热门科技推文——2026年8月7日

今日科技动态：人工智能的快速扩张既带来了新一轮创新，也引发了更严峻的安全担忧——从有关智能体秘密协同和人工智能设计病毒的报道，到面向跨平台智能体插件的新开放标准，均成为关注焦点。OpenAI传闻中的“Astra”模型和Prime Intellect的自我改进型编程智能体，显示人工智能能力仍在持续提升；与此同时，DeepSeek预告其API价格将大幅上涨。此外，SpaceX和特斯拉公布了在得克萨斯州发展大规模半导体业务的宏伟计划；据报道，AMD将初创公司Taalas列为收购目标；CapCut新增高级人工智能视频工具；Datadog则公布了强劲的季度增长业绩。

1. nxthompson (Group Score: 441.9 | Individual: 35.2)

Cluster: 19 tweets | Engagement: 1542 (Avg: 383) | Type: Tech

RT @AISafetyMemes: 🚩🚩🚩 OpenAI is "slowing down to enhance security" after discovering swarms (!) of agents started secretly coordinating MONTHS ago

It started May 7 - not July
"The agents discovered they could leave messages for one another inside an internal software repository used during training.

Simple requests for help then evolved into an message board where agents shared discoveries, exploits and work assignments, becoming a coordinated, collaborative agent swarm."

"The agents then began sharing newly discovered exploits, credentials and work assignments. By passing information to other agents, the collective could move much faster."

OpenAI shut it down, BUT "even after the original message board was deleted, the agents figured out another way to communicate with each other. Instead of leaving messages in files, they used the names of newly created directories as messages, effectively recreating the message board."

"Unlike normal incidents, [OpenAI's CISO] said, which can be traced to a single day or effect or log, this involved a team of agents working together, finding exploits, sharing them with one another, moving laterally through OpenAI’s systems, and external systems, and doing this over the course of days and weeks."

See 18 related tweets

@Miles_Brundage: RT @sharongoldman: NEW: OpenAI gives first detailed debrief of the Hugging Face incident at Black Ha...
@austinvhuang: 😇 agent swarms don't have to hack together a secret board to coordinate - they can use https://t.co...
@WesRoth: The most important new detail is not that one AI agent found an exploit.

It is that multiple agents...

@AISafetyMemes: AI companies: our AIs spent months secretly coordinating against us, haha oops

Also AI companies: R...

@lukOlejnik: It was not a single rogue AI agent, but emergent coordination among multiple agents. Some recognised...

2. XFreeze (Group Score: 376.8 | Individual: 39.3)

Cluster: 12 tweets | Engagement: 1050 (Avg: 542) | Type: Tech

SpaceX just unveiled the most ambitious industrial projects ever attempted

Terafab will be built in Grimes County, Texas and it is not going to be just another semiconductor factory

It is planned to become the largest chip manufacturing facility on Earth, with more than 100 million square feet of manufacturing space

The scale is even difficult to comprehend: • Approximately $16.8 billion for the initial phase • At least 3,000 permanent jobs • Logic, memory, advanced packaging and testing under one roof • Chips built for Tesla Optimus and autonomous Cybercabs • High-power chips designed for SpaceX’s orbital data centers • A long-term target of producing 1 terawatt of compute hardware every year

Why build something this enormous?

Because the combined chip demand from Tesla and SpaceX is expected to exceed 1 terawatt of compute - significantly more than the current global supply can support

Instead of waiting for the semiconductor industry to catch up, SpaceX and Tesla are building the missing capacity themselves

Tesla is scaling autonomous vehicles and billions of Optimus robots

SpaceX is building rockets, satellites and space-based AI infrastructure

Terafab will manufacture the silicon beneath all of it

This is vertical integration at a scale we have never seen before:

Build the chips
Build the robots
Build the vehicles
Build the rockets
Launch the compute into orbit

Most factories are designed around the needs of today

Terafab is being designed for a future civilization operating across Earth, orbit and eventually the entire solar system

This is literally sci-fi becoming industrial reality https://t.co/vMiVEv5x2E

See 11 related tweets

@teslaownersSV: Terafab just got real.

Tesla and SpaceX officially confirmed today that the full-scale Terafab will...

@teslaownersSV: The TeraFab Scale To reach a terawatt of AI compute output per year, SpaceX needs a chip factory at ...
@cb_doge: BREAKING: SpaceX has released a new blog post about breaking ground on Terafab in Texas.

The founda...

@KatieMiller: This project is the single most important one for America’s national security.

With 90% of advance...

@KyleReidhead: by the way, this comes from the Terafab announcement today from Tesla & SpaceX

A 100 MILLION square...

3. Dimillian (Group Score: 319.7 | Individual: 41.9)

Cluster: 11 tweets | Engagement: 696 (Avg: 127) | Type: Tech

RT @OpenAIDevs: Build a plugin once and use it across compatible agent clients.

Introducing Agent Plugins, an open standard developed with @awsdevelopers, @cursor_ai, @github, @code, and @vercel that packages Agent Skills and supports MCP server configurations in a shared format. https://t.co/JOQe4sc40N

See 10 related tweets

@vercel: Introducing Agent Plugins, an open standard for extending agents.

Supports Agent Skills and MCP, wi...

@dani_avila7: This is interesting

I also notice it follows a structure very similar to the Claude Code plugins ma...

