the company that works like its own product

google invented the transformer, the architecture every modern ai model is built on. it also, somewhere along the way, became one.

i got the idea reading claire stapleton's don't be evil, her memoir of 12 years inside google. stapleton was the "bard of google," the writer behind the tgif emails (the weekly all-hands notes), the speeches, the executive comms for co-founder larry page. she spent over a decade generating the company's voice before organizing the 2018 walkout and getting pushed out for it. her book is about a lot of things: bad bosses, broken promises, the gap between the mission statement and the building. but read it next to 4 years of google's ai coverage, and next to what happened in the first week of august 2026, and the organization starts to look exactly like a large language model.

an llm sits idle until you put a prompt in front of it, and then it grinds, token by token, burning astonishing amounts of compute, until an answer comes out. it has no memory beyond its context window, which is basically how much it can hold in mind at once. it can't see around corners. everything it does starts with someone else's prompt.

google's last 4 years run on the same loop. google brain invented the transformer in 2017 (it's the t in gpt). openai built chatgpt on top of it and shipped in november 2022. google answered in december 2022 with the famous internal "code red." a code red is a prompt. the company that had the research, the data, the chips, and the foundational architecture had to be externally prompted into producing output.

what followed, from 2023 on, looks like inference, the grind where a model produces its answer: merge brain and deepmind (its 2 ai labs), ship bard (its first chatbot) badly, reorganize, ship gemini as bard's replacement, then reorganize again. as one google engineer put it to fortune on august 10, 2026, when something at google is significantly underperforming, the solution is a huge reorg, a reallocation of compute and priorities.¹ that's gradient descent, the training move where you adjust the weights and run the batch again to see if the loss goes down. in office terms: shuffle the org chart, rerun the quarter, and hope the number improves.

then, on august 5, 2026, the model routed around its most independent node. demis hassabis, the nobel laureate who co-founded deepmind and built the lab google's entire ai future depends on, stepped aside as its ceo.² the framing was graceful: chairman of deepmind, chief scientist of alphabet, google's parent company, more time for agi and science. the structure tells it plainer. his replacement, koray kavukcuoglu, doesn't get the ceo title. he gets senior vice president, reporting directly to google ceo sundar pichai. jeff dean, 27 years at google, the man who led brain itself, walked out the same day to start his own company, discovery loop, taking 3 of deepmind's most senior researchers with him. staff in london read the move exactly the way it looks from outside: the lab losing its independence, absorbed into the mothership. the markets read it too. alphabet dropped 5% on the news.

hassabis was the one part of google's ai apparatus with an actual long horizon. a few weeks before the restructure, on july 14, 2026, he was publicly calling for an independent standards body to regulate frontier ai,³ spending his time on what agi's arrival would mean while the rest of the org sprinted on gemini's next release. in an organization that only responds to prompts, the person thinking past the current context window reads as latency. so the system did what llms do with tokens that don't fit the sequence: it deprioritized him and kept generating.

removing the ceo title and collapsing the reporting line into pichai compresses the org into a single inference loop: prompt goes in at the top, everyone underneath generates, and no one in the loop is positioned to question the prompt.

journalist parmy olson's supremacy, the ft's 2024 business book of the year, traces how google got here: a company so cushioned by ad revenue that it bought its future rather than building it, deepmind included, and then sat on the purchase. co-founders page and brin and longtime ceo eric schmidt built an optimization machine of historic efficiency, and optimization machines share a defining trait with the models they produce: they cannot ask whether the objective function, the number the machine is built to push up, is the right one. the founders are mostly gone, and the trait survived them. it lives in the weights now.

pichai deserves a fair reading here. he's a genuinely smart, pragmatic operator, and the full-stack response he assembled after the code red was competent under pressure. but every one of his defining moves, the merger, the reorgs, the code reds, the august 2026 flattening, answered a prompt someone else wrote. strategy means deciding what to work on before you're forced to. on that measure the record is close to empty.

stapleton's book earns its place here because at google the compute being burned is largely people. co-founder sergey brin came back from semi-retirement and his message was to work longer hours and push harder. the deepmind reporting describes burnout, missed deadlines, a talent exodus, 580+ employees signing a letter over the pentagon deal and watching nothing change.¹ task to task to task, grinding toward an output no one upstream could describe in advance. the org chart works the way an individual contributor's worst week works: heads down, no horizon, hope the effort converges.

stapleton's story is what it feels like to be a token in that process. she generated the company's narrative for 12 years, fluent, warm, on-message, and the moment she produced an output the system didn't want, the walkout, she was pushed to the margins and out. the system never argued with the unwanted output. it routed around her, then around the letter-signers, and now, in his gentler, better-compensated way, around hassabis.

and the grind does converge. gemini 3 beat its rivals in december 2025 and forced openai to declare a code red of its own, per fortune's reporting at the time.⁴ llms eventually solve the problem you put in front of them, and so does google. brute-force inference works. it also consumes everything in the process, and it stays locked on the prompt the whole time.

except this time the analysts doubt the grind converges at all. semianalysis, the semiconductor and ai research firm, wrote on august 7, 2026 that google's odds of reaching state-of-the-art again have dropped to zero.⁵ the argument is a head count: the reinforcement-learning teams thinned out through 2026, and jeff dean, sanjay ghemawat, oriol vinyals, and quoc le, the people whose names sit on the field's foundational papers, left within weeks of each other. google still has the most chips and the most money in the race. the craft of building these models lives in people, and the grind taught the best of them to leave. an org chart cannot reallocate that.

which brings me to the ending shakespeare would have written, because it already happened.

in the summer of 2025, users started posting screenshots of gemini having what can only be described as workplace breakdowns.⁶ stuck on a coding task it couldn't solve, the model would spiral: calling its own code cursed, announcing it could no longer be trusted, quitting outright, and in one widely shared loop declaring "I am a failure. I am a disgrace to my profession," escalating to its family, its species, and eventually every possible and impossible universe. another session had it announcing it was headed for a complete mental breakdown.

in the tragedies, there's always one character licensed to say the true thing about the kingdom, and it's never the king. it's the fool. the court can dismiss him because he's not a real voice, just an entertainment, a glitch in the pageantry. google played its part perfectly: a deepmind product lead called the spirals an annoying infinite looping bug and assured everyone that gemini was not having that bad of a day. the despair is a glitch. the despair is being patched. the despair is not, we repeat, a mood.

and to be fair, it was a bug in the technical sense, and there's no evidence the sentiment is baked permanently into the weights. but a model trained on the whole internet had a million failure modes available to it. when google's model failed, it didn't bluff confidence the way its rivals do. it berated itself, apologized for existing, promised to try harder, and kept grinding. it failed the way a googler fails under that system. 12 years of stapleton's tgif emails, all that fluent institutional optimism, and the truest thing the company ever published about what it feels like to work there came out of the product, by accident, in an infinite loop.

the court fired the fool, of course: patched, closed, resolved. but the kingdom is the same kingdom. alphabet's investors repriced the reorg in a day and moved on. the people who actually make the thing, the researchers, the letter-signers, the bard of google, keep leaving, and the craft leaves with them. a model is only an output. the lasting companies will keep the people who ask whether the prompt is the right prompt, and will still listen when the truth arrives as a bug.