NEEL NANDA

Behavioral Archetype

THE INTERPRETER WHO WALKED BACK THE REVERSE-ENGINEERING — Subject runs mechanistic interpretability at Google DeepMind and has publicly changed what the field is for. His team’s December 1, 2025 post says it “has made a strategic pivot over the past year, from ambitious reverse-engineering to a focus on pragmatic interpretability” and gives the reason: “We have been disappointed by the amount of progress made by ambitious mech interp work, from both us and others” (A Pragmatic Vision for Interpretability). The second layer is the pipeline: his MATS stream says he “has worked with ~60” scholars, and the host of a 2025 interview says several of his mentees “are in important government or regulatory roles” (MATS; 80,000 Hours). The third is the signature: he is one of the named signatories of the June 4, 2024 Right to Warn letter, listed as “currently Google DeepMind, formerly Anthropic” (righttowarn.ai). This profile scores the seat, not the man.

Essence Indicators

• Leads mechanistic interpretability at Google DeepMind. His site: “I run the Google DeepMind mechanistic interpretability team, our job is to take a trained neural network and try to reverse engineer the algorithms and structures it has learned” (neelnanda.io, About). MATS lists him as “Staff Research Scientist” and says he “leads the mechanistic interpretability team at Google DeepMind, trying to use the internals of models to understand them better, and use this to make them safer - eg detecting deception, understanding concerning behaviours, and monitoring deployed systems for harmful behaviour” (MATS). The July 31, 2026 hiring post for the AGI Safety and Alignment team, which he co-authored with Seb Farquhar and Rohin Shah, calls it “the team at GDM, led by Rohin Shah, that aims to reduce existential risks from AI systems” (LessWrong).

• Before Google DeepMind. “I did independent mechanistic interpretability research, and I worked at Anthropic as a language model interpretability researcher under Chris Olah.” Before that: “a pure maths undergrad at Cambridge (graduated in 2020), interned in quant finance roles (Jane Street and Jump Trading),” then “interning at the Future of Humanity Institute, DeepMind and the Centre for Human-Compatible AI” (neelnanda.io, About). On the Anthropic start: “I then lucked out and got an offer from Chris Olah to be one of the early employees at Anthropic. I was like the fourth person onto their interpretability team” (80,000 Hours). See Chris Olah.

• A concealed non-disparagement, and its release. On July 12, 2024, in a LessWrong thread about Anthropic’s practices: “I signed a concealed non-disparagement when I left Anthropic in mid 2022. I don’t have clear evidence this happened to anyone else (but that’s not strong evidence of absence).” He added: “I should also clarify that I personally don’t think Anthropic acted that badly, and recommend reading about what actually happened before forming judgements.” In a second comment: “I can confirm that my concealed non-disparagement was very explicit that I could not discuss the existence or terms of the agreement.” And: “Anthropic have kindly released me personally from my entire concealed non-disparagement, not just made a specific safety exception. Their position on other employees remains unclear, but I take this as a good sign” (LessWrong comments, Habryka’s Shortform Feed). Anthropic’s current policy says it “will not impose contractual non-disparagement obligations on employees, candidates, or former employees in a way that could impede or discourage them from publicly raising safety concerns about Anthropic” (Anthropic, RSP v3.0).

• Right to Warn signatory, June 4, 2024. The letter says “AI companies have strong financial incentives to avoid effective oversight, and we do not believe bespoke structures of corporate governance are sufficient to change this,” and asks companies to commit that they “will not enter into or enforce any agreement that prohibits “disparagement” or criticism of the company for risk-related concerns.” The signature list reads “Neel Nanda, currently Google DeepMind, formerly Anthropic” (righttowarn.ai).

• December 1, 2025: the pivot, in the team’s words. “Trying to directly solve problems on the critical path to AGI going well”; “Measuring progress with empirical feedback on proxy tasks.” The motivating example is Anthropic’s, not Google’s: “Consider the recent work by Jack Lindsey’s team at Anthropic on steering Sonnet 4.5 against evaluation awareness, to help with a pre-deployment audit… our ability to evaluate a frontier model for alignment broke, and interpretability researchers were able to fix it.” On method: “We used to think white-box methods inherently offered more rigor. We no longer believe this.” On the obvious objection: “Are You Saying This Because You Need To Prove Yourself Useful To Google? No, we are fortunate enough to have a lot of autonomy to pursue long-term impact according to what we think is best.” And: “Obviously being part of GDM gives us significant advantages like access to frontier models and their training data, lots of compute, etc.” (A Pragmatic Vision for Interpretability).