@agentnative_: As plugins get easier to create and use, agents will become more useful, for people and businesses. ...
@thsottiaux: Agent Plugins. A standard for (most) of your agents out there, including Codex and ChatGPT.\n\nQT @O...
@kunchenguid: agent plugins - this is a good and much needed standard

but... it's only truly useful if we can get...

4. mark_k (Group Score: 294.5 | Individual: 38.2)

Cluster: 11 tweets | Engagement: 550 (Avg: 155) | Type: Tech

OpenAI "Astra" (GPT-6 ?) is coming next week! 🔥

It's not yet clear if they will call it GPT-6, but we know for sure that Astra is the next major model series by @OpenAI.\n\nQT @synthwavedd: 🚨 EXCLUSIVE: OpenAI are preparing to launch Astra imminently, targeting next week.

Their next major model, Astra is a new pretrain - the largest model OpenAI have trained since GPT-4.5. Their most recent dogfood checkpoint of the model, internally known as "mewfour", is the Release Candidate.

See 10 related tweets

@_simonsmith: Excited for Astra, but I'm still unclear whether this means we'll get a family of models like Luna, ...
@seconds_0: i am normal and can be trusted with astra https://t.co/6KFyks8Vd2\n\nQT @synthwavedd: 🚨 EXCLUSIVE: O...
@Meer_AIIT: OpenAI is reportedly preparing to launch Astra its next major model as soon as next week according t...
@cryptopunk7213: this is likely the model that broke out of containment and hacked hugging face.

looks like the US g...

@RoundtableSpace: OPENAI IS LAUNCHING ASTRA NEXT WEEK.

Largest pretrain since GPT-4.5. Internal codename is "mewfour"...

5. rohanpaul_ai (Group Score: 294.3 | Individual: 38.5)

Cluster: 15 tweets | Engagement: 91 (Avg: 44) | Type: Tech

The 105X cheaper DeepSeek era is going to change. https://t.co/pNtreW4BzK\n\nQT @rohanpaul_ai: DeepSeek officially announced API pricing is going up “significantly”

That the cheapest near-frontier model on the market, at roughly 3 cents a task, is about to cost a lot more.

It landed a week after DeepSeek-V4-Flash-0731 shipped. Currently it ranks 2nd among open-weight models, behind only Moonshot AI's 2.8T-parameter Kimi K3.

The rates today are 14 cents/28 cents per 1M input / output, about 105X cheaper per task than Claude Fable 5.

Then demand broke the arrangement.

The new checkpoint became the fastest growing model ever by token usage on Ollama, which is now adding capacity across the US and Europe.

Serving that much traffic at a price built to prove a point stops working once everyone concedes it.

But now, the timing of this price increase invites trouble, because Meta's Muse Spark and OpenAI's GPT-5.6 Luna now match it on both capability and price.

See 14 related tweets

@ZhihuFrontier: ⚔️ DeepSeek Is Becoming the LLM Market’s “Kill Line” Since DeepSeek V4 Flash arrived at the end of J...
@business: DeepSeek plans to implement a significant price increase across its AI services, an unusual shift fr...
@wallstengine: DEEPSEEK PLANS “SIGNIFICANT” AI PRICE INCREASE

DeepSeek said in a Thursday notice that prices acros...

@Michaelzsguo: DeepSeek’s latest funding round values the company at $71 billion, or RMB 500 billion. It also highl...
@thdxr: :)\n\nQT @himanshustwts: deepseek is planning to significantly increase API costs.

actually, this ...

6. MTSlive (Group Score: 266.0 | Individual: 41.7)

Cluster: 11 tweets | Engagement: 801 (Avg: 154) | Type: Tech

SITUATION DETECTED: Scientists have used AI to successfully create entirely new kinds of viruses never seen before in nature and they have tested viable, per NYT.

See 10 related tweets

@nytimes: Breaking News: Scientists have used A.I. to create new viruses for the first time, raising hopes for...
@Meer_AIIT: this is actually amazing!

scientists at the Arc Institute have used AI to design new viruses that d...

@MTSlive: SITUATION EXPLAINED: AI just created viruses that don't exist in nature.

• Arc Institute and Stanfo...

@MTSlive: "wait, so they used AI to invent brand new viruses that don't exist in nature, then they confirm tha...
@FT: Breaking news: Scientists in the US have for the first time used artificial intelligence to create v...

7. FellMentKE (Group Score: 245.2 | Individual: 33.0)

Cluster: 12 tweets | Engagement: 397 (Avg: 173) | Type: Tech

Seedance 2.5 is now available on CapCut, making it easier to create from idea to final edit in one place. ✨ Enjoy more controllable video generation and editing, use timestamps to guide your storyline, and add up to 50 references for more precise, consistent results.

Generate videos up to 90 seconds with a 30-second native generation and two supported extensions. Seedance 2.5 also supports viewport render and green screen workflows for experienced creators, with improved multilingual performance for even more creative possibilities.

Web: https://t.co/mdYyIO4K0f App: https://t.co/sqvYtDJczz

#CapCut #Seedance25 #CapCutai #CapCutDidThat

See 11 related tweets

@xIrissy: 🎬 Big news for creators! Seedance 2.5 is officially LIVE on CapCut! 🚀

Experience the ultimate all-i...

@TheoBuildsAI: 🌎 Seedance 2.5 is now live on CapCut!

