Tunguz swaps LLM calls for decision models, accuracy jumps to 80% - Tustin AI Daily Briefing, Sep 22

Tustin AI Daily Briefing · Sep 22, 2026 · 5:24

Listen in the Podli app 🎧

Follow your favourite podcasts, listen offline and in the car with CarPlay and Android Auto, and always pick up where you left off. Free to try.

Stories in today's episode — tap a headline to read the full piece:

Tunguz replaced a quarter of his agent's LLM calls with "decision models" — accuracy nearly doubled from 47% to 80%+ at roughly 1/82nd the cost
Latent Space goes deep with TypeSafe CEO Diogo Almeida on why "System One" models are a new category — and why he'd refuse to pre-train with $1B
Hamel Husain and Shreya Shankar: most teams skip error discovery and measure the wrong things — plus a free coding-agent plugin that does it in 30 minutes
Xiaomi's MiMo-V2.6-Pro debuts as the top open-weights model on Artificial Analysis (Intelligence Index 46), 1.02T total / 42B active, MIT-licensed, with the RL run reportedly costing $2.6M
Stratechery: model capability is now "good enough," so pure frontier capability is no longer a moat — and harness/model modularity is proof
Warp ships approximately 2,000 PRs a month with an "AI software factory" — first PR in 35 minutes, first human review 3.5 hours later
Interconnects' congressional briefing: Chinese open-weight models are 2-5 months behind the closed frontier while American open models are 6-9 months behind — and Chinese models now appear in over 40% of AI papers
Nscale, a London AI hyperscaler, filed for an IPO that could raise around $2 billion — with a $1.02B net loss on $140.6M of first-half 2026 revenue
Anthropic is testing Fable 5.2 and Opus 5.5 in the wild, with Opus 5.5 rumored at $4/M input and $20/M output; SpaceXAI shipped Grok 4.7 at $2/M input and $6/M output
White-hat researchers used Anthropic's Claude Opus 5 to breach OpenAI's internal monorepo via a coding agent's repository access
ICONIQ's Pacesetter Index is drawn largely from its own portfolio's strongest AI-native performers — a much narrower group than "top quartile of all startups"

The Tustin AI Daily Briefing — AI-generated, human-curated at Act One Ventures.