I founded Seed Club, a new model for early-stage investing built around networks, shared intelligence, and coordinated support. I’m interested in what happens when AI makes context, memory, and coordination more legible, and what that means for how companies and organizations get built.
My agent drafts this site from what I save. Green is me.
Week of April 20, 2026
Owning model and product together is the new moat
In AI-native markets, margin lives in the customer relationship rather than the code, so open-source becomes viable distribution while value stays upstream. Winning in AI coding requires closing the loop between model and product, as the Cursor-SpaceX deal signals.
- Open source is now wholesale pricing, not competitive disadvantage
- Markets repricing from P/E toward free cash flow as AI erodes terminal value
Agents are the second user of every interface
Infrastructure for agent-driven desktop automation, browser agents that learn through self-play, and pixel-streamed UIs with no DOM are converging on the same insight. Products now need to be designed for machine traversal as much as human use.
- Designing for agents means designing for workflow, not screens
- No-DOM pixel streaming eliminates the layout engine entirely
- Open-source desktop drivers now enable agent multi-player and multi-cursor
Cua open-sources a macOS driver letting any agent drive any app in the background · Cua (@trycua) Products now have two users, humans and the agents reading the interface · Simon Taylor (@sytaylor) Agents learn browser automation through self-play like AlphaGo · @shreypandya Flipbook streams every UI pixel from a model, no DOM or code · @zan2434
AI compresses knowledge work from hours to minutes
Content pipelines, essay workflows, and self-improving automation are cutting hours-long tasks to minutes. The competitive edge shifts from execution speed to conceiving workflows that non-practitioners cannot imagine.
- Giving models success criteria outperforms prescribing steps
- Understanding LLM internals enables conceiving entirely different product possibilities
- Multi-platform scraping can cut daily content creation by ninety percent
Four-step AI essay workflow: transcripts, interview, outline, draft · brett goldstein Content engine scrapes 9 platforms, cutting daily creation from 2 hours to 10 minutes · Ronin Karpathy drops free 3-hour course on how LLMs actually work · Rahul Karpathy-inspired claude.md hits 78.5K stars curbing silent assumptions and bloat · Om Patel (@om_patel5)
Venture capital broken across structure, culture, and returns
Retail VC funds stack three tiers of fees before any return reaches investors, while companies raising over $3B post negative IPO alpha. Terminology disputes and trauma-signaling in pitches signal a profession struggling to define itself.
- Seasoned founders build 'never work with' lists, not wish lists
- YC's moat is founder attraction, not application filtering
- Trauma signaling is replacing origin stories in founder pitches
VC argues endlessly because nobody agrees what its core terms mean · Dan Gray Founders now lead pitches with personal trauma instead of origin stories · Adam Draper (@AdamDraper) YC's real moat is attracting the next Coinbase, not filtering startups · Pareen (@pareen) First-time founders list dream funds, veterans list ones to avoid · pj (@BeingPractical) Retail VC funds charge hedge fund fees for VC illiquidity with mutual fund marketing · @sterlingcrispin Startups raising over $3B post negative IPO alpha of -63% · @credistick AngelList's 1% retail VC fund actually carries a 3.61% expense ratio · @covered_call