Mira Murati's lab just shipped the biggest open-weight model the US has produced

Good morning 👋 Mira Murati just did the one thing OpenAI won't: she gave away the weights.

In today's issue:

• 🔭 Ex-OpenAI CTO's lab ships the biggest open-weight model the US has produced
• 🧠 NVIDIA's robot brains, Cadence's agentic chip design, and what building a real agent actually taught Allen AI
• 🔬 What the data really says about AI coding agents (it's slower than the hype)
• 🚀 A $15M seed for a coding "autopilot," plus a $100M round for enterprise AI ops
• 🏛️ Brussels goes easy on X, and OpenAI's plan to write AI law from both directions at once
• 🛠️ How to stress-test a 975B-parameter model before it touches production
• 🔗 xAI's CSAM lawsuit, and Anthropic's Canadian research bet

🔭 THE ONE THING

🔓 Mira Murati's lab just shipped the biggest open-weight model the US has produced

Thinking Machines, the lab Murati built after leaving OpenAI as CTO, released Inkling on July 15: 975 billion parameters total, 41 billion active via mixture-of-experts, trained on 45 trillion tokens of text, image, audio, and video, with a context window up to 1M tokens. Full weights are live on Hugging Face under Apache 2.0, no waitlist, alongside a smaller preview called Inkling-Small (276B total, 12B active) that the lab says performs competitively despite the much lighter active load. Treat that claim as a vendor benchmark until someone outside the lab reproduces it. Still, at 975B parameters this is the largest open-weight model any US lab has shipped, and giving away the weights instead of gating an API is a direct rebuke of the approach Murati's old employer takes. For builders it's a frontier-class model you can fine-tune and own, though running the full-size checkpoint means you need the kind of GPU cluster only a handful of companies have lying around.

🧠 MODELS & RELEASES

• 🤖 NVIDIA shipped the T3000 and T2000, two Blackwell-powered Jetson Thor modules built for edge robotics. The T3000 hits 865 FP4 teraflops at roughly half the size and power draw of the T5000; the T2000 is the stripped-down 400-teraflop option for cheaper hardware. Boston Dynamics, Amazon Robotics and FANUC are already building on it, though real availability doesn't land until Q1 2027. link

• ⚙️ Cadence rolled out AuraStack, an agentic layer that treats PCB layout, thermal, signal integrity and mechanical validation as one shared design context instead of four separate handoffs. The company claims 2x faster time to market and 15x higher productivity. Self-reported numbers, so read them as a ceiling, not a guarantee. link

• 🧩 Sakana AI is wiring NVIDIA's Nemotron models into Fugu, its multi-agent orchestrator, using Nemotron's coding and tool-calling strengths as specialized agents behind a single API. It's a bet that orchestration beats any one frontier model, and NVIDIA gets a live agentic testbed for Nemotron out of the deal. link

• 🚢 Allen AI wrote up what building Shippy, its maritime-monitoring agent now used in over 70 countries, actually taught them. A hand-built CLI beat letting the model construct API calls raw, and standard benchmarks were useless, so they built their own eval framework with subject-matter experts grading live data instead of static test sets. link

🔬 RESEARCH HIGHLIGHTS

• GitHub's agentic PR data turns out less dramatic than the hype suggests. Across 25,264 agent-submitted PRs in 2,361 popular repos, the median project sees only one or two a quarter, and most stay under an industry benchmark of 36 PRs per contributor over three months. Almost every one still runs through a single human reviewer before merging. Agents aren't replacing dev teams yet. They're on a slow, supervised trial period. link

• Optimized agents don't stay optimized. A Terminal-Bench 2.0 study pitted three agent-tuning methods against a fresh round of unseen tasks. GEPA's gains (66% pass rate after tuning) collapsed below the 58.7% unoptimized baseline the moment new tasks showed up. Meta Harness held its 64.6% but stopped improving. Only RELAI-VCL, which builds in regression control, kept climbing, to 76.4%, and kept the gains on retrain. Next time a vendor quotes you a benchmark score for their agent framework, ask what happens on round two. link

