Mistakes to Avoid

BAD: Treating Constitutional AI as a feature set rather than a negotiation process. A February 2024 candidate described “adding constitutional rules” as a Jira ticket. Anthropic’s safety team doesn’t work in tickets. The debrief note: “Candidate has not engaged with how principles are actually derived.”

GOOD: Describing the specific review process for constitutional rule changes, including which research leads have veto authority and how long the “principle incubation” period typically runs. Reference the “Collective Constitutional AI” blog post and its implications for democratic input in rule-setting.

BAD: Using “Google it” as a research strategy in the DeepMind loop. A September 2023 candidate, asked about Gemini’s multimodal capabilities, responded “I’d need to look at the latest blog post.” The interviewer—who’d contributed to that blog post—marked “lacks product curiosity.”

GOOD: Pre-deploying knowledge of Gemini 1.5’s specific context window capabilities (announced February 2024), including the technical constraint that audio and video processing shares the same token budget. Propose a specific enterprise use case (legal document review across languages) that exploits this constraint.

BAD: Negotiating compensation using startup logic at DeepMind or Big Tech patience at Anthropic. A candidate in March 2024 told Anthropic’s recruiter “I need two weeks to consider”—the role was filled by an internal referral before the deadline. Another demanded immediate equity refresh talks at DeepMind; Google doesn’t negotiate refreshers pre-hire.

GOOD: At Anthropic, citing specific valuation milestones and their implications for your equity’s potential value. At DeepMind, accepting the band and negotiating on sign-on, relocation, or starting bonus—elements with flexibility in Google’s system.

---

FAQ

Does Anthropic require a technical degree for PM roles?

No, but the non-technical candidates who pass have all done something specific: published analysis of Anthropic’s research, built in the Claude API ecosystem, or held policy-adjacent roles where they negotiated between competing safety frameworks. The February 2024 hire for “Constitutional AI Product” had a philosophy degree from Reed and had published a Substack series on “Claude’s refusals as political philosophy” that a research scientist had read. Your degree doesn’t matter; evidence of genuine engagement with their specific intellectual project does.

How long do these interview processes actually take?

Anthropic’s 2024 process averaged 47 days from recruiter screen to offer, with two candidates in my network experiencing 90+ day delays due to “founder review” bottlenecks—Dario Amodei personally reviews final offers above certain equity thresholds. DeepMind’s process, post-Gemini reorganization, averages 31 days but includes a mandatory “Google Hiring Committee” review that adds 10-14 days and has a 12% rejection rate even with loop support. Neither process rewards urgency; both punish passivity. The candidate who checks in strategically every 5-7 days advances faster than the one who waits.

What’s the single biggest differentiator between candidates who receive offers?

At Anthropic: demonstrating you’ve changed your mind about something important based on new evidence. In a March 2024 debrief, the hire-no hire split tracked exactly with which candidates could describe a specific Constitutional AI principle they’d initially disagreed with and now supported, with reasoning.

At DeepMind: demonstrating you can make a decision with 40% information and course-correct, rather than demanding research certainty. The L7 promote who shared her interview experience described being asked “how would you launch this with no user research” and responding with a specific “research debt” framework for rapid validation—she’d used it at Facebook in 2017.amazon.com/dp/B0GWWJQ2S3).