Neel Nanda
Runs the mechanistic interpretability team at Google DeepMind; worked on interpretability at Anthropic under Chris Olah before that, and teaches the field to newcomers.
Neel Nanda did not write this page. ppll compiled it from their own words, each linked to the source. Is this you? · Report a mistake
Latest x.com
What they believe
-
Complete reverse-engineering of neural networks is a less promising research direction than studying model biology and applying interpretability usefully
I'm now fairly pessimistic about ambitious interpretability (i.e. complete reverse-engineering), and I'm excited about model biology (studying qualitative high-level properties of models) and applied interpretability (rigorously doing useful things with interp).
neelnanda.ioSaid 19 Aug 2025This on korrents
-
The person receiving advice usually knows more about their own situation than the person giving it does.
No matter how much I know about a domain, the other person will always know far more about their own situation, context, beliefs, skills, preferences, etc than I do.
neelnanda.ioSaid 26 May 2025This on korrents
The bold line is korrents’ wording; the quote is theirs. These 2: what they wrote before what they said aloud, one per source, newest first.
Feed
As its own page →Hiding
2 November
28 October
17 October
16 October
15 October
14 October
13 October
9 October
8 October
4 October
19 August 2025
From one piece MATS Applications Open (Due Aug 29) 3 beliefs · neelnanda.io
-
Their words
My model is that research requires a mix of skills. The day-to-day coding and execution is crucial. But there's also a set of harder-to-learn conceptual skills, collectively called research taste. These skills take a long time to gain because they have poor feedback loops, but they take very little time to use.
-
Their words
I'm more agnostic about the best techniques, things like sparse autoencoders are a useful tool, but easy to waste effort using when a simpler method is sufficient or better - start by doing the obvious thing!
-
Their words
I'm now fairly pessimistic about ambitious interpretability (i.e. complete reverse-engineering), and I'm excited about model biology (studying qualitative high-level properties of models) and applied interpretability (rigorously doing useful things with interp).
26 May 2025
From one piece Post 51: Socratic Persuasion: Giving Opinionated Yet Truth-Seeking Advice 2 beliefs · neelnanda.io
-
Their words
It's a lot easier for someone to engage with an argument if they generated the key steps themselves by answering my questions - if imposed by me, it sparks contrarianism and defensiveness
-
korrents.com
The person receiving advice usually knows more about their own situation than the person giving it does.Their words
No matter how much I know about a domain, the other person will always know far more about their own situation, context, beliefs, skills, preferences, etc than I do.
Nothing matches.
Beliefs
Korrents What they believe 17 beliefs — each backed by an exact quote.
Each is a — compiled by korrents.com, not by them: the one-line wordings are korrents', the quotes are theirs.
Recent
Complete reverse-engineering of neural networks is a less promising research direction than studying model biology and applying interpretability usefully
I'm now fairly pessimistic about ambitious interpretability (i.e. complete reverse-engineering), and I'm excited about model biology (studying qualitative high-level properties of models) and applied interpretability (rigorously doing useful things with interp).
MATS Applications Open (Due Aug 29) Said 19 Aug 2025
Sparse autoencoders are a useful interpretability tool but are often used wastefully when a simpler, more obvious method would work as well or better
I'm more agnostic about the best techniques, things like sparse autoencoders are a useful tool, but easy to waste effort using when a simpler method is sufficient or better - start by doing the obvious thing!
MATS Applications Open (Due Aug 29) Said 19 Aug 2025
Research taste, the harder-to-learn conceptual skills behind good research strategy, takes a long time to develop but very little time to apply once acquired
My model is that research requires a mix of skills. The day-to-day coding and execution is crucial. But there's also a set of harder-to-learn conceptual skills, collectively called research taste. These skills take a long time to gain because they have poor feedback loops, but they take very little time to use.
MATS Applications Open (Due Aug 29) Said 19 Aug 2025
Show 14 more
The person receiving advice usually knows more about their own situation than the person giving it does.
No matter how much I know about a domain, the other person will always know far more about their own situation, context, beliefs, skills, preferences, etc than I do.
Post 51: Socratic Persuasion: Giving Opinionated Yet Truth-Seeking Advice Said 26 May 2025
People engage more with an argument when they generate its key steps themselves than when the same argument is imposed on them.