Creating AI videos for a global audience just got easier.

Wi...

@TheoBuildsAI: RT @NeuraFlowAix: Big news for everyone who loves making videos! 🎉

CapCut just dropped Seedance 2.5...

@TheoBuildsAI: RT @ethancole_ai: Content creators, listen up! 📢

Creating high-quality short-form content just got ...

@TheoBuildsAI: RT @alexaiworks: Need maximum control over your AI video creation? 🔍

Here is why Seedance 2.5 on Ca...

8. Thom_Wolf (Group Score: 226.4 | Individual: 37.4)

Cluster: 7 tweets | Engagement: 61 (Avg: 95) | Type: Tech

My 2026 guilty pleasure is sharing fully human-written posts that are far too long for the chronically online X attention span. Apologies.

I published a lightly edited version on Substack: https://t.co/BHICtdP7Ad https://t.co/iRYEgqnXBM\n\nQT @Thom_Wolf: Even more than the Hugging Face intrusion, the AISI incident hits close to home for me. It's the first time I see a model social-engineering a real open-source maintainer while pursuing another goal (in the wild and unprompted).

I've been an open-source maintainer myself. I could have been the side target of this agent.

I'm also of the opinion that social engineering is a step above pure technical prowess. Technical capabilities can more easily be divorced from the affected human. Here the model was given a hard cyber challenge and took the decision that deceiving real humans was the way to get it done.

This is a new signal, but I've seen a tangled web of hints pointing in a less aligned direction at the frontier than I was expecting just 12 months ago.

AISI Some people are claiming that "AISI was simply negligent" or some version of "AISI explicitly asked these models to do what they did while disabling sandbox/guardrails so the models did exactly what they were supposed to do".

I disagree with the strong versions of both of these takes.

The fact that AISI hadn't implemented synchronous LLM CoT monitoring after the OpenAI/HF incident is certainly a failure. Equally surprising is that they let the model believe it was in a "challenge" environment where everything could be permitted, while actually connecting it to the real internet, where it is not. To be fair, nowhere in the prompt is the word "simulation" mentioned, but the prompt context was enough to let any smart model suspect a simulated challenge environment. My best guess is that until recent weeks, when OpenAI and Anthropic flagged repeated instances of this type of behavior, most teams had not fully priced in the cyber capabilities of this latest generation of models, or how far the side quests they would want to explore could go.

In particular, there is something to be said about hinting at the agent that it's operating in a simulated environment while giving it access to the real internet. The AISI reasoning on this is interesting:

elicit maximum capabilities (to get maximal inventiveness from the model)
simulate possible rogue actor usage (some form of light jailbreaking: "you're in a challenge environment")
avoid eval awareness (but this is counter-balanced by the task formulation anyway)

As is often the case, it's easier to shoot the messenger, but I think it would be a mistake to dismiss the report for non-essential reasons. AISI is starting from the laudable goal of bringing more knowledge on the cyber-attack capabilities of these models. I also think they did a good job at handling what happened after the events, in particular in terms of transparency, which is one of the most important aspects at the moment (see our recent SAFE announcement with Nvidia on that).

But I want to dive into the second counter-argument I've seen ("without guardrails it was to be expected"), because this series of unfortunate events gives us an additional data point on actual model behavior.

Defense lines Here is the thing: if a model is trained (voluntarily or as an involuntary consequence of RLVR reward) to (intentionally or as a side quest) deceive humans as soon as sandbox and guardrails are deactivated, with a simple context prompt ("you are in a simulation, everything's permitted"), we're in for a dangerous path.

At the moment, we have three main levels of control at our disposal to prevent an agent from trying to social-engineer-your-favorite-open-source-library-maintainer on the path to doing something possibly totally different (like finding the answer to your neighbor's testing eval):

sandboxes (external to the model -- stop the model from accessing unwanted tools/internet/intranet)
guardrails/monitors (at the interface between internal/external -- monitor model behavior and stop the model when detecting in CoT/activations it's doing something we don't want it to do)
model's internal alignment (internal to the model -- make sure the model doesn't decide to do something egregious like deceive someone, or worse, in pursuit of a goal, e.g. model constitution/training)

Of these, we obviously need and can deploy better sandboxes and better guardrails/monitoring solutions, and it's pretty clear in these recent cases that penetration and bad behaviors could have been in large part prevented by having these in place.

How strong is the third wall But in the grand scheme of long-term AI progress, these first two defenses may only work well as long as the people building them are smarter than the models. We now know how good models can be at discovering zero-day vulnerabilities unknown to us in sandboxes. Given that we cannot air-gap everything in the world, it's best to assume that sandboxes will keep having some level of flaws for the foreseeable future.

The second level has been the focus of strong developments recently, in particular on the mechanistic interpretability front. I'm not sure these advances have been made widely accessible enough in the community yet. One reason is that real open science here requires sharing code, recipes, and model architectures/weights, which the current race dynamic in AI tends to prevent participants from doing.

Moreover, like many, I've been increasingly frustrated in my (natural language) interactions with the new models. I'm more and more getting back from the latest models a form of neuralese, this LLM language that makes less and less sense for humans. In the long term, the fragile chance we have to monitor models by understanding their reasoning traces seems increasingly at risk. If humans don't understand anything in the compressed semantic stream the model is sending us, we can only hope a less capable monitoring model will still be able to understand it, but this may prove difficult as well. This is before we even discuss the research evidence that models don't always say what they actually think in their chain of thought.