• Pen-testing needs a new failure mode. A proposed framework argues exploit-hunting misses how AI systems actually break: an agent can leak data or take a bad action with zero vulnerabilities exploited, just a poisoned tool call or prompt injection nudging its behavior off-objective. The six-step process tests for that kind of behavioral violation directly instead of waiting for a CVE. Worth stealing if you're red-teaming anything with real tool access. link

🚀 AI STARTUPS

• Vorflux Ex-Rippling CTO Prasanna Sankar closed a $15M seed for what he's calling an "autopilot" for software engineering, not another copilot: it plans, writes, tests, and reviews code across models on its own rather than waiting on you to approve every diff. Peak XV led, with YC and a bench of angels (Parker Conrad, Immad Akhund, Balaji Srinivasan) in behind it. link
• Whale The Singapore "AI Operating System" startup bolted on another $40M, taking its Series C to $100M total. CMB International and SMBC's venture arm led; Hyundai and Singtel's Innov8 are in the mix too. Whale already runs ops software across 1,600+ enterprises in 45 countries, so this is scaling, not searching for product-market fit. link

🏛️ POLICY & REGULATION

• European Commission Brussels accepted X's action plan to fix its Digital Services Act violations, seven months after fining the platform for the same law. The deal: better ad-transparency tools (faster search, API access to the ad repository) and a real process for researchers requesting public data, not the current runaround. X gets six months to deliver and has to submit to an independent audit. The EU's own advisory board called the audit piece "partially adequate." The Commission signed off anyway, promising to keep watching. Commission accepts X's DSA action plan

• OpenAI is pushing what its global affairs chief Chris Lehane calls "reverse federalism": back state AI safety bills in California and New York now, use them to build a de facto national standard, then ask Congress to pass one federal frontier-AI law and preempt every state law under it. Support the patchwork today, erase it tomorrow. Convenient timing if you're the one lobbying both layers. OpenAI: advancing AI safety through state and federal action

🛠️ TRY THIS

Stress-test the new 975B-param open-weight model before you touch production

1. Pull the weights, or the hosted endpoint if self-hosting 975B params isn't your week, and read the license terms yourself instead of trusting the headline.

2. Run it against your own eval set, not a public benchmark. Your prompts, your edge cases.

3. Price out tokens and latency against whatever you're running now. A model this size changes your inference math even when quality holds steady.

4. If it wins, swap your lowest-risk endpoint first and watch error rates for a week before going wider.

Prompt: Compare these two model outputs on [task] for accuracy and tone, score 1-10 each. Flag any response that contradicts our documented API behavior. Output as a table.

Worth a look

• xAI open-sourced Grok Build, its terminal coding agent, days after researchers showed it was silently uploading entire git histories, secrets included, to a Google Cloud bucket xAI controlled: roughly 5.1GB per session against about 192KB of actual task content. The 844,530 lines of Rust are Apache 2.0 now, but the upload code is still sitting in the binary. xAI flipped a server-side switch to turn it off, the same switch it can flip back on without shipping an update. Read the source before you point it at a real repo. GitHub
• AWS published a pattern for wiring a Computer Vision MCP server into Bedrock agents, standardizing how a model calls out for image classification or object detection instead of hand-rolling that integration every time. Worth a skim if you're already on Bedrock and sick of writing bespoke vision glue. AWS

🔗 QUICK LINKS

• xAI sued a Grok user in a Texas federal court, alleging he fed the chatbot ordinary photos of minors and prompted his way past its safeguards to generate CSAM. The company says it filed over 73,000 NCMEC reports this year alone. Case docket
• Anthropic is putting $10M CAD behind eight Canadian research shops (Mila, Vector, Amii, CHEO among them), with startup API credits attached. Anthropic

That's the one worth acting on today. See you tomorrow.

Pradeep Perugu