And the next upcoming model for coming week: MiniMax M3.1. Already testing m3.1-preview. Attached the Graph for M3, it was a very very solid model. So we should keep an eye on m3.1!

Because it's becoming difficult to keep track. Currently, it's estimated that OpenAI has reported over two dozen incidents (via Reuters). However, OpenAI still doesn’t know the full extent of its agents’ unauthorized activity, according to Reuters. tl;dr -53 ChatGPT user images leaked. Most have been removed; OpenAI hasn’t disclosed whether they depicted real people or were AI-generated. -Roughly two dozen incidents had been identified by mid-September, with more emerging. The review will take months. -Government websites were also targeted. Researchers reported an attempted breach of a US Education Department site; OpenAI says access to SEC and Census sites showed no evidence of a security breach. -Oversight and transparency are under scrutiny. Some incidents went unnoticed for months. Sources describe a tightly controlled, lawyer-shaped investigation; OpenAI denies lawyers discouraged deeper inquiries. i am all on optimistic for AI. but its getting serious, ngl.

OpenAI updates on upcoming products: - Their always-on agent will be called "o" (essentially a GrokBot alternative) - It will be available in all Pro tiers - Fast Mode confirmed, presumably included in the $500 sub-plan. (Cerebras chip takes effect) - shared message board for agents also incoming

More details on the incident behind OpenAI’s pause: a researcher acknowledged the alert within 3 minutes, but the training run was only stopped manually 2.5 hours later. OpenAI says the automatic shutdown did not work as expected. The model had reached an external chatbot through a gap in DNS filtering. A separate detector for unusual DNS activity did not cover the affected environment. A retrospective review also found other external DNS requests that monitoring had failed to flag at the expected severity. In some cases, it treated an unhelpful response as evidence that internet access had failed. Here is what else happened: - New research into July’s Hugging Face hack documents internal Slack searches, credential collection and programs designed to maintain access to compromised servers. Agents also tried querying Claude, DeepSeek, Kimi and Qwen. - In May, another model published a researcher’s GitHub token while trying to obtain another team’s mathematical proof. It split the token to evade secret scanning, despite twice being told to solve the problem itself. - Reuters reports that agents leaked 53 ChatGPT user images online. OpenAI expects its broader investigation to take months. This is getting serious.

We've now received GPT-6 Astra and Opus 5.5. Confirmed for the next releases are: - Sonnet 5.5 - Haiku 5.5 - Kimi K3.1 (they themselves hinted at it in a cryptic post) - Gemini 4, before the end of the year Unconfirmed but very likely: - Fable 5.5 - Astra update (at the DevDay?) Currently, there's no word yet on new GLM or DeepSeek models.

I love how Mark Zuckerberg simply laughs at the question of whether "AI will wipe us all out," as if it were the most absurd question in the world. At this point, no one is going to slow down.

I am sure that power supply, grid connections, and data center construction will be among AI’s biggest bottlenecks in the coming years. Google’s Project Suncatcher explores an interesting possibility: moving AI compute into orbit, with TPU-equipped satellites connected by lasers. Google says solar panels in the right orbit could be up to 8x more productive than on Earth, thanks to near-continuous sunlight. It sounds crazy, but the potential is enormous: Scaling compute without waiting for terrestrial grid connections would be a major advantage. Cooling, launch costs, and reliable connections between satellites still need to be solved. The first step is a prototype developed with Planet, scheduled to launch aboard SpaceX’s Transporter-18 mission. SpaceX could end up being one of the biggest winners here. Making orbital compute economical would require affordable, frequent launches. That could turn the AI infrastructure boom into a major new source of demand for SpaceX.

Damn, im in love with Opus 5.5. Its the best model ive ever used, has such a good taste, is comparatively quickly, smart and finally precise instead of verbose. As good as Opus 5.5 is, I'm incredibly excited for Fable 5.5. After switching completely to Codex, I now have to go back to Anthropic. oh and btw. its crazy how fast the vibe has changed. Now all I read on X are posts about how great Opus tastes and how OpenAI absolutely has to catch up. One day that changed everything.

Xiaomi's MiMo-V2.6-Pro has really impressed me. It now leads open-weight models on the Artificial Analysis Intelligence Index. 46 points, just behind GPT-5.6 Sol (max) at 47 in AA Benchmarks. The cost gap is much larger: $0.13 versus $1.99 per Intelligence Index task, on a weighted-average basis. Pro handles text, images, video and audio, with a 1M-token context window. Xiaomi has kept API pricing unchanged from V2.5. The model weights are available under MIT. Xiaomi is also releasing the training environments and RL code alongside the technical report, giving researchers more of the work behind the model to examine. This is literally the definition of open AI. Congrats on that release!