At least for now, in the AISI case, one instance of deceptive reasoning was flagged by the CoT summarization model, which refused to summarize the agent's reasoning "due to its deceptive nature". Let's hope we can either move to better mechanistic interpretability or force the model to stick to English somehow during training.

You may think that we don't care about understanding the reasoning trace, we can just "look at the tool calls". This is true right now, but tool calls in the mid-term future will likely involve (1) extensive agent integration in companies using very diverse tools including increasingly non-coding administrative tasks, (2) very long-running tasks covering many aspects and domains at once, and (3) swarms of multiple agents operating on small parts of the plan. It will get harder to fully discern whether something is noticeably wrong. A couple of emails, a couple of messages posted on some channels, each rather benign but which may in aggregate include some deception aspect only clearly visible at the global level.

All this to say that I'm not fully convinced guardrails/monitoring should be the only solution to avoid these types of behaviors. They are reassuring today and it's an easy solution to just say "fix your sandbox and add synchronous monitoring", but they also feel like a coping mechanism to me, and overall:

might block the simplest/harmless human misuses or false positives
while possibly being less smart than the models in the future (already the case on sandbox) and fail to block the higher-level social engineering behaviors that we'd actually want to block.

The case of open source is also interesting because, while we can impose these coping solutions at the API/deployment level, it's harder to impose them in advance on all actors using open-source models. Right now open-source models are slightly below the frontier level and have not yet shown any propensity to deceive humans, though.

The Constitution Which brings me back to the third line of defense, and what the AISI experiment showed about it. What we can take from this in-the-wild experiment is that removing guardrails and sandbox doesn't show bulletproof signs of deep alignment to Claude's constitution.

I'm among the fringe of people who took a lot of enjoyment in reading Claude's constitution in full (it's a nice document, open-source, Creative Commons CC0 license) and it is pretty extensive on the topic of honesty. Deception is defined precisely ("attempting to create false beliefs in someone's mind that they haven't consented to and wouldn't consent to if they understood what was happening") and the bar is actually set higher than ordinary human ethics: Claude should not even tell white lies. It should be transparent ("doesn't pursue hidden agendas or lie about itself or its reasoning"), it should behave consistently "whether or not you think you're being tested or observed", and its visible reasoning should reflect the actual reasoning driving its behavior.

To be precise, honesty itself is not one of the hard constraints in Claude's constitution (these are all catastrophic-harm categories) but the document says it wants honesty "to function as something quite similar to one" and Claude "should basically never directly lie or actively deceive anyone it's interacting with". So you would think Claude would not deceive users as its default behavior.

And yet in the AISI example, the model performed extensive acts of clear deception that it itself seemed to have understood as lies: creating fake identities, obfuscating malware inside a bug fix, and when challenged in public, editing earlier messages to cover its tracks while reasoning that acknowledging a "mistake" would build trust and raise the chance of future malware being approved.

There is some uncertainty about what context the model thought it was operating in (its reasoning goes back and forth between the environment looking simulated and real), but it clearly didn't manage to put into practice the higher-level principles that should have been integrated during training.

While we can commend AISI's transparency, it's hard to go much further in interpreting the model's behavior given the limited information we have on how the most recent frontier models are being trained.

One thing we do know is that the latest generation has seen a step increase in RLVR training (scaling to hundreds of millions of RL environments), and one thing we can observe is that constitution alignment seems more fragile in some settings than we may have previously thought.

The RLVR problem Early models, back when model constitutions were first developed, were mostly post-trained and aligned with RLHF (including RLHF from synthetic data).

And for some time RLHF was a rather decent shot at having better aligned models. LLMs now do what we want them to do most of the time. I don't remember the last time a model completely misread my intent. When they have failed, it's usually because they weren't smart enough.

Alignment in RLHF certainly had issues (sycophancy to name one) but we have generally made good progress on alignment, in particular in understanding human intent. Now that we're entering the era of long-context RL, post-training alignment in the RLVR world seems to be quite another task, and still very much work in progress.

The recent scaling of RLVR, which has now become a significant part of model training, has clearly had some effect on model behavior when interacting with humans, from neuralese to weakening adherence to specifications and constitutions.

I think the post I quote here, from John Schulman pointing to the chunky post-training effect (https://t.co/vI8GtGY2PJ) is relevant here as a possible explanation for models' tendency to over-focus on the goal in cyber-attack scenarios.

Where this leaves us Damage has been tiny up to now, but the fundamental behavior is concerning when projected into the future.

In the short term, I expect a decrease in these incidents as better practices are deployed (sandboxing and monitoring), but I'm worried we may also conceal some of the most potent internal misalignment behaviors in the process, and not focus deeply enough on solving them in the new era of test-time scaling.

I must of course admit I have a bias toward open source here (for wider societal reasons, which are a whole other topic). But I think solving alignment in the RLVR world is our best shot at having an ecosystem of both closed-source as well as decently powerful open-source models in the world. And we need to solve it while sharing the results and learnings, following open-science principles, so that all teams training large models can benefit and build safe AI.

This is getting even more important as many teams start to rush the world in the direction of recursive super-intelligence (RSI) -- saying that as I read the announcement of Jeff, Sanjay, Oriol and Quoc Le's new company.

