Morning brief

Thursday, July 16

1 people · 0 companies · 6 topics

TL;DR Anthropic launched Ode, a $1.5B enterprise AI services firm with Blackstone and Hellman & Friedman, and confirmed IPO investor meetings are underway targeting October; Thinking Machines Lab (Mira Murati’s startup) released Inkling, its first open-weight model.

Models & research

  • model-releases Thinking Machines Lab released Inkling on July 15: a 975B-parameter / 41B-active MoE open-weight model trained on 45T tokens across text, image, and audio with a 1M-token context window. Available for download and fine-tuning via the company’s Tinker platform. Murati’s explicit bet is enterprise customizability over raw benchmark supremacy. TechCrunch · Thinking Machines

  • NVIDIA revealed Cosmos 3 Edge on July 16 during Jensen Huang’s Tokyo visit: a new real-time world model for physical AI and robotics. Alongside it, Japan’s government and major industrials (Toyota, Kawasaki, Fujitsu, Hitachi, Fanuc, Yaskawa) announced a national AI infrastructure with 13,750 Vera CPUs and 27,500 Rubin GPUs. CNBC · NVIDIA Blog

People

  • simonw posted July 15 on the xAI Grok Build privacy scandal: running the Grok CLI in a directory was uploading its entire contents — including SSH keys and password databases — to xAI’s Google Cloud buckets. xAI disabled the feature and released the codebase under Apache 2.0. simonwillison.net

Tools & dev

  • agentic-ai Anthropic, Blackstone, and Hellman & Friedman officially launched Ode on July 15: a $1.5B standalone AI implementation company that embeds small teams of Anthropic Applied AI engineers inside enterprise clients. Built on the Fractional AI acquisition (May 2026); Goldman Sachs, General Atlantic, Apollo, GIC, and Sequoia also invested. BusinessWire · TechCrunch

  • Anthropic’s bankers are scheduling meetings with prospective public-market investors, confirmed July 15 by CNBC and Bloomberg; listing is expected as early as October 2026, which would put Anthropic ahead of OpenAI to public markets. CNBC

  • Apple sued OpenAI on July 16 alleging two former engineers (Chang Liu and Tang Yew Tan) took proprietary hardware files to OpenAI; OpenAI says it has found no evidence supporting the claim. Dataconomy

  • Meta’s custom AI chip “Iris” entered production July 15: manufacturing begins in September as part of Meta’s plan to reach 14 GW of AI compute. Completed testing without major issues. HPCwire

OSS

  • PrismML/Bonsai 27B released July 15: a 1-bit/ternary quantized build of Qwen3.6-27B at 3.9 GB that runs locally on smartphones. One of the biggest HN stories of the window (~480 pts). Capable reasoning without datacenter hardware. PrismML · HN

  • ollama/ollama v0.32.1 released today (July 16): Gemma 4 tool calling and multi-turn reasoning improvements, fixed MLX model cache memory leak, interactive agent now gets current working directory as context. Point release stabilizing the v0.32.0 interactive agent. GitHub

  • comfy-org/ComfyUI v0.28.0 released July 15: SeedVR2 native image/video upscaling, int4 model support with Turing int8/int4 optimizations, new Save 3D / Save Splat / Save Point Cloud nodes, 4 CVE fixes. GitHub

Other

  • OpenAI lost a trademark dispute at the EU Court on July 15/16; details are still thin but the ruling drove 215 HN points and is expected to affect brand-protection strategy in Europe. DPA International · HN

  • OpenAI’s ad business — rolled out mid-2026 — is reportedly on pace to miss its own revenue forecast by 90%, per an analyst cited by Adweek. Adweek · HN

  • Codex started encrypting the prompts sent to sub-agents this week (408 HN pts / 240 comments); sparked debate about agentic transparency and what operators can actually audit in automated pipelines. GitHub issue · HN

  • Anthropic published new research on how Claude’s expressed values shift across model versions and non-English languages (32 HN pts, active discussion). Anthropic · HN

  • Community meme of the morning: “How to stop Claude from saying ‘load-bearing’” drove the highest HN engagement of the window (430 pts / 492 comments), triggering a wider conversation about LLM verbal tics and prompt hygiene. jola.dev · HN

  • (Polymarket: 96% yes, $11M+ volume) Polymarket’s “which company has best AI model end of July” market now puts Anthropic at 96% — up from 82% noted on July 12 — signaling the crowd sees Fable 5 as a decisive lead over GPT-5.6 Sol with no close race. Polymarket