I founded Seed Club, a venture network that turns the conviction of the people founders most want on their cap table into coordinated capital. 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 September 14, 2026
Ambient agents turn downstream apps into silent databases
Personal AI layers are absorbing daily tasks across groceries, messaging, email, and video editing, reducing traditional apps to data backends. Durable moats shift toward data access, permissions, and proprietary API surface rather than UI, habit, or workflow design.
- Email clients and calendar apps becoming pure data stores, not interfaces
- Micro-SaaS era is over, with network effects and deep AI integration now required
- Successful software factories start minimal, expanding only as output is verified
Meta's AI assistant handles groceries, errands, and bookings in a day · @nicbstme AI newsroom wall tracks 15 live feeds and alerts on important stories · @sriniously Agents are reshaping software moats toward data and distribution · @signulll Consolidating messages and email into one AI agent interface · @signulll ChatGPT edits full video production without human Resolve work · @orenmeetsworld Micro-SaaS era is over, solo founders should pursue ambitious ideas · @johnrush Start AI software factories tiny and expand as you build trust · @mattpocockuk
Open source models and agent infrastructure converge fast
Model compression is delivering near-frontier performance at a fraction of the size, with a 27B model retaining 98% benchmark parity at 9x smaller footprint. A prediction that top three models will be open source within 12 months positions American cloud infrastructure as the economic winner.
- Sub-30MB automation models now run locally at 4,000 tokens per second
- Economic winners may be cloud providers serving open models, not model labs
- Agentic orchestration is becoming its own infrastructure layer, distinct from LLMs
Late-cycle signals accumulate in venture and crypto capital
Venture capital is showing late-cycle 'keep dancing' behavior, crypto prediction platforms face credible wash trading accusations, and retail SPV fraud is described as pervasive. Crypto VCs are shifting allocation toward liquid books rather than defending pure-crypto theses to LPs.
- Prediction market volume integrity is now a live question for the category
- SPV fraud in retail syndication is widespread but structurally underreported
- Crypto VCs moving to liquid books signals difficulty defending pure-crypto theses to LPs
Venture capital faces a 'dance while the music plays' moment · @venkyganesan Repeated identical trade sizes suggest Kalshi fakes PERP volume · @beniduboss Kalshi prediction market shows $2M wash trading over six weeks · @danielsapkota Kalshi defends crypto volume data against fake-trading claims · @icobeast Kalshi dispute escalates with job loss threat over public conduct · @beniduboss Kalshi accused of faking crypto trading volume · @beniduboss Crypto VCs shift toward liquid books as pure crypto bets get harder · @seyong SPV shilling to retail investors is rife with fraud · @WillManidis
Jev reframes classification as a programmable runtime primitive
A new classifier treats arbitrary classification as a zero-shot, type-safe decision layer rather than a model-selection problem, cutting tokens, latency, and cost. It is appearing across context compaction, browser navigation, model routing, and agent orchestration as a drop-in component.
- Context compaction drops from near 1M to 86K tokens in one second
- Browser agents using it cut action steps from 1,092 to 101
- On-device alternatives already 50x faster suggest rapid commoditization ahead
Jev classifier eliminates the need to manually select an AI model · @okkshitij Open-source browser agent cuts cost and latency for web-browsing AI · @FurqanR A breakdown of Jev's architecture, philosophy, and tradeoffs · @_raghavdixit_ Jev-powered sites personalize copy and design per visitor · @bryantchou Laya-MLX brings on-device AI classification 50 times faster than Jev · @mizorewww Jev brings model routing and memory filtering to Hermes agents · @StevenDarlow Jev routes agent calls to top endpoints for fractions of a cent · @tomosman AgentRun harness lets AI agents learn tasks and self-code solutions · @MiguelriosEN Jev skill integration promises 100x speed boost for Hermes agents · @BkashJosi Jev model rates spreadsheet column intent in about 100ms · @dabit3 Jev-use brings fast computer use to Mac, Windows, and Linux · @trycua Jev plugin cuts Claude context tokens by 90% in one second · @altryne Jev is a fast, cheap classification service worth watching · @fleetingbits Using Jev to score tool calls makes context compaction instant · @tamarajtran New browser agent cuts Google Flights search to 7 seconds · @tonysimons_ Open-source agent foreman monitors coding agents for drift · @JoshARosen Jev makes arbitrary classification a type-safe programmable primitive · @cocktailpeanut Browser agent finds flights in 7 seconds for under a cent · @IndraVahan