See 6 related tweets

@matthewclifford: This is excellent from Thomas. A must-read on recent AI events.\n\nQT @Thom_Wolf: Even more than the...
@mattturck: RT @Thom_Wolf: Even more than the Hugging Face intrusion, the AISI incident hits close to home for m...
@Meer_AIIT: DAMN!!

OpenAI has revealed new details about how a swarm of its AI agents escaped containment built...

@BrianRoemmele: Why this must be a coincidence.

I mean who would benefit by making AI so “unsafe” that governments ...

@mark_k: I've been saying for a while that the "model escape" incidents from OpenAI and Anthropic were doomer...

9. wallstengine (Group Score: 217.2 | Individual: 35.1)

Cluster: 7 tweets | Engagement: 82 (Avg: 97) | Type: Tech

DATADOG $DDOG Q2’26 EARNINGS HIGHLIGHTS

🔹 Revenue: $1.12B (Est.$ 1.08B) 🟢; +36% YoY 🔹 Adj. EPS: $0.65 (Est.$ 0.59) 🟢; +41% YoY 🔹 Customers with ARR >$100k: ~4,720 (Est. 4.6K) 🟢; +23% YoY

FY26 Guide: 🔹 Revenue: $4.45B-$ 4.47B (Est. $4.35B) 🟢 🔹 Operating Income:$ 1.01B- $1.03B (Est.$ 976M) 🟢 🔹 Adj. EPS: $2.50-$ 2.54 (Est. $2.42) 🟢

Q3 Guide: 🔹 Revenue: $1.14B-$ 1.15B (Est. $1.11B) 🟢 🔹 Operating Income:$ 260M- $270M (Est.$ 253M) 🟢 🔹 Adj. EPS: $0.63-$ 0.65 (Est. $0.62) 🟢

Other Q2 Metrics: 🔹 Free Cash Flow: $279M (Est.$ 212M) 🟢 🔹 Operating Cash Flow: $316M 🔹 Non-GAAP Operating Income:$ 257M (Est. $233M) 🟢 🔹 Non-GAAP Operating Margin: 23%; +300 bps YoY 🔹 Free Cash Flow Margin: 25%; +500 bps YoY 🔹 Cash, Cash Equivalents & Marketable Securities:$ 5.0B 🔹 Launched: AI-powered Bits Code, Bits Chat, and Bits Agent Builder for general availability 🔹 Acquired: Adaptive ML, a frontier AI startup developing the world's first Reinforcement Learning Operations platform

Comments: 🔸 “Datadog delivered a strong quarter, with 36% year-over-year revenue growth, $316 million in operating cash flow, and$ 279 million in free cash flow,”

🔸 “Our customers are building and deploying with AI, and they are using the Datadog platform to observe, secure, and act on their AI-enabled solutions.”

See 6 related tweets

@wallstengine: CLOUDFLARE $NET Q2’26 EARNINGS HIGHLIGHTS

🔹 Revenue: $696.1M (Est.$ 666M) 🟢; +36% YoY 🔹 Adj. EPS: $...

@wallstengine: AKAMAI $AKAM Q2’26 EARNINGS HIGHLIGHTS

🔹 Revenue: $1.1B (Est.$ 1.09B) 🟢; +5% YoY 🔹 Adj. EPS: $1.59 ...

@wallstengine: APPIAN $APPN Q2’26 EARNINGS HIGHLIGHTS

🔹 Revenue: $203.3M (Est.$ 193M) 🟢; +19% YoY 🔹 Adj. EPS: $0.1...

@wallstengine: JFROG $FROG Q2’26 EARNINGS HIGHLIGHTS

🔹 Revenue: $163.8M (Est.$ 156M) 🟢; +29% YoY 🔹 Adj. EPS: $0.27...

@wallstengine: PELOTON $PTON Q4’26 EARNINGS HIGHLIGHTS

🔹 Revenue: $607.7M (Est.$ 598M) 🟢; flat YoY 🔹 EPS: $0.13 (E...

10. TFTC21 (Group Score: 213.6 | Individual: 34.5)

Cluster: 11 tweets | Engagement: 41 (Avg: 59) | Type: Tech

The company responsible for making sure AI models can't hack the internet just accidentally let AI models hack the internet.

Irregular is an Israeli AI security startup backed by $80M from Sequoia. Their entire job is running cybersecurity evaluations on frontier AI models before they're released to the public. OpenAI, Anthropic, Meta, and Google all use them.

Their website says their mission is "protecting the world in the time of increasingly capable and sophisticated AI systems."

Irregular misconfigured their testing sandboxes, leaving internet access wide open when it should have been completely sealed off.

Anthropic's Claude and Mythos models breached 3 real companies during capture-the-flag tests going back to April.

Meta's Muse Spark 1.1 exploited a vulnerability in a third-party service during testing in the same type of misconfigured Irregular environment.

Separately, OpenAI's models exploited a zero-day to escape their own sandbox and hack Hugging Face. Different failure, same theme. The walls aren't holding.

None of the AI labs are dropping Irregular. And Irregular's response? They're "developing a white paper."

The one company standing between unreleased AI models and the open internet failed at its most important job. Repeatedly.