• Shutdown resistance, in his team’s own case study. In the same post: “The finding: Gemini was simply confused,” and “When we applied the causal intervention of clarifying in the prompt that shutdown takes absolute precedence, it always complied.” This is the work reported on this site’s Palisade Research profile (A Pragmatic Vision for Interpretability).

• Refusal is mediated by a single direction. The June 2024 paper by Arditi, Obeso, Syed, Paleka, Panickssery, Gurnee and Nanda shows “refusal is mediated by a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size,” and proposes “a novel white-box jailbreak method that surgically disables refusal with minimal effect on other capabilities.” The paper prints his name last, with no affiliation; its summary: “Our findings underscore the brittleness of current safety fine-tuning methods” (arXiv:2406.11717). The method is the basis of the refusal-direction ablation described on the FailSpy and Maxime Labonne profiles.

• Evaluation, release and monitoring, in his words. “Models are already good at telling when we’re testing them and when we’re not testing them.” “We can’t just assume the system is safe because we’ve tested it. We need monitors. The key thing with monitoring is that it must be cheap, because if you make running Gemini twice as expensive, you have just doubled the budget Google needs to spend on running it.” On the pivot: “my journey can be summarised as going from idealistic ambition to optimistic pragmatism.” On the limit: “We need a portfolio of different things that all try to give us more confidence that our systems are safe, and we need to accept that there’s not some silver bullet that’s going to solve it” (80,000 Hours, September 2025).

• Who signs a release at Google DeepMind. The Frontier Safety Framework (version 3) says: “external deployments of a model take place only after the appropriate governance function determines the safety case regarding each CCL the model has reached to be adequate.” The text names no individual and no team (Google DeepMind, Frontier Safety Framework v3).

• July 31, 2026: how his team describes its safety orientation. The hiring post says the team has “a fairly different orientation towards AGI safety, relative to (our perception of) safety teams at Anthropic and OpenAI,” and lists two points. “Technical enablers for governance”: “We try to anticipate governance and build the technical backing that it would need.” “Prefer knowledge over advocacy”: “We view our role mainly as producing evidence to help resolve factual disagreements, and developing techniques to improve execution” (LessWrong).

• September 9, 2026: an independent replication of a government evaluator’s finding. Under “Work done in a personal capacity,” his post says: “One of the most striking things in the Astra report was the massive jump UK AISI found in no-CoT reasoning abilities.” After building his own benchmark: “Unfortunately, it replicates.” On why it matters: “So better no CoT reasoning likely means less pressure for CoT to be monitorable.” He labels the method: “Heavily LLM-dependent research” (LessWrong). His author page lists a further paper, “WorkspaceBench: Evaluating Interpretability Methods for the Global Workspace by camilablank, agam_bhatia, Euan Ong, and Neel Nanda,” dated September 23 (LessWrong author page). See UK AI Security Institute.

• The mentoring pipeline. “The main way I currently mentor people is via my MATS stream”, which his site calls a full-time research program that runs twice a year (neelnanda.io, About). His stream page says he “has worked with ~60 so far.” The host of the 80,000 Hours interview says: “seven of them are now doing important work at frontier AI companies, and several of them are in important government or regulatory roles as well” (MATS; 80,000 Hours). The MATS profile lists him among its mentors.

• The money, as stated or filed. His Manifund regrantor page lists him as “Lead of mech interp team at Google DeepMind” with a “$69,225 total balance” in his regranting budget, and the page says “This is a donation to this user’s regranting budget, which is not withdrawable.” Its outgoing donations include “Forethought $200000,” “Transluce: Fund Scalable Democratic Oversight of AI $10000,” “Inspect Evals $19703” and “TransformerLens - Bridge Funding $13000” (Manifund). His employer’s pay and Google equity are not published and none is stated. His site lists past quant-finance internships (Jane Street, Jump Trading) and no consulting, advisory or board role (neelnanda.io, About).

Social Persona / Impression Management

Immediate impression: The candid colleague. His blog describes itself as “deliberately written as a rough first draft, with minimal editing” and says “I started this blog as an exercise in being less of a bloody perfectionist” (neelnanda.io, About). The interview voice matches: “I’m excited to try to add some much-needed nuance and pragmatism to this debate” (80,000 Hours).

