Issue #161: America’s AI Manhattan Project Begins

• White House launches Genesis Mission as AI science Manhattan Project → DOE builds a national platform linking supercomputers, labs, and datasets for AI agents in biotech, fusion, and chips. Targets 20 challenges with demos in nine months. Read more here →

• SF Compute raises $40 M Series A to scale its AI compute marketplace → The round was co-led by DCVC and Wing Venture Capital, with participation from existing backers including Alt Capital. The funding will accelerate SF Compute’s growth as a liquid marketplace for GPU capacity, letting AI projects rent high-end compute flexibly and resell unused capacity rather than locking in long-term commitments. (Disclosure: I invested in this round.) Read more here →

• Anthropic releases Claude Opus 4.5, its strongest model for coding and agents yet → The update tops benchmarks like SWE-Bench Verified at 80.9% and leads in tool use and problem solving, with new features for Excel integration, Chrome browsing, and endless chats via memory compression. Read more here →

• Google's Nano Banana Pro dominates X with viral use cases → Built on Gemini 3, it excels at infographics, multilingual text, 4K edits, and much more. Users showcase everything from blueprints to 3D renders and social thumbnails. Read more here →

Nano Banana Pro’s structural breakdowns are insane:

Impressive demo showing that Nano Banana Pro is not just an image generator but also uses the intelligence of Gemini 3:

This post is going mega viral on X - more than 70M views so far. A reminder that from now on everything you see on social media might not be actually real.

• Factory integrates Claude Opus 4.5 as default for AI software dev → The new model boosts reasoning, judgment, and planning on complex tasks like migrations and prototypes, with a tunable effort parameter for speed vs depth. Now live in Droid at Sonnet prices for two weeks. Read more here →

• Ilya Sutskever: Scaling era ends, research drives next AI leap → In rare interview, SSI founder calls 2020-2025 the scaling age; now research differentiates breakthroughs. Predicts 5-20 years to superhuman learners that care for sentient life via novel tech paths. See more here →

• OpenAI rolls out ChatGPT shopping research for personalized buys → The tool asks clarifying questions, scans the web for prices and reviews, adapts via feedback like "more like this," and builds buyer guides. Unlimited use through holidays for all plans. Read more here →

• Black Forest Labs launches FLUX.2 for pro image gen and edits → The suite handles 4MP outputs, multi-ref blending up to 10 images, sharp text, and prompt fidelity for ads and mockups. Pro/flex via API, dev weights on Hugging Face. Read more here →

• DeepSeek Math V2 hits IMO gold as top open math model → The 685B MoE uses generator-verifier loops to score proofs on rigor, acing Putnam and ProofBench while fixing its own flaws. Weights out now for agentic math workflows. Read more here →

• OpenAI eyes 220M paid ChatGPT users by 2030 → Internal projections show 8.5% of 2.6B weekly users subscribing, up from 35M today at 5%. Plus and Pro plans drive growth amid broader adoption. Read more here →

• Anthropic study: Claude cuts task times 80%, doubles US productivity growth → Analysis of 100K convos shows 90-min tasks drop to 18 mins on average, with healthcare at 90% savings. Could add 1.8% annual labor boost over decade. Read more here →

That’s it for this week’s Lore Brief.

See you next week!

-Nathan Lands