Good morning. Today’s digest has a clear through-line: OpenAI is shedding people and products as it narrows its focus, while the open-source crowd keeps finding new ways to squeeze intelligence into smaller packages — with plenty of skepticism to go around.
OpenAI’s executive exodus continues as the company trims “side quests.” Chief Product Officer Kevin Weil, Sora creator Bill Peebles, and enterprise CTO Srinivas Narayanan are all leaving, according to reports from TechCrunch, Wired, and The Verge. The moves follow last month’s shutdown of Sora, which was reportedly burning around $1 million a day in compute, and the winding down of the OpenAI for Science initiative (Prism), which is being folded into the Codex desktop app. Peebles used his exit post to push back on the strategic direction, arguing research labs need room for exploratory work separate from the core roadmap — a pointed note as OpenAI consolidates around enterprise, coding, and a planned “everything app” ahead of an IPO.
Anthropic launched Claude Design, and Figma’s stock noticed. The new Claude Design research preview, powered by Opus 4.7, turns conversational prompts into interactive prototypes, wireframes, presentations, and marketing assets, with automatic brand system integration for teams. Figma shares reportedly started sliding right as the announcement went live, and HN commenters see it as aimed squarely at Figma and Lovable. The reception from designers is mixed — one 25-year veteran reported burning $50 in tokens on small tasks with little to show for it, and several worried that AI-generated design will accelerate the homogenization already visible across the modern web.
Ternary Bonsai claims top-tier intelligence at 1.58 bits — the LocalLLaMA crowd isn’t buying it. The release showcases an 8B model using aggressive 1.58-bit quantization, but commenters quickly pointed out that comparing full-precision 8B/9B baselines against a heavily quantized Bonsai model is somewhat unfair — a Q4 quant of those same baselines would close most of the size gap. A separate post went further, titled “Bonsai models are pure hype” and showing Bonsai-8B trailing Gemma-4-E2B badly. The more sympathetic read from the community: Bonsai is a proof of concept for extreme quantization, and the real payoff would come from applying the technique to something like Qwen 27B or a 35B+ model that actually fills consumer GPU memory.
Qwen 3.6 is being benchmarked — and hitting a nasty CUDA bug. A widely-shared GGUF benchmark post digs into quantization tradeoffs across providers, but the more useful finding is buried in the comments: CUDA 13.2 is producing gibberish output on 4-bit quants across the board, not just from one vendor. NVIDIA has confirmed a fix is coming in CUDA 13.3. Meanwhile, another user showed Qwen 3.6 35B running at Q2_K_XL with surprisingly usable results — one tester got 30 tokens/sec on 8GB of VRAM, though someone else got a prompt about PDF book processing answered with a generated gambling website. Small models, big personalities.
DeepSeek is reportedly raising outside money for the first time. Crypto Briefing reports the company is seeking at least $300 million at a $10 billion valuation, a break from its previous stance of relying solely on parent company High Flyer Capital. That valuation sits far below OpenAI’s ~$852B and Anthropic’s ~$380B, and commenters noted The Information floated similar rumors back in February, so a healthy dose of wait-and-see applies. Some argued $10B actually undersells DeepSeek given the splash R1 made earlier this year.
Google’s TPU v7x Ironwood gets benchmarked against Nvidia’s B200. A chart comparison made the rounds on LocalLLaMA, though the post itself is light on written analysis. Worth keeping an eye on as Google pushes Ironwood availability through Vertex — the hardware competition story underneath the model releases tends to get less attention than it deserves.
That’s the morning. OpenAI’s narrowing, Anthropic’s widening, and the open-source world is arguing about bits per weight — about as normal as Thursdays get right now.