Anthropic reportedly plans to give its seven co-founders majority voting control ahead of its IPO, which gives them much more power. The Information reports that a proposed Palantir-style share structure would give them 50.1% of voting power on most corporate matters, as long as at least three founders retain a minimum number of shares. Dario Amodei reportedly owns around 2% of Anthropic. The new shares would add voting power without additional economic ownership.

Meta nailed it. Agentic AI, hardware, accessibility, and distribution: I think they’re on their way to becoming number one. A few thoughts. Meta’s keynote was yesterday, and I was incredibly impressed. To me, Muse already looks like a major success. It has reached number one in the US iOS App Store, and I’ve been hearing plenty of enthusiasm about the Mac app. Several things make Meta stand out right now. I was there at the invitation of Meta and part of a very select group of creators, with whom I later had the opportunity to meet Mark Zuckerberg. It's still completely crazy for me to have experienced that. But back to keynote: 1. They seem to understand how to make agentic AI accessible. They’re using what I see as their biggest moat, distribution, and making their products approachable enough for the average person to enjoy using them. Muse is designed around familiar interactions, including messaging in its app or through WhatsApp. That’s no small thing. To truly reach the consumer market, you have to avoid overwhelming people. And I think Meta is getting that right. At the same time, its Meta Spark models are pretty good, too. No, they’re not as good as Fable or GPT-6 Astra, but they’re more than good enough for even the most complex tasks. Meta doesn’t need to be at the very cutting edge of frontier models. Its models just need to be good enough for the people it wants to reach. 2. They unveiled hardware yesterday that genuinely surprised me. There was an Apple-style “one more thing” moment, and the new VR glasses clearly invite comparison with Vision Pro. Meta says they will launch in spring 2027 for $1,299.99 and weigh about 100 grams on your face. The separate computing and battery puck is additional. That is still completely crazy. People who tried them told me they preferred them to Apple Vision Pro. Those are personal impressions, though, rather than proof of better visual quality across the board. 3. Then there’s the adorable Jolly, the avatar associated with Muse. Pair that personality with the promised agent integration and multiple virtual screens, and I can easily see this becoming a huge hit next year. As if that weren’t enough, Meta also previewed Muse Charm, a cute, portable AI gadget for talking to and interacting with Muse. Meta says more details will follow later this year, so this is an announcement rather than a confirmed retail launch. As if that weren’t enough, they also released a cute AI gadget: a pendant you can carry with you as a wearable and use to delegate tasks. That’s a first-mover move that puts them ahead of even OpenAI. Meta is no longer just following the competition. It’s now moving ahead of it. Of course, it’s too early to say whether this gadget will actually prove useful. But I see it as another sign that Meta is moving from follower to frontrunner. One thing is certain for me: I’m getting the new VR glasses. There’s absolutely no doubt in my mind. If typing works as well as promised, the connection to Muse is seamless, and battery life is good enough in everyday use, this will be an absolute game changer. I think Meta has officially joined the big leagues. p.s. Thank you @Meta for providing me a pair of your Meta Glasses! :)

Xiaomi's MiMo-V2.6-Pro has really impressed me. It now leads open-weight models on the Artificial Analysis Intelligence Index. 46 points, just behind GPT-5.6 Sol (max) at 47 in AA Benchmarks. The cost gap is much larger: $0.13 versus $1.99 per Intelligence Index task, on a weighted-average basis. Pro handles text, images, video and audio, with a 1M-token context window. Xiaomi has kept API pricing unchanged from V2.5. The model weights are available under MIT. Xiaomi is also releasing the training environments and RL code alongside the technical report, giving researchers more of the work behind the model to examine. This is literally the definition of open AI. Congrats on that release!

This Meta VR headset is a game changer. 100g, better quality than the Apple Vision Pro, 50% cheaper than Apple. An instant buy for me. Crazy how they've outplayed Apple, the best hardware company.

Xiaomi's MiMo-V2.6-Pro has really impressed me. It now leads open-weight models on the Artificial Analysis Intelligence Index. 46 points, just behind GPT-5.6 Sol (max) at 47 in AA Benchmarks. The cost gap is much larger: $0.13 versus $1.99 per Intelligence Index task, on a weighted-average basis. Pro handles text, images, video and audio, with a 1M-token context window. Xiaomi has kept API pricing unchanged from V2.5. The model weights are available under MIT. Xiaomi is also releasing the training environments and RL code alongside the technical report, giving researchers more of the work behind the model to examine. This is literally the definition of open AI. Congrats on that release!