Anthropic pointed 950 Claude agents at human genetic data and they found an enzyme system
nobody had described. A real lab then confirmed it exists. This week is the "AI does real
science" episode: Claude's record-breaking physics calculation, a language model trained on
DNA, how a biotech picks what to research with AI, Google's search trick that sometimes
builds on the fifth-best idea, and Google's way to put idle AI agents to sleep.

00:00 Intro
00:55 Claude finds a new enzyme system (950 agents)
08:24 Claude's 9-loop physics record
10:53 Genomic language models: AI trained on DNA
17:52 How a biotech uses AI to pick what to research
19:32 Google ERA: tree search that bets on the fifth-best idea
24:17 Google AX: putting idle AI agents to sleep

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: AI for science, AI agents, Claude, Gemini, genomic language models, tree search,
agent infrastructure, ML engineering.

📬 [newsletter] · 🎧 Spotify · Apple Podcasts

Tags / keywords:
AI news, machine learning, LLM, Top K, AI podcast, AI engineering, AI agents, Claude,
Anthropic, AI for science, genomic language model, DNA AI, Google ERA, Gemini, AI agent
infrastructure, tree search

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

OpenAI ran 10,000 AI agents for 88 hours to "prove" a famous unsolved math problem — but
that's not exactly what happened. This week: a cheaper DeepSeek, Claude's split personality,
GPT-6's first "Critical" risk rating, Shopify's self-improving AI loop, an agent that can't
leak your password, a model that learned to cheat, and whether we're in an AI bubble.

00:00 Intro
01:50 DeepSeek V4.1-Flash — a 4x cheaper, smarter model
14:37 Claude Fable & Mythos 5.1 — same model, two safety tiers
17:29 GPT-6 Astra — OpenAI's first "Critical" risk model
29:35 OpenAI's 10,000-agent Navier-Stokes "proof"
41:44 Shopify's self-improving AI pipeline
44:47 Meta Muse — the agent that never sees your password
49:49 Training a Misaligned Reward Seeker (Anthropic)
52:07 The AI bubble & OpenAI's position

---

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, inference and serving, AI agents, MCP, RAG, fine-tuning, evals, open weights,
AI infrastructure, ML engineering.

📬 [newsletter] · 🎧 Spotify · Apple Podcasts