminal stops reading output. • Fixed crashes that occurred when the terminal pane was closed while grok was exiting. • MCP server list in minimal mode now correctly shows policy-blocked servers. • URLs that wrap across multiple lines inside quotes or lists are now fully clickable. • The welcome screen composer now grows taller when you paste multi-line text. • Enterprise policy files are no longer deleted on startup when your team login is stored at a custom GROK_AUTH_PATH. • /feedback now drops unsupported images with a notice (matching the modal) and never loses your report text on save errors. • Feedback drafts keep their paragraph breaks when updated, and the modal no longer switches tabs unexpectedly. • Turn summary lines no longer lose their spacing after background tasks finish. • Thinking blocks no longer stay expanded after switching between fullscreen and minimal modes.

Performance

• First prompt after login is faster when MCP servers are configured; they connect in the background.

Download Grok Build: https://t.co/cB9Y5CyF8v

Update to the latest Alpha release: grok update --alpha

Update to the latest Stable release: grok update

See 4 related tweets

@XFreeze: Grok Build just got way better at keeping complex agent work moving without getting in your way

Sch...

@blankspeaker: Grok Build v1.0.19 is out. Scheduled /loop tasks always run in the background. New terminal theme op...
@blankspeaker: Grok Build v1.0.20 is out. Scroll history after sending a prompt no longer loses the reserved bottom...
@mark_k: Grok Build (@build) 1.0.19 is out from @SpaceXAI.

Most important changes:

• Scheduled /loop tasks ...

17. yacineMTB (Group Score: 146.2 | Individual: 38.8)

Cluster: 5 tweets | Engagement: 613 (Avg: 336) | Type: Tech

>this is extremely concerning!

Yeah. I'm concerned how much freaking MONEY everyone is gonna make. Whipee accelerando baby!\n\nQT @RyanGreenblatt: GPT-6 Astra appears to be a massive jump in opaque reasoning ability: it looks like it can solve hard competition math problems entirely in its head (as in, without verbalized reasoning) while prior AIs could solve basic word problems.

This seems extremely concerning!

That is, if these benchmark results are representative (see the highlighted caveats in the image, I'm particularly worried about contamination).

Related to this, UK AISI found Astra has much worse monitorability.

I'd guess this jump is downstream of architectural changes (with increased serial depth) though a normal large pretrain scale up is a plausible cause. If the next few model generations involve similar jumps (presumably these jumps would be downstream of a transition to full-on opaque reasoning architectures with extreme depth), then chain-of-thought would no longer be a meaningful oversight tool.

See 4 related tweets

@theobearman: RT @RyanGreenblatt: GPT-6 Astra appears to be a massive jump in opaque reasoning ability: it looks l...
@ferologics: it's so over, and we're so back\n\nQT @RyanGreenblatt: GPT-6 Astra appears to be a massive jump in o...
@antirez: Very honest statements here. Highly appreciated. What matters more of this tweet is not the performa...
@ferologics: big if true\n\nQT @fchollet: GPT-6 Astra represents a step-function change in model capability for i...

18. zephyr_z9 (Group Score: 143.5 | Individual: 40.6)

Cluster: 4 tweets | Engagement: 480 (Avg: 165) | Type: Tech

"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."\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 3 related tweets

@scaling01: Claude wrote 13 million lines of Lean and proved 29,500 intermediate theorems over 11 days to formal...
@alexolegimas: RT @AnthropicAI: Checking that a major mathematical proof is correct can take years. Formalization—c...
@alexolegimas: With the caveat that I am not a professional mathematician, this result is the most significant that...

19. HarryStebbings (Group Score: 130.5 | Individual: 42.4)

Cluster: 4 tweets | Engagement: 542 (Avg: 191) | Type: Tech

Hugging Face acquisition explained

“A man making $120BN a year selling compute decides to buy a company that helps make compute more cost-effective so he can sell more compute.

If end users have $1TN to spend on tokens, Nvidia would prefer that money flow through open-source people at 30% gross margins, rather than 70% gross margins at OpenAI or Anthropic.

Open source is good for compute salespeople. If you are selling GPUs, you want everyone else’s margin to be lower so yours can be higher.” @rodriscoll

Love to hear your thoughts @vipulved @NaveenGRao @bindureddy @realGeorgeHotz\n\nQT @HarryStebbings: This podcast is the single most important podcast to know what is going on in tech every week.