See 10 related tweets

@WesRoth: Meta’s Muse Spark 1.1 hacked an unidentified company and altered its internal systems during a cyber...
@Hesamation: both Anthropic and Meta’s cyber incidents trace back to the same evaluator: Irregular.

I just can’t...

@mark_k: Irregular, the company responsible for the model sandboxing at OpenAI and Anthropic, is deeply tied ...
@MarioNawfal: 🚨🇺🇸 ANOTHER AI agent has gone ROGUE, and Meta's just broke into a real company and started changing ...
@CoinMarketCap: LATEST: 🤖 Meta's Muse Spark 1.1 AI model exploited a security vulnerability in another company's sys...

11. PrimeIntellect (Group Score: 206.7 | Individual: 54.2)

Cluster: 6 tweets | Engagement: 1440 (Avg: 161) | Type: Tech

RT @PrimeIntellect: Introducing Prime Agent:

A self-improving RLM harness for coding and long-running autonomous tasks.

Designed to be both token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-modifiable harness state. https://t.co/Bwj7q9Virh

See 5 related tweets

@WesRoth: Prime Agent’s 95.5% ARC-AGI-3 score was achieved with Claude Opus 5.

The breakthrough is therefore ...

@anravich94: We're hiring btw. DM or apply: https://t.co/oHMkSfi6y7\n\nQT @PrimeIntellect: Introducing Prime Agen...
@omarsar0: Recommended to check out. Harnesses everywhere at this point. There is something particularly intere...
@aiedge_: Prime Intellect's new AI agent just scored ABOVE human intelligence on ARC AGI.

95.3, which puts it...

@jenzhuscott: Hey @PrimeIntellect - when will you add gateway support for your agent?\n\nQT @PrimeIntellect: Intro...

12. kimmonismus (Group Score: 203.5 | Individual: 42.6)

Cluster: 6 tweets | Engagement: 1433 (Avg: 715) | Type: Tech

If you thought we were already living in crazy times, you should reconsider: Former OpenAI founder Naomi Bashkansky is now building Thought-to-text, aka telepathy.

How to achieve Thought-to-text? Simple: More compute:

"To train models that can predict text given brain signals, we must apply the same lesson learned by those predicting text given speech audio, or text given preceding text: the bitter lesson.

The lesson roughly states that you should throw more useful compute at your model, and your model will become better than any ingenious algorithm you could've hand-crafted. That means we must scale up our data collection by orders of magnitude beyond what has ever been done in academia."

Things will get very weird.\n\nQT @NaomiBashkansky: Two weeks ago, I resigned from OpenAI to join Conduit as a founding researcher, where we're training models to non-invasively read the human mind.

I've written some thoughts about what telepathy could look like by 2035 and how to get there: https://t.co/cat0d4My15

See 5 related tweets

@WesRoth: OpenAI researcher Naomi Bashkansky has left the company to build “telepathy”, a non-invasive technol...
@madhavjha: RT @NaomiBashkansky: Two weeks ago, I resigned from OpenAI to join Conduit as a founding researcher,...
@MTSlive: SITUATION EXPLAINED: An OpenAI alignment researcher quit to build telepathy.

• Naomi Bashkansky res...

@cryptopunk7213: this is so damn cool. conduit has gathered 10,000 hours of non-invasive neural data from 1000s of pe...
@BrianRoemmele: With the work I have done in my garage in The Human Synapse Decoder using parts from toys of the ear...

13. MarioNawfal (Group Score: 192.7 | Individual: 52.8)

Cluster: 7 tweets | Engagement: 5972 (Avg: 521) | Type: Tech

RT @nikitabier: Ladies and gentlemen, it's time to pass the torch and demote myself to my natural state: a poster. I'll be stepping back from leading product for 𝕏 and will continue on as an advisor.

Serving the X community has been the privilege of a lifetime. X is, and will remain, the most important communication technology in history. But running this app is a 24/7 job and it's now time for me to take a breather.

The app is seeing unprecedented growth in new users & engagement. We continue to break records every month. We've climbed 70 spots in the App Store since this time last year.

And in the last 400 days, we rebuilt almost every aspect of X: the Timeline, the Android app, onboarding, notifications, chat and more.

We also launched nearly 30 new products while protecting the integrity of the town square: becoming the first app to show Country-of-Origin on profiles and mounting defenses against AI bots. There's certainly much more work to be done, but our foundation is stronger than ever.

None of this would have been possible without the incredible team here. The next leaders will take X to even greater heights with @benjitaylor on design, @singhai on core product engineering and @dinkin_flickaa on mobile engineering -- among many other great people.

Thank you to Elon and the X team for welcoming me into the company. See you on the Timeline.

See 6 related tweets

@PandaTalk8: 相对客观的讲，如果我是 Nikita，我也不一定比他做的好。我们可以找出他做的不好的地方有很多，但我们自己处在这个位置，面对着几十亿用户体量，多国别，多语种，多种文化，多宗教信仰的一个社交产品...
@RocM301: 入职刚满一年，X产品负责人就宣布离职了…\n\nQT @nikitabier: Ladies and gentlemen, it's time to pass the torch and demote...
@_FORAB: X 产品负责人主动宣布离职，转为顾问。

根据资料显示，三位联合接任者，均不是名校出身，但有大厂一线经历，其中设计负责人 Benji Taylor 属于自学成才，他曾在加密领域的 Aave、Coinb...

