NVIDIA writes a ~$5B check into Ilya Sutskever's superintelligence bet

Good morning 👋 Ilya Sutskever left OpenAI in 2024 with no product and no revenue, just a bet on scaling. NVIDIA just put roughly $5B behind that bet.

In today's issue:

• 🔭 NVIDIA writes a ~$5B check into Ilya Sutskever's superintelligence bet
• 🧠 A 2.8T-parameter open model just moved the goalposts
• 🔬 Catching hallucinations in one forward pass, plus two agent-safety papers
• 🛠️ Chunk your RAG by task, not by document
• 🔗 Data centers, chip diplomacy, and Claude going enterprise

🔭 THE ONE THING

💰 NVIDIA writes a ~$5B check into Ilya Sutskever's superintelligence bet

NVIDIA and Safe Superintelligence announced a long-term strategic partnership Monday. NVIDIA's own release stays vague ("substantial"), but Bloomberg and Reuters both put a number on it: roughly $5B, valuing SSI at $32B. The real prize for Sutskever is access to NVIDIA's next-gen Vera Rubin platform, which the two companies say expands SSI's compute by an order of magnitude. He left OpenAI in 2024 with no product, no revenue, and a bet that scaling alone gets you to superintelligence; NVIDIA just backed that bet with the one resource it can't get anywhere else in this volume. Read the other side too: a multi-billion-dollar stake that flows back into GPU purchases is also NVIDIA financing its own demand, the same structure it's run with OpenAI and xAI.

🧠 MODELS & RELEASES

• 🌙 Moonshot AI open-sourced Kimi K3, a 2.8T-parameter MoE model with 104B active params, native vision, and a 1M-token context window. Built on a new Kimi Delta Attention scheme, it's the kind of open weight that makes the frontier a moving target for closed labs. arxiv.org/abs/2607.24653

🔬 RESEARCH HIGHLIGHTS

• D-Score flags hallucinations from a single forward pass, no retrieval, no second generation needed. The method counts how many singular directions in the hidden-activation matrix stay nearly as strong as the top one; conflicting internal evidence spreads the signal across more of them, and that spread is the tell. Cheap enough to run inline. The catch: the paper tests it on FAVA-Annotation and RAGTruth and doesn't publish a hit rate, so treat it as a promising mechanism, not a shipped detector yet.

• APPA goes after a real agent tax: read one untrusted document under standard taint tracking and your whole context gets contaminated, utility included. This permissions algebra spins off a labeled child process to absorb the taint, then a sanitizer hands clean results back to the parent. Across four models, exfiltration success dropped from 31-50% to 0-7%, and three of the four recovered most of the utility strict tainting used to cost. Relevant if you're building agents that touch tool output you don't fully trust. Read the paper.

• Generate-test-revise loops, the default pattern in coding agents, don't hold up the way builders assume. A sealed HumanEval study found correctness on the live patch falling from 82.0% after one revision to 67.3% after two, even as the "ever found a correct fix" rate climbed to 84.7%. The agent stumbles onto the right answer, then talks itself out of it on the next pass. The proposed fix isn't more iteration, it's binding verifier evidence to the exact code state so a good checkpoint can't get silently overwritten. Worth a look (arXiv) if you're shipping an autonomous repair loop.

🛠️ TRY THIS

Stop chunking your RAG pipeline by document. Chunk it by task.

AWS's write-up on task-aware knowledge compression (TAKC) makes the case plainly: flat RAG chokes once your knowledge base spans hundreds of documents and every query gets the same treatment. Bigger context windows don't fix that. Task-aware compression does.

1. Group your knowledge base by task type (lookup, comparison, troubleshooting, summarization), not by source folder.

2. Pre-compress each group into its own representation before query time. Don't rebuild it fresh on every call.

3. Cache two or three fidelity tiers per group: a thin summary, a mid-detail pass, the full source.

4. Route each query to the cheapest tier that answers it correctly. Escalate only when the model flags missing context.

Prompt: List the top 5 query patterns hitting our knowledge base this month, grouped by task type. For each, note the minimum fidelity tier needed to answer correctly, and flag any where flat RAG is over-fetching context we don't need.

Worth a look

• opensquilla A coding agent tuned for token efficiency: same context budget, denser reasoning per call. Worth a look if your agent spend is creeping up. GitHub
• T3MP3ST elder-plinius's autonomous red-teaming harness, multiple agents coordinating an attack on a target. Built for authorized offensive-security work, not something you point at a system you don't own. GitHub

🔗 QUICK LINKS

• Meta / BlackRock $14B joint venture, 1GW, El Paso. BlackRock takes 80%, Meta keeps 20% and signs a lease that can run two decades. Online in 2028. Meta newsroom
• Hut 8 A second 352MW lease at Beacon Point pushes the Texas campus to $19.6B in contracted value, $50.2B if every renewal option gets exercised. The tenant's identity is still the one thing nobody will say out loud. PR Newswire
• Anthropic Dario Amodei says the company has never wanted a ban on open-weights models. What it wants instead: chip export controls and mandatory safety testing on anything capable enough to matter, regardless of who releases it. Anthropic
• Cognizant 30,000+ associates now Claude-certified, and Cognizant just became a Global Premier Partner. One live number: contract review time down 40% at a biopharma client. Anthropic
• Meta AI Segment Anything and DINO are the vision stack behind Pitt's assistive robots, an ARPA-H-funded project chasing the 100,000 wheelchair-related ER visits the US racks up every year. Meta AI
• Guardoc Health Amazon Nova models are now writing up clinical documentation for long-term care facilities, via Bedrock. AWS

Keep an eye on how fast Vera Rubin actually ships. That's where this story gets tested.

Pradeep Perugu