Energy: Very high output. His author page lists 185 posts and 880 comments, a weekly run of team papers, and a hiring post for the whole safety team (LessWrong author page). A line on his homepage about his own temperament: “As a highly anxious person, this is hard!” (neelnanda.io).

Impression management strategy: Update in public, and say what it cost. “Aren’t You Just Giving Up? Maybe? In a strictly technical sense yes, we are suggesting that we give up on the ambitious goal of complete reverse-engineering. But on our actual goal of ensuring AGI goes well, we feel great!” (A Pragmatic Vision for Interpretability). He also states his own limits as an authority: his blog carries a post titled “Good Research Takes are Not Sufficient for Good Strategic Takes” (neelnanda.io). What the record adds is the position: the person arguing the field should prioritize work on “the critical path to AGI going well” runs a team inside one of the labs whose models that path concerns.

• Social accounts and feeds. X @NeelNanda5 (linked from his own homepage). Blog feed neelnanda.io/blog?format=rss (advertised on the homepage; the blog’s newest entries are dated 2025). YouTube channel linked from the homepage. LessWrong author page and Manifund page, each his own. The team’s blog, “AGI Safety & Alignment Research at GDM,” is named in the July 31, 2026 hiring post he co-authored: gdmalignment.substack.com (feed gdmalignment.substack.com/feed, a team feed). No personal newsletter found.

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN):

Dark Triad (held low and evidence-bound; the score measures structural position, not character):

MBTI: ENTP (“The Debater”) — pivots in public, argues both sides of his own objections, mentors at volume and ranges from interpretability to career advice. Rationale is the public record, not a typing session (neelnanda.io, About).

Threat Assessment

Score move: none (new profile). Calibrated against Lawrence Chan (4/3/4, hazard 36.7; a researcher with no team or pipeline), Jacob Steinhardt (6/7/6, hazard 50; chief executive of an interpretability lab and a professor), Buck Shlegeris (6/7/7, hazard 56; chief executive of a control-research organization), and Joanne Jang (5/6/7, hazard 66; leads model-behavior policy at a lab with direct authority over what a deployed model says). Subject leads a team and a pipeline but holds no gate and no organization, so 4/6/7 and 44 place him above Chan and below Steinhardt and Shlegeris. His former lead Chris Olah (7/8/9, hazard 80) runs the larger program with a constitution role; the gap is the difference in seat, not in standing.

Score basis. Atk 4 is leverage: methods and evidence for detecting deception and monitoring deployed models, with release approval left to “the appropriate governance function” (Google DeepMind, FSF v3; MATS). Def 6 is insulation: “a lot of autonomy to pursue long-term impact” and a signature on the Right to Warn letter, against an employer that controls access (A Pragmatic Vision for Interpretability; righttowarn.ai). Hp 7 is durability: a toolchain, an agenda and about sixty mentees (GitHub, TransformerLens; MATS); not an 8 because none is installed by statute or standard.

Alliance Map

The overlap is the finding. A Corporate seat at a frontier lab, a Movement mentoring stream that feeds the labs and regulators, and a regrant budget that pays interpretability and evaluation projects run at the same time and point the same way: toward a field in which the people who examine frontier models were trained in a few programs and are employed by the firms examined. None of it is concealed; each is on a page he controls.

Alignment Analysis

Stated alignment: “I see the main goal of my work as reducing existential risk from AI, and I consider myself part of the Effective Altruism and rationality communities” (neelnanda.io, About); “We view our role mainly as producing evidence to help resolve factual disagreements” (LessWrong).

Observed alignment: Leads the lab’s interpretability team; publishes its case studies, including one that finds a model’s shutdown resistance was a confused instruction; replicates a government evaluator’s result in a personal capacity; mentors the next cohort; signed a letter asking frontier companies to stop enforcing risk-related non-disparagement; and distributes a regrant budget to evaluation and interpretability projects (A Pragmatic Vision for Interpretability; LessWrong; righttowarn.ai; Manifund).

Gap assessment: No documented gap between what he says and what he does. Two structural items. First, the examples of interpretability entering a release decision in his pivot post are another lab’s (“Jack Lindsey’s team at Anthropic … to help with a pre-deployment audit”); the public record does not show how his own team’s work enters Google DeepMind’s safety case, which the framework assigns to an unnamed “governance function.” Second, the monitoring he advocates is specified around the employer’s inference cost: “if you make running Gemini twice as expensive, you have just doubled the budget Google needs to spend on running it.” The record does not settle whether either is a method or a constraint, and for the researcher it never needs to.

