Oracle, Silver Lake and other investors in TikTok's US business agreed to give the government several multibillion-dollar payments for brokering the deal.
70+ developer interviews. Startups: ~100% AI-written code. Google: ~50% AI-written. Junior dev jobs (22-25) down 16% since 2022. Boris Cherny: 100% of his Claude codebase contributions written by Claude. Kent Beck back to coding after quitting — AI output "addictive, in a slot-machine way."
AI model providers are today's walled gardens. The open layer emerging: open source models, x402 payments, MCP, CLI tools, skills frameworks. "The meteoric growth of OpenClaw suggests the large pull of these approaches."
Agent edits train.py, human edits program.md. Fixed 5-min training runs, ~100 experiments overnight. "You're programming the program.md" — markdown as research strategy. Pattern generalizable beyond ML.
Animated viz showing OpenAI's long lead, Meta's rise and collapse, xAI's catch-up and stagnation, Chinese open-weight LLMs entering. PhDs get 65-70%, models approaching that ceiling. "We are probably near saturation."
Coding agents already driving massive inference demand — and that's a small % of knowledge work. Next: agents across all knowledge work with full computer access. "The run on inference capacity is coming."
1,692 physicians surveyed. Most common uses: medical research summarization and clinical care documentation.
900K+ New Yorkers uninsured. 92% of low-income legal problems unaddressed. "This isn't consumer protection. It's cartel protection."
Captures: @daniel_mac8 (Manus ex-lead: CLIs > tool calling), @garrytan (markdown files = power), @morganlinton/Perplexity CTO (moving away from MCPs), @karpathy (autoresearch: markdown as programming)
Why it matters: The agent control surface is converging on plain text and CLIs, not protocols — validates Dan's OpenClaw approach and the entire skills-as-markdown thesis. The Manus insight is especially sharp: Unix text streams from the 1970s are in LLM training data, making CLIs native to how models think.
Captures: NYT "Coding After Coders" (Thompson), @emollick Codex demos, @levie inference explosion, @illscience pre-AI transition playbook
Why it matters: The mainstream narrative shifted from "will AI replace coders?" to "here's exactly how it's happening." Junior dev jobs down 16%, Google at 50% AI-written, startups at ~100%. The Jevons paradox framing (cheaper coding = more software, not fewer jobs) is the smart take — and it's the one Dan's been saying.
Captures: USV "History Rhymes", @garrytan (open GitHub repos), @scottbelsky (talent density + alignment)
Why it matters: USV putting institutional weight behind the walled-garden → open-protocol thesis for AI. Explicitly names OpenClaw. This is the investing framework that validates the entire open skills ecosystem Dan's building in.
Captures: Alex Karp (zero-sum AI), Daniel Pink (6 irreplaceable skills), @illscience (transition playbook: "be the point of economic diffusion"), AMA physician survey
Why it matters: Direct fuel for Blank Metal's FDE model. The illscience playbook is practically a client pitch deck — "reorganize radically, put the model in charge of a business unit, humans handle exceptions." 80% physician AI adoption validates the healthcare vertical bet.
| Deal | Status | Signal |
|---|---|---|
| EQT | 🟢 Intro call 3/13 | Hot — PE firm, heavy ChatGPT, scaling AI fluency firm-wide. Token mgmt pain. |
| Jason Dehler | 📧 New intro email | Unread email "Intro blankmetal" — needs reply |
~40 leads tracked in pipeline. Only Granola auto-syncs in last 24h (no major repo activity).
Opportunity Cost — Breadth vs Depth. USV alignment doc, Omnia AI analysis, EQT intro, BM skills repo, research captures — five distinct ventures touched, none advanced past strategy/document stage. Munger would ask: which has the highest expected value per hour, and why isn't it getting all of today?
Second-Order Thinking — Skills Repo Bootstrap. Creating BlankMetal/skills as standalone repo is first-order good (portable, shareable). Second-order: does a separate repo create maintenance drag across two agents that outweighs distribution benefit? If no client uses it within 30 days, it's inventory, not infrastructure.
Incentive Analysis — Analysis as Procrastination. Eight Omnia AI opportunities beautifully ranked is satisfying work. But incentive structure rewards producing the document, not executing the top item. The ranking existed intuitively — SEO at scale was always #1.
"Invert, always invert."
New neolab: AI-driven R&D for biology and materials science. a16z + Kleiner Perkins co-leading $175M round. Rare Anthropic research departure.
Kalanick launching new AV company with Uber backing. Discussing acquiring Levandowski's Pronto.ai (autonomous mining/industrial).
Dividing bank roles: Morgan Stanley + Goldman on institutional, BofA + Citi on retail, JPMorgan advising. Not the typical 1-2 bank setup.
This week's bundle: Integration and AI (Copilot Cowork vs Anthropic's own integration), Oracle crushing earnings on AI wave, MacBook Neo review, Hollywood deep dive. Key insight: companies best able to harness model capabilities are the model makers themselves.
Kieran Klaassen turned a prompt into a working app in an hour, step-by-step walkthrough.