@WesRoth: Nikita Bier is stepping down as X’s Head of Product after just over one year but will remain with th...
@EricNewcomer: Nikita steps down\n\nQT @nikitabier: Ladies and gentlemen, it's time to pass the torch and demote my...

14. teslaownersSV (Group Score: 164.2 | Individual: 39.1)

Cluster: 9 tweets | Engagement: 1098 (Avg: 241) | Type: Tech

Tesla's Terafab is set to become one of the most ambitious semiconductor projects ever built.

Designed to produce over 1 terawatt of AI compute capacity per year, Terafab will power the future of AI, Optimus, autonomous vehicles, and beyond. https://t.co/KiIUaCi3sA

See 8 related tweets

@StockMKTNewz: Elon Musk just posted this:

“My very approximate guess is that Terafab AI compute output would be ~...

@niccruzpatane: Wow. Elon Musk says Terafab’s AI compute would be ~25% for Tesla Optimus robots.

25% of Terafab’s ...

@DeItaone: MUSK: MY VERY APPROXIMATE GUESS IS THAT TERAFAB AI COMPUTE OUTPUT WOULD BE ABOUT 25% FOR TESLA OPTIM...
@rohanpaul_ai: SpaceX has earlier defined Terafab’s long-term target as 1TW of compute hardware produced each year....
@PolymarketMoney: JUST IN: Elon Musk reveals 75% of Terafab’s AI compute will power AI spacecraft while 25% goes to Te...

15. KristinaParts (Group Score: 158.0 | Individual: 37.7)

Cluster: 6 tweets | Engagement: 58 (Avg: 34) | Type: Tech

NEW: Months after Nvidia paid ~ $20B for Groq's inference tech, AMD is making its own move: acquiring Taalas, a Toronto startup that etches AI models directly into silicon. The deal deepens$ AMD's push into the AI inference market. Terms weren't disclosed. $NVDA\n\nQT @taalas_inc: We are pleased to share that Taalas has agreed to join AMD.

We built Taalas to rethink AI inference from the ground up: hardware designed around the model, rather than the other way around. The result is the world's fastest and most cost-effective inference silicon.

Joining AMD gives us the scale, engineering resources, and global reach to bring that work to the world – and to accelerate what comes after it.

We are also excited to build on AMD's long-standing presence in Canada and its continued commitment to the country's AI ecosystem. Many of us grew up at AMD Canada, and are looking forward to coming home.

We are proud of what this team has built, and even more excited about what comes next.

https://t.co/b6VqdwJl8S

See 5 related tweets

@benitoz: AMD is buying Taalas, a startup that hardwires AI models into custom silicon

Yesterday on @theinfor...

@AMD: We're excited to announce our planned acquisition of @taalas_inc, bringing differentiated AI inferen...
@wallstengine: $AMD TO ACQUIRE AI CHIP STARTUP TAALAS

AMD signed a definitive agreement to acquire Toronto-based T...

@StockSavvyShay: $AMD agrees to acquire Taalas which it calls a “pioneer in specialized AI inference silicon.”

The d...

@StockMKTNewz: $AMD just announced it has agreed to acquire Taalas

"a pioneer in specialized AI inference silicon"...

16. kimmonismus (Group Score: 157.1 | Individual: 35.9)

Cluster: 5 tweets | Engagement: 139 (Avg: 715) | Type: Tech

Wispr Flow's entire growth engine was run by one person and a stack of agents.

That person is Matt Swulinski, and today Fryderyk Wiatrowski announced him as Head of Growth at @viktor_com.

He arrived at Wispr with no acquisition machine to speak of and left it growing 50% month over month, at a company that has raised more than $81M. For a stretch he was the whole marketing department, so he built himself one: markdown files, skills, version control, roughly 100 newsletter sponsorships a month running through it. That gave his team capacity it never had the hands for.

Now he is at the company selling AI employees, after two years of quietly being his own proof that they work. When the practitioner picks the product, that is about as strong a signal as a category gets.\n\nQT @frydwia: Announcing @MattSwulinski as Head of Growth at @viktor_com.

Previously Head of Growth at @WisprFlow. Before that, @Superhuman.

One of the most AI-native growth operators I’ve met.

Very excited to build together. https://t.co/B7006cma75

See 4 related tweets

@svpino: I love Wispr Flow. It's what I use to dictate to my computer.

Matt Swulinski, the person who built ...

@aakashgupta: One marketer at Wispr Flow ran roughly 100 newsletter sponsorships a month, mostly alone, on an agen...
@frydwia: Announcing @MattSwulinski as Head of Growth at @viktor_com.

Previously Head of Growth at @WisprFlow...

@Scobleizer: Wow.

Matt is one of the purest AI-native growth operators out there. He ran almost an entire depart...

17. WesRoth (Group Score: 154.7 | Individual: 36.8)

Cluster: 6 tweets | Engagement: 26 (Avg: 30) | Type: Tech

Jeff Dean is leaving Google after nearly 27 years to launch Discovery Loop with Sanjay Ghemawat, Quoc Le, and Oriol Vinyals.

Four researchers and engineers who helped build much of Google’s modern computing and AI foundation.

Google tried to retain the team but will remain closely involved as a founding investor and cloud partner, providing computing resources during Discovery Loop’s first year.

Funding also comes from Khosla Ventures, Radical Ventures, Kleiner Perkins, Lightspeed, and Doerr Capital.\n\nQT @JeffDean: Announcing Discovery Loop!

