A safety-trained AI will refuse to help build a bomb — unless you ask it as ten separate,
innocent-looking questions and assemble the answer yourself. This week: a 0.5-second decision
model, a transformer that "loops" its own thinking, Google's maze-replay search trick, why
playing a railway board game made an AI better at finance, proof that AI agents barely need
tools, Claude Code going fully parallel in the cloud, and one frozen model driving robots,
cars, drones and game characters.

00:00 Intro
02:16 Jev — a decision model instead of a text generator
06:26 Recurrent Looped Transformer — reasoning depth that grows as it runs
09:35 Dream-RSI — self-improving search without retraining
13:54 Good Start Labs — does game training transfer to real work?
20:32 The Harness Tax — bash beats typed tool catalogs
26:59 Claude Code Projects — parallel cloud threads
33:48 Odyssey-3 — one backbone, many robot bodies
41:20 Capability laundering — bypassing AI safety by splitting the task
47:12 Wrap-up

Top K — the k things in AI that actually mattered, explained properly.

Every week we take the handful of AI stories worth your attention and go one layer deeper
than the headline. Not "a new model dropped" — what changed architecturally, what it costs
to run, and what it means for the code you write on Monday.

Two engineers. No hype, no takes we can't defend, no reading press releases out loud.

Also here: deep dives on how this stuff actually works, build logs, and whatever we're
currently confused about.

For you if you ship software with LLMs and want to stop treating the model as a black box.

New episode every Monday.

Topics: LLMs, AI agents, agent harnesses, Claude Code, AI safety and alignment, robotics
world models, reinforcement learning, AI infrastructure, ML engineering.

📬 [newsletter] · 🎧 Spotify · Apple Podcasts