Exclusive: I have been told directly that SSI is not launching this month, Gavin is fantastic and incredible source nonetheless - it happened to be a small mistake in a long podcast.

JUST IN - For the first time, AI has designed complete viral genomes, producing 16 functional viruses that infect bacteria and "pose no threat to people." — BBC

We’re making better intelligence easier to access in ChatGPT for everyone: - GPT-5.6 Sol now powers both Instant and deep reasoning for Plus & Pro users, delivering more factual, focused responses. - Free & Go users get unlimited text chats with GPT-5.6 Luna starting tomorrow.

Holy: Leo is one of the best leakers. He says that OpenAI's Astra (GPT-6) will supposedly be released as early as next week. Astra, aka GPT-6, is the new foundation model with completely new pretraining. Internally codenamed "Mewfour," it's already been tested. Rumors have long suggested that GPT-6 would be released relatively soon after version 5.6, at the end of July or beginning of August. The main goal is to have a model that consistently outperforms Anthropic's Fable 5. I can't get any more excited!

JACK JUST ORDERED 2,000 HUMANOID ROBOTS. SOFT FABRIC WAS SUPPOSED TO BE THE ONE THING ROBOTS COULDN’T HANDLE.🤖👕 Sewing has always been the automation-proof job. Cloth wrinkles, shifts, deforms — machines are built for rigid parts, not soft ones. Now watch the clip. Alignment deviations for cut pieces like collars and pockets have to stay within ±2mm. Cutting and sewing motions demand 0.3-0.5mm precision — tighter tolerances than most metal assembly work. Robots on this line are already hitting 98% fabric-separation accuracy, running 3-second operations, and reaching 80% of human efficiency. They’re not alone at the machines. Real workers sit right beside them, running the same Juki and Jack stations, same shift. A Jack executive was upfront: the robots can currently perform individual sewing operations, not run a complete production line independently. This is task automation, not a fully autonomous tailor. But it’s real, and it’s running on an active factory floor. Jack Technology has now ordered 2,000 of these units for apparel manufacturing — aiming to lift manufacturing efficiency by at least 30% across the line. Ten years ago, “robot-proof job” was the whole pitch for garment work. That pitch just expired. Would a 2,000-robot rollout change where your clothes get made?

Just a quick reminder: There is no universally accepted definition of superintelligence, just as there isn't one for AGI. This is problematic because the terms are used in different ways. Dario Amodei writes about superintelligence in his blog, but elsewhere it's understood more as general intelligence. Therefore, I continue to advocate for a unified definition. And that means we need to examine precisely what Ilya Sutskever means by superintelligence.

Ilya Sutskever’s company is going to release a model in August. I took another look at their website. They state it very clearly: their only goal, their only mission, is to develop superintelligence. That is their sole focus. In that sense, this model release will be far more than “just” another model release. In some form, they must have achieved superintelligence if we take their statement seriously. "SSI [Safe Superintelligence] is our mission, our name, and our entire product roadmap, because it is our sole focus. Our team, investors, and business model are all aligned to achieve SSI."

Figure's secret weapon is vertical integration so deep that almost no one else on Earth can catch them. "I don't think there's any group in the world that designs more parts than we do on the robot." "We design the motors, basically every part within there, the rotor, stator, everything." He self-funded the whole company upfront, hit $1M/month burn in 4 months, and built a 40-person team in under 5 months working 100 hours a week. FT @MollySOShea @sourceryy @adcock_brett.

As I was saying …

AI is a supersonic tsunami

🔥Let's talk about Qwen! #AMA

🚨 ¡ESTO VA A ROMPER INTERNET! Elon Musk acaba de soltar la bomba más loca de Neuralink: “En los próximos 6 a 12 meses vamos a poner los primeros implantes de visión. Aunque estés 100% ciego de nacimiento, vamos a escribir directamente en tu corteza visual… y vas a ver.” “Y eso es solo el comienzo. A largo plazo tendrás resolución ultra HD y superpoderes reales: verás en infrarrojo, ultravioleta y hasta radar. Literalmente como un superhéroe.” El futuro ya no es ciencia ficción. Es Neuralink.

Google 发布了第八代 TPU 最核心的变化：以前一颗芯片什么活都干，这次拆成了两颗，各管一头。 - TPU 8t — 专门用来训练模型 - TPU 8i — 专门用来推理，跑模型 为什么要拆？ 因为这两件事要的东西已经彻底不一样了： 训练要的是"能一次吞下海量数据"，慢一点无所谓； 而跑模型要的是"反应快"——你问一句它得立刻答上来。 现在 AI 进入了Agent时代（AI 自己一步步干活、反复思考），一个任务要来回想很多轮，每轮慢一点点，加起来用户就等不了了。 TPU 8t（训练）：定位为：万亿参数级大模型预训练， 一组 9600 颗芯片，121 exaflops 算力、2 PB 共享内存， 峰值性能是上一代的 3 倍 TPU 8i（推理）： 后训练 + 实时推理，通信密集任务延迟最多少一半

Today's Superintelligence is a guest essay by @ccatalini - Founder of the MIT Cryptoeconomics Lab. His argument: almost every moat you think you have is a lease. Routing. Distribution. Marketplace liquidity. Your app-layer feature set. All of it gets weaker every single time the models improve. Exactly one kind of network effect gets stronger instead. He calls it verification-grade, and it comes down to a question most founders never ask: who owns the measurement nobody else has? Highly recommend you check out his article, link down below (Photo Christian Catalini):

DeepSeek-V4-Flash 正式版发布：可原生接入 Codex 九项 Agent 测试全部反超自家更大的 V4-Pro 预览版 登顶开源模型前三 输出比 Opus 4.8 便宜 89 倍 DeepSeek 针对 Codex 还专门做了官方适配：连 Codex 自己的系统提示词都一并配好了 接入教程：一条命令配好，原有设置不动 详细内容：best.xiaohu.ai/article/deeps…

SpaceX is hiring exceptional engineering & skilled trades talent to build & operate the most powerful AI supercomputer clusters on & off Earth. If interested, please email 3 bullet points demonstrating evidence of exceptional ability & your resume to datacenters@spacex.com.

Seedance 2.5 京剧表演 更多介绍和案例：best.xiaohu.ai/article/seeda…