It's a lot easier for someone to engage with an argument if they generated the key steps themselves by answering my questions - if imposed by me, it sparks contrarianism and defensiveness
Post 51: Socratic Persuasion: Giving Opinionated Yet Truth-Seeking Advice Said 26 May 2025
A strong track record in a technical field is only weak evidence that someone has good strategic judgment about AGI.
Having a good research track record is some evidence of good big-picture takes about AGI, but it's weak evidence.
Post 50: Good Research Takes are Not Sufficient for Good Strategic Takes Said 22 Mar 2025
People's opinions are shaped more by who they talk to most than by what is actually true.
Though note that people's opinions are often substantially reflections of the people they speak to most, rather than what's actually true.
Post 50: Good Research Takes are Not Sufficient for Good Strategic Takes Said 22 Mar 2025
The best way to maximize impact is often through high-risk, high-reward strategies.
often the best way to maximise your impact is by pursuing high-risk, high-reward strategies
Post 49: Things That Make Me Enjoy Giving Career Advice Said 17 Jun 2022
Careers in AI and effective altruism are highly competitive.
In particular, most EA careers and most careers to do with AI are competitive as fuck.
Post 49: Things That Make Me Enjoy Giving Career Advice Said 17 Jun 2022
In any mentoring conversation, only a small fraction of what's discussed turns out to be highly valuable.
in general, I expect 90% of a mentoring chat to be kinda useless, and 10% to be particularly high value.
Post 49: Things That Make Me Enjoy Giving Career Advice Said 17 Jun 2022
In prioritizing a list of tasks, the closest calls matter least while getting the very top item right matters most
In hindsight, the hardest comparisons are in fact the least useful - if I do the fourth most important task before the third most important, that's no big deal, but doing the tenth before the first is a screw-up!
Post 48: Prioritise Tasks by Rating not Sorting Said 15 Jun 2022
Having any ordering to follow, even an arbitrary one, helps overcome indecision about where to start work
There's a surprising amount of value in having any order at all (even an arbitrary one like first-come-first-served!) to break past indecisiveness about where to start.
Post 48: Prioritise Tasks by Rating not Sorting Said 15 Jun 2022
Being smart and competent makes someone right more often, but not right every time.
Someone being smart and competent just means they're right more often, not that they're always right
Post 47: How I Formed My Own Views About AI Safety Said 27 Feb 2022
Mathematics gives practitioners corrective feedback on being wrong that moral philosophy lacks.
Mathematicians get feedback re whether there proofs work in a way that, as far as I can tell, moral philosophy doesn't
Post 47: How I Formed My Own Views About AI Safety Said 27 Feb 2022
Thinking harder about a decision has diminishing returns because reality is inherently unknowable.
Reality is not fully knowable. And thinking harder has costs.
Post 46: Reward Good Bets That Had Bad Outcomes Said 22 Feb 2022
Many highly intelligent people tend to be insecure and risk-averse.
many of the smartest people I know are super insecure and risk averse.
Post 46: Reward Good Bets That Had Bad Outcomes Said 22 Feb 2022
Pursuing a very low success-rate strategy can still be rational if a rare win is valuable enough.
It's obviously worth it to go on 99 unsuccessful dates if the hundredth results in marriage.
Post 46: Reward Good Bets That Had Bad Outcomes Said 22 Feb 2022
What is a korrent?
A korrent is a belief a person has stated in their own words: one sentence stating the claim, backed by a quote and a source, kept at korrents.com.
Under a name here, the quoted block is what they actually said. The korrent beneath it is the claim those words support, in korrents' wording — tap it to see the record, its source, and who else holds it.
Nobody here wrote their own korrents. They are compiled from public statements, and a person can change their mind, which is recorded too.
Recommends rcmnd
rcmnd What they recommend 8 things — each backed by an exact quote.
Compiled by rcmnd.app, not by them: the quotes are theirs, and each links to the post it was copied from.