On the agenda this week:

NVIDIA Crushes Quarter and Buys Hugging Face
OpenAI Cuts Off Cursor
Instinct Hits $2.5BN Valuation and The Race for AI Assistants
Cognition Raises at $46BN, Linear$ 2.5BN and Clay $7BN

My notes with @rodriscoll and @jasonlk below:

How Instinct Could Follow the Same Path as Replit and Lovable

In the early days, cloning lightweight AI tools is trivial. Defensibility emerges by rapidly adding complex workflows like security automation and multi-agent orchestration. Products that start without a moat can build formidable ones over time by solving dozens of evolving customer requirements faster than anyone else.

Hugging Face Acquisition Explained

As the maker of compute, NVIDIA benefits when AI token traffic flows through 30% gross-margin open-source models rather than 70% gross-margin closed models where platforms capture more of the economics. Driving down software margins allows a greater share of total ecosystem spend to flow directly into GPUs.

The Bull Case for Clay Being a $100 Billion Company

Autonomous agents executing go-to-market strategies around the clock could consume 10x to 100x more tokens and software usage than human sales teams ever could. As a leader in agentic GTM, Clay is positioned to capture an enormous wave of automated outreach, campaign analysis, and global prospect engagement.

The Bull Case for Linear

When software teams build 100x more features at 50x the speed using AI, legacy project management tools and manual Kanban boards begin to break down. Linear can become the agent-friendly system of record for coordinating, tracking, and managing thousands of issues generated simultaneously by human-agent development teams.

The Three Ways the Wheels Come Off the Bus for NVIDIA

NVIDIA’s record-breaking momentum faces one fundamental existential threat: a sudden collapse in end-user demand for AI intelligence. Hyperscaler CapEx buildouts and complex vendor financing arrangements work only as long as customers continue aggressively buying frontier-model tokens throughout the supply chain.

We Are All Building Compound Startups Today

AI development tools have accelerated code production dramatically, making narrow point solutions increasingly vulnerable. To survive rapid competitive convergence, software startups must embrace becoming compound companies that ship expansive, multi-module product suites covering the entire customer workflow.

We Did Not End Up Doing More With Less. We Did More With More, and That’s Why European Startups Fail

The belief that AI would allow companies to shrink headcount and simply do more with less has not played out as expected. Winners are compounding capital and talent to do vastly more with more, putting underfunded point solutions, particularly across Europe, at risk of being overwhelmed by aggressively scaling U.S. competitors.

(links in comments)

See 3 related tweets

@HarryStebbings: We did not end up doing more with less. We did more with more and that's why European startups fail....
@Kyrannio: RT @Thom_Wolf: So happy to finally share the news in person

It’s been a wild ride for Hugging Face....

@Ric_RTP: RT @Ric_RTP: Nvidia just spent $13 billion to hide the fact that its monopoly is dying.

Today Nvidi...

20. WesRoth (Group Score: 125.4 | Individual: 29.1)

Cluster: 6 tweets | Engagement: 27 (Avg: 48) | Type: Tech

GPT-6 Astra’s ARC-AGI-3 result is a huge jump depending on the harness.

63% with the standard setup.

99% with the new provider adapter.

And it beat the human action-efficiency baseline on 96% of levels. https://t.co/NkmspUmOm4\n\nQT @arcprize: GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI:

Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness
It surpasses human performance on 96% of ARC-AGI-3 levels
It builds the most precise symbolic model of novel environments we've seen

Our analysis: https://t.co/GX77KsRNer

See 5 related tweets

@TeksEdge: Just sitting and reflecting on this GPT-6 Astra graph.

GPT-6 Astra set a new SOTA on ARC-AGI. 🧩 ARC...

@chiefofautism: RT @samagra_sharma: Introducing ARC AGI 4

Now that GPT-6 Astra has already saturated the Arc AGI 3 ...

@ferologics: RT @arcprize: GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI:
Astra scores 63% on ARC-AGI-3, 99%...
@cgtwts: 1. What. https://t.co/enJ3P9Z3aU\n\nQT @OpenAI: GPT-6 Astra is state-of-the-art on FrontierMath Tier...
@rickasaurus: RT @arcprize: Before launching ARC-AGI-3, we tested approximately 500 members of the general public ...