要点
- Anthropic’s Alignment Science Blog announced the Anthropic Residency: paid fixed-term placements of six weeks, three months, or six months; this round closes 22 Nov 2026, with US and UK application links.
- Targets professors in quantitative fields, technical founders, and staff-level engineers; ML background, PhD, or prior papers are not required. Pay is $350,000–$500,000/year equivalent; San Francisco preferred, with remote possible especially for shorter stays.
- Framed as a senior adaptation of the Anthropic Fellows Program (prior fellow work cited: subliminal learning, Petri auditing, persona vectors); potential mentors include Ethan Perez, Evan Hubinger, Nicholas Carlini, Jascha Sohl-Dickstein, and others.
The announcement is not another intern funnel. It deliberately moves recruiting toward people who have already led ambitious technical agendas—professors, founders, staff engineers—and states outright that machine-learning background is not required.
Falsifiable read: if this Residency round fails, within a reasonable window, to produce Fellows-comparable empirical work or to move most residents into full-time AI research (at Anthropic or elsewhere), treat the “senior cross-field transfer” story as high-pay talent PR; if output and transition paths both hold, raise the weight of “agenda judgment over ML résumé.”
[1]Hard facts from the official post: paid, fixed-term; six weeks to six months; deadline 22 November 2026; compensation $350,000 to $500,000/year equivalent; San Francisco preferred, remote possible for shorter stays; aim that every Resident can continue full-time AI research afterward (Anthropic or elsewhere).
Interest areas span scalable oversight, adversarial robustness and AI control, model organisms of misalignment, interpretability, safeguards, AI welfare, plus frontier red-teaming and ML/RL systems. Named potential mentors sit on Alignment Training, Alignment Science, Alignment Stress-Testing, Scalable Oversight, Model Psychology, and related teams.
[1]Evidence is institutional self-description of a hiring/training design, plus pointers to Fellows outputs (subliminal learning, Petri, persona vectors). No cohort size, admit rate, historical conversion rate, or independent evaluation is attached.
Steelman: as capabilities rise, the scarce skill looks more like framing agendas and judging what is worth doing—muscles many senior quantitative people already have; high pay, mentorship, and compute lower the switching cost versus hiring only people who already train models. Beyond the steelman, $350k–$500k/year equivalent also reads as a bidding signal in the talent market, not only as safety-research capacity expansion.
[1]Strongest steelman: Fellows already showed newcomers can contribute quickly; shifting the same funnel to people who have led ambitious work should raise research-judgment density per unit time and soften the ML-résumé wall.
Steelman risk: success metrics (papers, internal tools, conversion, external roles) are undefined; six-week versus six-month learning curves differ sharply; “most full-time research roles require San Francisco relocation” tensions with remote short stays—brief exposure may not become a portable full-time path.
[1]For senior academics and industry technologists, a high-compensation, low-ML-barrier on-ramp into alignment research; for Anthropic, productizing the Fellows model as a “senior transfer” brand. For observers, the variables that matter after 22 Nov 2026 are admit mix, public research output, and post-residency destinations—not the salary band alone.
[1]Near term, judge by this round’s public work and destinations: if high-pay stays mainly buy brand exposure without auditable output, file it as a talent-market signal; if Fellows-grade work and full-time transitions arrive together, “senior cross-field into safety” becomes a copyable 2026–2027 lab expansion template.
【No.2】
[1]