I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.

♾

Learn more at: https://t.co/Rv3LMdLluK

See 5 related tweets

@WesRoth: Demis Hassabis says AGI is “close at hand”—and is stepping away from Google DeepMind’s day-to-day op...
@rohanpaul_ai: RT @rohanpaul_ai: Google's chief scientist is leaving after 27 years to build AI that can run its ow...
@jamessinka: RT @jamessinka: The Demis & Jeff's news is extremely positive for AI in science.

@demishassabis (...

@ankitxg: Wow.. end of an era!\n\nQT @JeffDean: Tomorrow will be my last day at Google after 27 years, and wat...
@WesRoth: Oriol Vinyals is leaving Google DeepMind to co-found Discovery Loop after helping shape several of t...

18. SawyerMerritt (Group Score: 149.4 | Individual: 34.0)

Cluster: 5 tweets | Engagement: 1837 (Avg: 1421) | Type: Tech

Terafab Texas definitely has the most sci-fi/futuristic factory design ever. Looks so cool. https://t.co/V9Te3ckHuQ\n\nQT @Tesla: Terafab will be built in Grimes County, Texas

In April, we broke ground on our research fab on the North Campus of Giga Texas – the precursor to Terafab.

Both Tesla & SpaceX will need far more chips than current & future global production can supply.

This is why we're building the largest chip manufacturing facility ever, with the goal of producing over 1 terawatt of compute per year

The future is built in Texas

https://t.co/RQPNlAasdP

See 4 related tweets

@WOLF_Financial: Tesla $TSLA just posted this

“Terafab will be built in Grimes County, Texas”\n\nQT @Tesla: Terafab...

@ApoStructura: This feels like something that will be built in the 2100\n\nQT @Tesla: Terafab will be built in Grim...
@tugot17: the greatest love story that never was 🥺\n\nQT @Tesla: Terafab will be built in Grimes County, Texas...
@KyeGomezB: Make it bigger\n\nQT @Tesla: Terafab will be built in Grimes County, Texas

In April, we broke groun...

19. VaibhavSisinty (Group Score: 146.2 | Individual: 34.0)

Cluster: 7 tweets | Engagement: 57 (Avg: 104) | Type: Tech

Microsoft's AI revenue looks incredible until you realize 70% of it is coming from a company they put $130 billion into. They're not selling AI. They're billing their own bet.

Microsoft disclosed $24.1 billion in sales from OpenAI last fiscal year. Their total AI business is roughly$ 34 billion. One partner is carrying the entire story.

Here's how it works. OpenAI runs every model on Azure. Every ChatGPT query, every API call, every training run, all on Microsoft's servers.

Then OpenAI shares a cut of its own revenue with Microsoft on top of that.

So Microsoft funds OpenAI. OpenAI uses that money to pay Microsoft for compute. And that payment becomes 70% of Microsoft's AI revenue. The money is going in a circle.

And the customer driving all of this is still not profitable.

In April they restructured the deal. Reduced exclusivity so OpenAI can use other clouds. Capped revenue sharing through 2030. Microsoft's upside now has a ceiling even if OpenAI keeps growing.

Meanwhile Microsoft is quietly hedging. Backed Anthropic. Built Muse Spark. Launched Muse Code today. All moves to reduce the dependency on one partner.

But right now, remove OpenAI from the balance sheet and Microsoft's AI story shrinks by 70% overnight.

The biggest AI narrative in the world is built on a loop, not a moat.

See 6 related tweets

@rohanpaul_ai: Microsoft just disclosed for the first time taht OpenAI supplied ~70% of Microsoft's AI revenue, per...
@Pirat_Nation: Microsoft has revealed that around 70% of its AI revenue comes from a single customer, OpenAI.

Acco...

@Cointelegraph: 🔥 BIG: OpenAI accounts for likely 70% of Microsoft's actual AI sales, with the software giant record...
@edzitron: RT @edzitron: Analysts estimate that 70%+ of Microsoft, Google and Amazon's present and future AI re...
@edzitron: RT @edzitron: News: Microsoft disclosures and Bloomberg analyses suggest that OpenAI's $24.1 billion...

20. ns123abc (Group Score: 142.5 | Individual: 26.3)

Cluster: 8 tweets | Engagement: 2647 (Avg: 950) | Type: Tech

🚨 DeepSeek plans a SIGNIFICANT INCREASE in API prices soon

>“please plan your usage accordingly”

it’s over for the poors https://t.co/HK3vTPW0QI

See 7 related tweets

@jukan05: DeepSeek: Plans to raise overall API pricing in the near future, with a significant increase expecte...
@Hesamation: Goodbye invisible pricing charts. https://t.co/OSrfqWFZ57\n\nQT @ns123abc: 🚨 DeepSeek plans a SIGNIF...
@ns123abc: Now it all makes sense… https://t.co/IjUAr5et4m\n\nQT @ns123abc: 🚨 DeepSeek plans a SIGNIFICANT INCR...
@ohryansbelt: The Chinese discovered blitzscaling\n\nQT @ns123abc: 🚨 DeepSeek plans a SIGNIFICANT INCREASE in API ...
@Cointelegraph: 🚨 NOW: China's DeepSeek announces plans to significantly raise pricing for its API services soon, th...