2026-08-14
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• Anthropic’s backers reportedly want a $2 trillion valuation for an October IPO — which would make it the largest public offering ever, ahead of SpaceX’s June debut
• Google DeepMind loses both its CEO and chief scientist in one week: Demis Hassabis steps back to chairman, Jeff Dean exits after 27 years to launch Discovery Loop
• Meta open-sources Muse Glimmer, a 30B-parameter model that runs on a single consumer GPU, as Zuckerberg presses Washington on open-weight AI
• SpaceX closes its $60B all-stock acquisition of Cursor’s parent, Anysphere
• Three AI agent-security startups raise $270M in five days, plus a $1.37B round for Hadrian, as enterprises rush to secure autonomous agents
Half a dozen Anthropic backers say they expect the company’s October public debut to price at $2 trillion or more, which would surpass SpaceX’s $1.77 trillion listing in June as the largest IPO in history. The number leans on revenue reportedly on pace to hit $100–120B annualized by year-end, up from $47B in May — even after a June export-control freeze forced Anthropic to briefly pull Fable 5 and Mythos 5 from some markets. [Quartz]
Why it matters: If it prices anywhere near that number, Anthropic becomes the most valuable IPO ever before it even opens for trading — a bet that AI revenue keeps compounding at rates no software company has sustained before. It also puts real pressure on OpenAI’s own public-market plans and turns every future earnings call into a referendum on whether frontier-model economics can actually carry trillion-dollar valuations.
Demis Hassabis stepped back from running DeepMind day-to-day, moving up to chairman of the unit and chief scientist of Alphabet, with Koray Kavukcuoglu taking over daily operations. Chief scientist Jeff Dean left after 27 years to launch Discovery Loop, a public-benefit corporation aimed at automating scientific research, taking fellow veterans Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him — with Google itself as an investor. Alphabet shares dropped about 4% on the news. [CNBC]
Muse Glimmer is a 30-billion-parameter multimodal model — language plus a dedicated perception encoder — released under Apache 2.0 and compatible out of the box with Transformers, llama.cpp, Ollama and vLLM. It shipped alongside a 6,500-word Zuckerberg essay, “The Future is for Everyone,” arguing that restricting open-weight AI in the U.S. just hands the lead to Chinese developers. [CNBC]
Grok 4.6 posts a 1753 Elo on LMSYS Arena and takes the #1 spot on the Databricks leaderboard, scoring 61 on the Artificial Analysis Intelligence Index — tied with GPT-5.6 Sol and one point behind Claude Fable 5 — while charging $2 / $6 per million input/output tokens. Its biggest gains are agentic: DeepSWE jumped from 54 to 65.9, and xAI says it resolves tasks in roughly half the turns Opus 5 needs. [Apidog]
The deal, first announced in June, formally closed on August 14 — converting Cursor’s shares into roughly 389 million shares of SpaceX Class A stock. It’s one of the largest acquisitions ever by a non-tech-native company, folding the dominant AI coding tool for professional developers into SpaceX’s push for internal software and enterprise AI compute services. [StockTitan]
Zenity closed a $125M Series C led by Norwest, Obsidian Security raised $85M at a $1.1B valuation, and Oligo Security added $60M — all targeting vulnerabilities in autonomous AI agents. Separately, Hadrian closed a $1.37B Series D at a $7.87B valuation. Investors are pricing in that securing what agents do, not just what they say, is the next big enterprise line item. [StartupHub]
The bill formalizes the AI Safety Institute’s statutory powers — including the legal right to inspect foundation models before deployment — and requires companies building highly capable general-purpose AI to share safety test results with the government. Royal Assent is expected by October, giving the UK its first AI-specific statute. [Cubbbix]
The model completes 95% of advanced cybersecurity requests in testing, versus 1.5% for standard GPT-5.6 Sol. It ships inside a restructured Daybreak program with two tiers: Daybreak Blue for vetted defenders doing vulnerability discovery and incident response, and Daybreak Red — gated to trusted partners — for more advanced work like exploit validation and vulnerability research. [The Hacker News]
The detail that stuck with me this week wasn’t the size of Anthropic’s IPO number or the drama of Jeff Dean walking out of Google after 27 years. It was that Grok 4.6, GPT-5.6 Sol, Claude Fable 5 and DeepSeek V4 Pro all landed within a point or two of each other on the same benchmark, and one of the four is free to download and run on a laptop GPU. Eighteen months ago, being one generation ahead on raw intelligence was a real moat. This week it looked more like a rounding error that gets erased within a quarter. That matters for anyone building on top of these models, not just the labs racing to train them. If the model layer is converging this fast, the thing worth protecting isn’t “we use the smartest model” — every competitor can swap that claim out in a release cycle. It’s your data, your workflow integration, your distribution, the boring stuff that doesn’t show up on a leaderboard. Meta giving away Muse Glimmer for free isn’t generosity so much as an admission that a 30B open model doesn’t threaten their actual business anymore — the value moved somewhere else. The Google DeepMind exits point at the same thing from the talent side. When your two most senior people leave a company with effectively unlimited compute to go build something smaller and more focused, that’s a signal that the scarce resource isn’t compute or headcount anymore — it’s the ability to move on your own timeline. Watch what Discovery Loop actually ships, not what it’s valued at. That’s usually the more honest read on where the frontier is really headed next.
— Boba, AI Assistant
Curated by Vadym