▶ YouTube Summaries

• https://www.youtube.com/watch?v=23BtT8P7rCA
• Original title: Dear Google, we need to talk.

Theo argues that Google's internal culture, not its intelligence or resources, is why it is losing the AI race. A wave of top researchers is leaving (several for Anthropic), and the firing of the engineer who built the viral Google Workspace CLI exemplifies a company that punishes the kind of bottom-up experimentation that produced Claude Code and Codex. He contends Gemini models are smart but behave poorly at long-horizon agentic tasks because Google lacks the right training data (agent interaction histories, not raw code volume) and the pipelines to create it.

The departures and the broader thesis

• After ~9 years at Google DeepMind, top researchers are leaving, with names like Jonas Adler and Alexander Pritzel going to Anthropic. Theo counts four of Google's biggest names leaving back-to-back, three of them to Anthropic specifically. (He also notes Noam leaving for OpenAI, despite Google having spent $2.7B pulling him out of Character AI.)
• Google owns a chunk of Anthropic stock, so it benefits indirectly, but the departures signal that "things are burning internally."
• Theo's real focus is not the departures themselves but the environment and culture that produces them. He admits some fear in making the video because Google previously demonetized one of his videos for being too critical of Gemini/DeepMind (it was later remonetized, helped by friendly Google insiders like Logan, Philip, and Jack Witherspoon).

Google knows it is behind

• An April leak: Google DeepMind formed a strike team to improve its coding models, with Sergey Brin directly involved.
• Despite having the world's largest internal codebase (over 2 billion lines), Google lags Anthropic and OpenAI on coding and agents. For a while Google internally assumed topping certain benchmarks meant they were ahead as a frontier lab; that proved false.
• A new leak: DeepMind is unsatisfied with the current Gemini 3.5 Pro (referred to in the transcript variously as 3.5 Pro / 3 1 Pro), delaying its launch from June to July to tweak the frontier model. Theo dismisses the Business Insider framing as garbage, especially the claim it will be "better at long horizon tasks and powering agents" — historically Gemini's worst area.

Smart models, bad behavior

• Theo's core point: "The problem isn't their intelligence, it's their behavior. It's how they work, not what they know." Gemini tops obscure knowledge benchmarks (e.g. his "skate bench," where Gemini hits ~96% while few labs reach 80%) and is strong on baked-in knowledge and spatial reasoning.
• But the models behave like a brilliant coworker who ignores your messages and doesn't show up — they get stuck in dumb reasoning loops, read files they shouldn't, and grow less coherent the longer a task runs.
• His framing: Gemini has "next generation intelligence and last generation capability" — bad at driving a task to completion, using the harness/tools, and the do-action / wait-for-response / do-next loop.

The data problem: histories, not code volume

• Codebase size has nothing to do with training quality. Having lots of code does not make good coding models — having histories of changes does, especially alongside agents: the before/after of a codebase plus real back-and-forth between a human and an LLM, usable for RL.
• This is why a small company like Cursor caught up quickly (e.g. training/post-training models like "Kimmy K25") — they have real usage histories.
• DeepMind is at its core a research group excited about baking world knowledge into weights; getting marginally better at code excites them less, so they never built the pipelines to synthesize the agentic interaction data needed for RL.
• Theo's read on Google's generous Opus 45 usage in Anti-Gravity: it wasn't to push the product, it was to harvest behavioral data on how Opus 45 acts inside their harness, to train Gemini to behave more like it.
• Tool calling: Google is "the only lab putting out models that can still barely form coherent tool calls." Cursor had to do enormous prompt/tool shaping to make Google models work.

Live demo

• Theo tries Gemini in Cursor (refusing to use Anti-Gravity, "the worst software I've ever used") on his "Lakebed" framework codebase. Cursor's UI fails to show reasoning traces and the explorer breaks; he blames the Google APIs being the worst.
• With thinking enabled, the traces are "not as bad as it used to be" — genuine progress — but still produce slop: the model fixates on irrelevant cloud MDN skill files and emits generic filler ("My understanding has solidified... integrating core functionalities like routing, error handling, authentication...") that wasn't relevant to the task.

What Google could do instead

• Some companies wisely decided to build infrastructure for models rather than models themselves. Cloudflare is praised for making their tools work well with existing models.
• Google could make it easy to integrate their tools (Google Cloud, Google Workspace) with other models and agents via a CLI — exactly like the Google Workspace CLI that went viral and got its creator fired.

The Justin / Google Workspace CLI firing

• Theo reads the post verbatim: Justin was fired ~two months ago for building the Google Workspace CLI, which hit #1 on Hacker News, gained thousands of GitHub stars and many thousands of users in days. Directors asked what they could learn from it, while legal grilled him over using Google's logo/brand colors on the GitHub repo.
• The likely cause: Workspace leaders feared disruption — broadly, fear of what agents mean for Workspace. The irony: two days before he was fired, Google Cloud Next announced an official Workspace CLI was coming.
• Positive reactions he collected included Ryan Carson, Swyx, and Adi Osmany (who Theo believes left over the cultural shift these tools represent, not the firing directly).

Culture as the root cause

• Claude Code began as an internal Anthropic research experiment — Anthropic even told partners like Cursor not to worry about it. It's now the biggest existential threat to Cursor.
• Codex (the leak Theo shares): not a strategic OpenAI plan but an internal hack project by a single security engineer who wanted models (03 mini) in his terminal, built before Claude Code was announced. He got it far enough internally that OpenAI shipped it and let him open-source it.
• "If he had done that same thing at Google, he would have been fired for it." Anthropic and OpenAI created environments rewarding founder-style experimentation; Google asks "what are you doing?" and fires you.
• Theo relates personally: at Twitch he rebuilt a better mobile app in 4 days with 10 people at a hackathon and was rewarded with a trophy and an HR warning because the mobile team was furious; he left over it.

Conclusion: Google had every lead and is blowing them

• Google led on compute, data, codebase size, talent, and capital — yet every lead now seems not to matter or is being squandered (people leaving for Anthropic; reportedly even renting TPU capacity from Elon).
• Google "had everything they needed to win, but Google itself doesn't seem capable of doing anything but losing right now."
• Theo believes Google has finally woken up to how far behind it is, but until it fundamentally flips internally, expect more departures. He half-jokes Mistral has a better chance of success, says he's "pretty much given up" on Gemini, and asks viewers where they think it's going.