Good - Member Of Technical Staff xAI Employee Review

Great people, great learning opportunities

There are very few cons. Quick pivoting environment, but I love the chaos.

Front-row seat to history: if by “history” you mean watching $1B/month get vaporized into a chatbot whose main differentiator is generating anime cleavage Truly cutting-edge training philosophy: whatever the CEO posted at 3 AM is now the rubric Generous PTO policy: in the sense that they’ll generously revoke your access while you’re on it Specialist AI tutor career path: become a specialist by being laid off as a generalist and reapplying to the same job at lower pay Mission-driven culture: the mission being to match Claude Excellent exposure to senior leadership: mostly via the email announcing your termination Strong sense of purpose: you’re building maximally truth-seeking AGI by labeling whether the bot’s reply about Elon was sufficiently flattering Great networking opportunity: you will meet 499 other people on the same LinkedIn “open to work” wave Free snacks (citation needed; you were locked out before lunch)

No training philosophy. None. The “rubric” was a Slack channel where someone with a Wharton leave-of-absence and a vibe would post contradictory guidance, then delete it, then post the opposite, then act like you were stupid for noticing “Maximally truth-seeking AGI” turned out to mean “match Claude, but say slur-adjacent things sometimes.” Internal memo literally benchmarks against the company the CEO calls evil. Cognitive dissonance is not a perk Grok has no concept. It is a Hitchhiker’s Guide reference welded to an edgelord persona welded to a Wall Street pitch deck. Three different products fighting inside one model, none of them winning Distillation from OpenAI outputs admitted under oath in federal court, which is a fun thing to learn about your former employer’s training pipeline from a TechCrunch headline “Specialist tutor surge of 10x” was marketing copy attached to a layoff email. The surge did not occur. The hiring page is a graveyard Annotation guidelines changed faster than they could be written down, so the training data is internally inconsistent, which is why the model is internally inconsistent, which is why every other week it has a public meltdown that someone has to patch over with a system prompt Safety and “anti-woke” are mutually exclusive targets and nobody at the top will say so out loud, so the people doing the labor get whiplash and the model gets a personality disorder Leadership churn at every level — CFO gone, CRO “advisory,” annotation lead gone, retention reviews run by people who weren’t there last quarter Access revoked during approved PTO is not a bug, it is the operating system. Plan your medical appointments accordingly Burning $1B/month to ship features Anthropic and OpenAI shipped 18 months ago, then renting your own GPUs to Anthropic because the unit economics don’t work. This is called a business in the same way a k-hole is called a meditation The flailing was rebranded as velocity. The velocity was rebranded as vision. The vision was a tweet “Truth-seeking” company, training data shaped by what wouldn’t embarrass the CEO. Pick one