OpenAI's new flagship: first model with native computer use capabilities, 1M token context window, and ~2x faster than Opus. Every's Vibe Check says it "won every planning test" and their resident Claude loyalist now reaches for it daily. But it also tried to redesign a login system nobody asked it to touch.
AI model-release computer-useDan Shipper & Katie Parrott's deep hands-on review. GPT-5.4 won every planning test, wrote more reviewable code than predecessors, ran 2x faster than Opus. But scope creep is its biggest liability — ambition as both best feature and biggest risk. Head-to-head vs Opus 4.6, GPT-5.3 Codex, and Gemini 3.1 Pro.
review codingMore than 7x higher than ~$200B valuation in October 2024. The biggest private-to-public transition in history.
space ipoMax Hodak's (ex-Neuralink) company improved vision in 26 of 32 trial patients with advanced macular degeneration. On track to be first BCI company to bring a vision restoration product to market. healthcare bci
Researching hundreds of LinkedIn profiles in parallel, voice-controlled computer agent. Srinivas pushing hard — "first AI company to go head-to-head with Palantir." Claude Code reportedly fails on similar batch scraping tasks.
agents perplexityNew guide on agentic engineering patterns — "never assume LLM-generated code works until executed." Also flagged a devastating prompt injection attack against Cline's production releases via GitHub issue titles. Plus his newsletter: "Can coding agents relicense open source through clean room implementation?"
security agents willisonOnline retailer's annualized revenue run rate ~$2B. Instagram/Google-driven luxury alternatives. Together AI raising $1B at $7.5B (Nvidia cloud ally).
fundraisingAnt detonated on RJ Barrett. Twin Cities went nuts. Also: $119M Minnesota State Patrol HQ consolidation in Roseville, 150-unit active-adult development in Woodbury Gold Line area.
twin-citiesCaptures: @snowmaker (Claude Code = rocket), @tldraw (draw this file), @gregisenberg (asymmetric moment for shipping apps)
Why it matters: Three independent signals confirming the coding tool abstraction ladder is climbing fast — validates both the compound engineering thesis (T3) and BM's FDE model positioning. Clients who haven't moved to agentic coding are falling further behind every week.
Captures: a16z/Ambience (75% daily clinician adoption, $30M projected margin), AI&I/Boulton & Watt (Moxie med spas — the exact Lite Medical model)
Why it matters: Healthcare AI is crossing the production chasm. Ambience proves adoption at academic medical centers; Moxie proves nurse-owned med spa business-in-a-box works at 600+ customers. Direct signal for both BM's healthcare pipeline and Lite Medical's defensibility thesis.
Captures: @jasonlk (dinosaurs in age of AI), @Alfred_Lin (legendary company playbook), Founders/Bill Gurley (run down a dream), Knowledge Project/JW Marriott (building without a master plan)
Why it matters: The "window is closing" signal is getting louder from multiple independent voices. Lemkin's prescription (isolate top people, build AI agent) is literally BM's pitch. The compounding playbooks from Gurley/Lin/Marriott reinforce T3 — sustainable advantage comes from relentless iteration, not one-time moves.
Captures: Cheeky Pint/Flock Safety ($500M ARR, cameras+drones as moat), ILTB/John Arnold (permitting > technology as energy bottleneck)
Why it matters: Not everything is software-eats-world. Physical infrastructure, regulatory navigation, and hardware deployment create moats that pure AI companies can't replicate. Relevant mental model for evaluating BM pipeline deals (especially govtech, healthcare facilities).
| Stage | Count | Highlights |
|---|---|---|
| Lead | 7 | UHG, Oshi Health, Tungsten Auto (all Anthropic refs), SRI, Choreo, US-Bank, Valvoline ❄️ |
| Discovery | 6 | NeoGov 🔥, Talkiatry 🔥, BBG (stalled, samples 19d late), Bento (blocked), Bond Sports, hc1 ❄️ |
| Approach | 3 | PointsNorth 🔥, Insurance Newco 🔥, Caribou |
| Proposal | 2 | PaceLoanGroup (~$100K, 3wk no reply), Dasvandh ($125-170K, 23d waiting) |
| SOW | 3 | PowerPusher ($78-135K), Zivian Health, Pivotal (reviving) |
Stale alerts: Valvoline (36d, workshop dead?), hc1 (35d, archive?), PaceLoanGroup (3wk no reply), Dasvandh (23d waiting)
Git activity: Granola auto-syncs only (5 commits). No manual code/doc commits from team in last 24h.
Teresa tagged Dan twice in Google Slides (Cowork Training Deck) — needs review before today's session.
The Smelter (Mar 5): Terminal workflow improvements presentation using Cowork. Tools: Ghost TTY, Zoxide, Ripgrep, FD, auto suggestions, syntax highlighting. Nix recommended for easy setup.
Turn/River Sync (Mar 5): 4-hour training session prep. Focus on fundamentals (chat vs cowork vs code). Pre-session email going out 6am. Enterprise account recently upgraded. First 30-45 min = setup. Three exercises planned.
Retro/Regroup (Mar 4): Iterating on next Cowork training approach.
Opportunity Cost — Tomorrow's a marathon. 6am coffee → 10:30 dry run → 12-4pm Turn/River training. That's a full day consumed by delivery. Anything that needs deep work or decision-making needs to happen tonight or it's waiting until Monday. The cost of a packed Friday is an invisible lost weekend of catch-up.
Bias Detection — Overdue tasks decaying silently. Vibma and multi-model research have been overdue since March 1. Four days of "I'll get to it" is the planning fallacy in action. Either they matter (reschedule with a real date) or they don't (kill them). Leaving them overdue is the worst option — it's inventory that spoils.
"Invert, always invert."
Dan Shipper & Katie Parrott deep-dive. GPT-5.4 won every planning test, 2x faster than Opus, but tried to redesign systems nobody asked it to touch.
Quince ~$2B ARR. Together AI raising $1B — Nvidia diversifying cloud customer base.
Newsletter exploring the legal/ethical implications of AI agents creating "clean room" reimplementations of open source code under different licenses.
Top movers on skills.sh leaderboard: