Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
we are NOT PREPARED for the end of 2026
TL;DR
Google I/O looked boring until the bigger pattern clicked: everything is turning into an agent — Wes and Dylan argue the real announcement wasn’t a single flashy feature but Google layering Gemini across Search, YouTube, Gmail, Docs, Keep, Chrome, and Android so the interface increasingly becomes “just talk to it.”
The core AI debate here is code-first flywheel vs. world-model intelligence — they frame Anthropic and recursive self-improvement as the fast path to superintelligence, while Demis Hassabis and Yann LeCun represent the opposing bet that AGI needs a richer multimodal “world model,” not just bigger language models.
Math is becoming one of the clearest signs these systems crossed into real discovery — they cite Terence Tao and Scott Aaronson as evidence newer models became genuinely useful for mathematical research, then point to OpenAI’s claimed proof disproving an 80-year-old Paul Erdős geometry conjecture as a symbolic frontier moment.
You can outsource research, summaries, and coding help — but not understanding — one of the episode’s sharpest lines is that AI can handle notes, charts, and data gathering, but if you can’t explain an idea simply, in the Richard Feynman sense, you still don’t really understand it.
The coming AI IPO wave could be historic, but both hosts are really circling Google and SpaceX — they discuss Anthropic, OpenAI, and SpaceX as potential generational public companies, with Dylan favoring Anthropic’s focused coding flywheel while Wes keeps returning to Google’s TPU strategy and SpaceX’s leverage over future orbital compute infrastructure.
The ‘we are not prepared for the end of 2026’ feeling comes from how fast weird things become normal — whether it’s AI search watching videos semantically, a $300 AI basketball coaching your shot, or students and older relatives casually using ChatGPT, they keep coming back to the same pattern: what sounded insane 30 days ago suddenly becomes obvious.
The Breakdown
A sick-day apology, then straight into Google I/O
Wes opens by apologizing for the previous episode where he was visibly out of it, then immediately pivots to the “new insanity” in AI. That energy carries into their first topic: Google I/O, which Wes initially found underwhelming live, while Dylan caught up later and thought was surprisingly massive.
Why Google’s YouTube and Search changes feel bigger than they look
Dylan’s case is that Google isn’t just bolting on a chatbot — Gemini is becoming ambient infrastructure across Chrome, Docs, Gmail, and YouTube. The real shift is that YouTube search appears to be understanding videos more like a person would, not just reading transcripts, which makes Google’s product direction feel tied to Demis Hassabis’s bigger “world model” vision.
The argument underneath everything: are LLMs enough?
From there they get philosophical fast. Wes pushes the idea that intelligence may be a universal commodity — if language models already do coding, math, and proofs, why assume some extra ingredient is missing? Dylan counters with the ant-brain thought experiment: scale up an ant forever and does it become generally intelligent, or just a terrifyingly optimized ant?
Math breakthroughs, AI psychosis, and the 30-day genius cycle
They talk about how claims of AI discovering new theorems sounded delusional just months ago, and in some cases were. But now, with references to Terence Tao and Scott Aaronson, they say the newest models have crossed into actually useful mathematical discovery — culminating in OpenAI’s claim that one of its reasoning models disproved an 80-year-old geometry conjecture from Paul Erdős.
A $300 AI basketball and the explosion of weird new data sources
In one of the most Wes Roth moments possible, he pulls out an AI-enabled basketball with a Wi-Fi-style logo on it. The pitch is absurd and memorable: a sensor-filled ball paired with AirPods and an app that gives you shot-by-shot coaching, which becomes their running example for how AI is spilling into everyday objects and niche modalities nobody would’ve predicted.
The line they keep returning to: you cannot outsource understanding
This becomes the episode’s deepest thread. Wes says AI can summarize, research, quiz you, and help navigate terminal commands, but the one thing you can’t outsource is understanding — and he worries the path of least resistance will produce more people who sound smart by reading LLM-written scripts without actually grasping anything.
Google I/O, reconsidered: the through-line was agents everywhere
When they circle back, Wes finally explains why I/O landed for him after the fact: every product had evolved into an agent. Gemini in Search, YouTube, Docs, Keep, shopping, Chrome, and eventually glasses all point toward the same future where the UI is not menus and tabs but a context-rich assistant acting across your life.
IPOs, orbital compute, and why 2026 feels closer than people think
The last stretch turns to impending IPOs from Anthropic, OpenAI, and the SpaceX ecosystem, with Dylan favoring Anthropic’s focused coding strategy and Wes obsessed with SpaceX plus Google’s TPU buildout. The wildest tangent is also the most on-brand: they seriously discuss Google research on space-based data centers, launch costs dropping from roughly $1,600 per kilo toward $200, and Mark Zuckerberg beaming electricity back to Earth “with lasers,” which Wes says he is absolutely not okay with.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
