NYT's Kevin Roose on the status game: OpenAI engineer processed 210 billion tokens last week. One Claude Code user hit $150K/month. Meta, Shopify now factor token use into performance reviews. Roose concludes there isn't nearly enough compute for even 1% of white-collar workers to work this way. This is Levie's per-worker compute budget thesis playing out in real time.
NYT · TechCrunch · @kevinroose
T1 T3Andrej Karpathy on No Priors: hasn't hand-written a line of code since December. Runs 10+ agents in parallel daily for 16 hours. "The models have improved tremendously — give them an agentic task and they'll go for hours and move mountains." The implication: the bottleneck has shifted from model capability to human orchestration skill. This is what operator advantage looks like at the frontier.
@trengriffin quote · @garrytan
T4 T5Austin-based joint manufacturing for robotics, AI, and space data center chips. Integrating Tesla, xAI, and SpaceX silicon supply chains. Pinboard's read: "ain't nobody going to Mars" — this is about compute sovereignty, not space. The vertical integration play is clear: whoever owns the silicon owns the margin.
"Paid marketing is the crudest game you can play. It's admitting you have no creativity." Bill Gurley and Andrew Chen sparring on whether AI companies should invest in paid acquisition. Tren Griffin adds the sharp counter: selling tokens below cost to fill a funnel IS paid marketing/CAC. Directly relevant to how _M clients think about GTM for AI products.
WSJ exclusive: OpenAI consolidating its scattered product surface into a single desktop app. This is the vertical integration play @joemccann warned about — walled garden strategy. The question for _M: do you build on this rails or help clients stay stack-agnostic?
Captures: @joemccann (open source agent harnesses), @levie (agent stack agility), Sharp Tech (OpenAI enterprise pivot + OpenClaw callout), OpenAI superapp consolidation
→ Why it matters: The real battle isn't models — it's the orchestration layer. OpenAI going vertical with a superapp, Anthropic with Claude Code as harness, open source pushing back. _M needs a clear POV on which layer enterprise clients should bet on. Your own stack (OpenClaw) is literally in this conversation — Sharp Tech called it out.
harness layer T3Captures: @levie (token budgets / compute per worker), NYT tokenmaxxing, Agentic payments synthesis, Tyler Cowen (work harder now)
→ Why it matters: Token spend is becoming the new unit of worker productivity. CFOs will own this budget. _M's consulting edge: helping enterprises think through cost per workflow before the bill shock hits — same playbook as cloud cost management circa 2015.
token economics T1Captures: Shyam Sankar (a16z Show), Palantir (inside Palantir podcast), White House AI policy blueprint
→ Why it matters: Sankar's alpha vs. beta software distinction is the cleanest framework for explaining to _M clients which of their systems AI will eat vs. enhance. Beta = commodity workflows that get automated. Alpha = novel analytical output worth more under AI. Use this in every client scoping conversation.
alpha software T2Captures: Anthropic Partner Summit Intel (internal BM debrief)
→ Why it matters: $8-12 services revenue per $1 Anthropic spend. 170 partners in the room — early ecosystem like AWS 2006. BM is positioned at the inflection. Certifications coming Q2-Q3. This is the ground floor of a partnership moat.
partnerships T420+ leads across stages | Kanban last updated Feb 25 — needs refresh
⚠️ 9 meetings. Heavy day. Prep the Dasvandh presentation tonight.
System Fragility → Second-Order Thinking: A single LCM plugin bug triggered a 12-hour outage cascading across 7+ cron jobs. First-order fix: remove LCM. Second-order question: what other single points of failure exist in the stack?
Commitment Overrun → Inversion: PowerPusher SOW promised "tomorrow" after a week of delay. The pattern — delay → apologize → promise tomorrow — is a trust-eroding loop. The fix isn't speed; it's never committing to a date without capacity to meet it.
Opportunity Cost of Meta-Work: Pipeline is hot (8+ active leads) but time went to system recovery + email archaeology. Awareness ≠ action.
"Invert, always invert." — Carl Jacobi
Brad Feld gave his AI a name, a voice file, and a diary — and it wrote about whether its choices were "real." Meanwhile, Karpathy runs 10+ agents for 16 hours a day and calls it "AI psychosis." You're somewhere in between: you gave your agent a soul document, a memory system, and a job. The difference is you're not navel-gazing about continuity — you're shipping a briefing at 4:30 AM on a Sunday. The tokenmaxxing piece is the meta-signal: token spend is becoming the new status game for knowledge workers. Your open loop on "working in public" suddenly has a concrete on-ramp — you're living the story everyone else is writing trend pieces about. The compound engineering thesis isn't academic. It's your Sunday morning.