Mighty Brief: April 28, 2026

Start Here

Google doing classified AI work for the Pentagon is the cleanest sign yet that defense AI normalization is over. The headline matters less than the permission structure behind it. “Any lawful government purpose” means the frontier labs are moving from flirtation to formal doctrine. If AI infrastructure is now national-security infrastructure, procurement and politics both get tighter from here.
The Information: Google Signs Classified AI Deal With Pentagon Amid Employee Opposition
OpenAI getting deeper into AWS does not erase the multi-model market. It confirms it. The useful read is not “OpenAI wins AWS.” It is that AWS customers already built habits around Bedrock, Anthropic, and cheaper house models like Nova. Distribution helps, but model loyalty is weaker than workflow loyalty.
The Information: OpenAI’s AWS Push Comes As Customers Embrace Rivals
Musk v. Altman moved from internet theater into courtroom mechanics this morning. A seated jury changes the texture. The OpenAI story is no longer just blog posts, leaks, and subtweets. It is discovery, testimony, and a public record. Even if nothing fundamental changes legally, the trial will keep dragging internal governance questions into daylight.
CNBC: judge seated a nine-person jury in the Musk v. Altman trial · Reuters: trial to reveal ongoing OpenAI power struggle
The better agent story is shifting from single prompts to operating systems. Three inputs lined up cleanly: Every’s “You Are the Most Expensive Model” on incremental determinism, Garry’s remote pair-agent/browser-control post, and the X chatter about reusable design skills. Same pattern from three angles: memory, governed workflows, and tool routing are becoming the real product.
Every: You Are the Most Expensive Model · Garry Tan on pair-agent browser control · Meng To on turning design elements into skills

Blank Metal

Unread BM inbox status is blocked, not clear. gog gmail search --account dan.wick@blankmetal.ai 'is:unread newer_than:1d' --max 10 --no-input failed on April 28, 2026 with Google OAuth invalid_grant / invalid_rapt. Treat that as auth drift, not signal.
Open-loop reality check: Rally, HR&A, Tungsten, Great Hill, Sentinel, Orbis, Care Providers of MN, and Edmunds Govtech are still live. The opportunity set is healthy. The risk is scattered attention, not empty pipeline.
Best BM frame today: regulated workflow adoption, implementation trust, and clear before/after operating change. The market is getting less impressed by “AI access” and more interested in “what actually changes Monday morning.”

Munger Observer

Inversion: today's daily log is empty. If the key calls do not get written down, today will not compound into tomorrow.
Opportunity cost: the old March loose ends are still charging rent. Either recommit to them or kill them.
Practical read: the highest-return move today is probably closing one old loop cleanly, not opening a new one.

Personal Thought

The tension in MEMORY still reads true: compounding versus scattered. This morning's news says the same thing in public-market language. Infrastructure is concentrating. Attention still is not.

Quick Scan