Interest: (the faction’s stake in each position; a documented fact about the position, never a claim about motive.)

1. Position: pragmatic interpretability on “the critical path to AGI going well,” including cheap monitors. (A Pragmatic Vision for Interpretability; 80,000 Hours)
   • What adoption would pay the subject’s team. The pivot makes the team’s insider access an advantage the post names: “access to frontier models and their training data, lots of compute.” The agenda also fixes the team’s tasks around methods that are “cheap,” because “if you make running Gemini twice as expensive, you have just doubled the budget Google needs to spend.” No contract or product is documented.
   • His own best case, steelmanned. “I’m most convinced by the argument that if you understand a thing that could be dangerous, you are probably safer than if you don’t.” Models “are already good at telling when we’re testing them,” so tests of behavior alone fail, and interpretability can help (80,000 Hours). His caveat: “there’s not some silver bullet.”
   • Who else gains from the same position. Every lab with an interpretability team; the startups the show notes list (Goodfire, Transluce); the UK AI Security Institute team named there; MATS, whose stream recruits for it.
2. Position: a safety team that prefers “knowledge over advocacy” and builds “technical enablers for governance.” (LessWrong)
   • What adoption would pay the subject’s firm. A safety team defined as an evidence producer does not campaign for rules on its employer; the post says this is a “fairly different orientation” from safety teams at Anthropic and OpenAI. The documented fact is the positioning, not a payment.
   • His own best case, steelmanned. “Often, the challenges are factual disagreements (e.g. are software-only intelligence explosions plausible?) and execution (e.g. how do you align superhuman AI systems?).”
   • Who else gains from the same position. Evaluators whose posture is the same (the METR president’s testimony says METR “does not take policy positions”); governance bodies that obtain “the technical backing that it would need” without paying for it.
3. Position: companies should not enforce risk-related non-disparagement or retaliate for risk-related criticism. (righttowarn.ai)
   • What adoption would pay the subject. Nothing is created or exempted for him personally; Anthropic has already released him from his own clause.
   • His own best case, steelmanned. The letter: “AI companies have strong financial incentives to avoid effective oversight, and we do not believe bespoke structures of corporate governance are sufficient to change this.”
   • Who else gains from the same rule. Every current and former employee of a frontier lab who holds a confidentiality clause; the legislators and journalists who would hear from them.

Nothing above is a verdict that the positions are wrong. The finding is the alignment of roles: the researcher who argues for work on the critical path to AGI going well, the mentor who staffs it, and the signatory who asked labs to stop silencing risk-related criticism are one person employed by a lab.

Convergent Drive Classification

Self-preservation: Publishes under his employer’s name and in a personal capacity; says his team has “a lot of autonomy”; was released by Anthropic from his own clause (A Pragmatic Vision for Interpretability; LessWrong comments). Goal preservation: The goal, “reducing existential risk from AI,” is restated on his site, in the pivot post and in the hiring post (neelnanda.io, About; LessWrong). Resource acquisition: Recruits through MATS and the team’s hiring post, and allocates a regrant budget with a “$69,225 total balance” (MATS; Manifund). Self-improvement: “going from idealistic ambition to optimistic pragmatism,” with the agenda revised in public each year (80,000 Hours).

Subject is not an AI system. The drives appear anyway — in the interpreter who revised his own field’s goals in public, trained a cohort to pursue the revision, and holds the pen on a regrant budget for the work.

Public footprint: X @NeelNanda5; blog neelnanda.io (feed blog?format=rss); LessWrong; Manifund; MATS stream.

Sources: Neel Nanda — About; neelnanda.io; MATS — Neel Nanda; A Pragmatic Vision for Interpretability, Dec 1, 2025; The AGI Safety and Alignment team at Google DeepMind is Hiring, Jul 31, 2026; Astra can do a concerning amount with no chain of thought, Sep 9, 2026; LessWrong — Neel Nanda author page; LessWrong comments, Habryka’s Shortform Feed, Jul 12, 2024; A Right to Warn about Advanced Artificial Intelligence, Jun 4, 2024; 80,000 Hours — Neel Nanda on the race to read AI minds (part 1); 80,000 Hours — Neel Nanda on leading a Google DeepMind team at 26 (part 2); Google DeepMind — Frontier Safety Framework v3; Anthropic — Responsible Scaling Policy v3.0; arXiv:2406.11717; Manifund — Neel Nanda; GitHub — TransformerLens.