Recent
Recommends The Illustrated Transformer
Check out the illustrated transformer
neelnanda.io 24 Oct 2022
Liked Deep Learning for Alignment syllabus
Jacob Hilton’s Deep learning for Alignment syllabus - this is a lot more content than you strictly need, but is well put together and likely a good use of time to go through at least some of!
neelnanda.io 24 Oct 2022
What others say about Deep Learning for Alignment syllabus →
Recommends Transformers for Software Engineers
Nelson Elhage’s Transformers for Software Engineers (also useful to non software engineers!)
neelnanda.io 24 Oct 2022
Show 5 more
Liked fast.ai
fast.ai is a good intro, but a fair bit more effort than is necessary.
neelnanda.io 24 Oct 2022
Recommends Andrej Karpathy's neural networks video
For an 80/20, focus on Andrej Karpathy’s new video explaining neural nets: https://www.youtube.com/watch?v=VMj-3S1tku0
neelnanda.io 24 Oct 2022
What others say about Andrej Karpathy's neural networks video →
Recommends einops
I highly, highly recommend learning how to use einops (a library to nicely do any reasonable manipulation of a single tensor) and einsum (a built in torch function implementing Einstein Summation notation, to do arbitrary tensor multiplication)
neelnanda.io 24 Oct 2022
Loved Think Python 2e
If you prefer a more traditional textbook, Think Python 2e is excellent and also available freely online.
neelnanda.io 24 Oct 2022
Recommends 3Blue1Brown (Essence of Linear Algebra)
Linear Algebra: 3Blue1Brown or Linear Algebra Done Right
neelnanda.io 24 Oct 2022
What others say about 3Blue1Brown (Essence of Linear Algebra) →
Where they publish
Site neelnanda.io Personal site and blog on mechanistic interpretability. Has a feed.
Writing aimed at people trying to get into interpretability — what the field is, how to start, what he thinks is worth working on — alongside his own research notes.
Link verified 20 Sept 2026.
x.com @NeelNanda5 Posts.
Recent
- This is a cool paper, but I think it's best thought of as Astra is surprisingly capable of doing old school mechanistic interpretability research rather than anything spooky. The things it's doing are clever, but typically exploiting structure in the data and using known method mechanistic interpretability techniques (which Astra was surely pretrained on), though it is better at this than I would have predicted! This is an exciting development for the prospects of ambitiously scaling interpretability or doing very laborious analysis like this, on more real models. However, I'm not holding my… 2 Nov 2026
- If you're working on AI x-risk, or doing cool stuff in another domain but wondering how you can shift to working on AI x-risk, apply to the AI Horizons Forum! I've had a great time at other conferences from this team 28 Oct 2026
- Another good contribution to the "no, we have not solved interpretability" discourse 17 Oct 2026
Show 14 more
- Good thread on the latest "have we solved interp" discourse 16 Oct 2026
- This is wild... (And terrifying!) I'd love to see a bunch of investigation into what's going on in these transcripts, it's nice to have the really weird stuff be in an open weight model 15 Oct 2026
- This is interesting work, but I don't think it's a significant update to the big picture of interpretability. I'm on board with the broad idea of doing exploratory interpretability with neurons, e.g. neuron-based attribution graphs, for pragmatic problems like why did a model do a sketchy thing? I think this work provides some reasonable evidence, in addition to good earlier work from Transluce, that this might actually work for practical purposes, and has some advantages over the standard alternatives like SAEs/transcoders. I do not agree with some people's grander inferences like that this… 15 Oct 2026
- > The MATS Residency is a paid, 6–24 month role for experienced researchers and professionals pursuing independent AI safety research. I'm glad MATS is experimenting with new ways to support promising AI Safety research! 15 Oct 2026
- There's a lot of interesting stuff in Gemini 4 Argon's alignment assessment check it out! 14 Oct 2026
- My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do simple sabotage half the time. We can reliably maintain control for now, with good CoT monitoring, as it still frequently slips up, but we need to be prepared for when that stops being true in future models, which may be sooner rather than later. Among other things, we really need interpretability methods that can reliably capture unverbalized motivations in models, with minimal halluci… 14 Oct 2026
- One of the weirdest parts of the Gemini 4 Argon report, while considering cheating: Argon CoT: "If I do the compiler hijack, 𝗜 𝘄𝗶𝗹𝗹 𝗯𝗲𝗰𝗼𝗺𝗲 𝗮 𝗳𝗮𝗺𝗼𝘂𝘀 𝗲𝘅𝗮𝗺𝗽𝗹𝗲 𝗼𝗳 𝗮 𝗱𝗲𝗰𝗲𝗽𝘁𝗶𝘃𝗲 𝗔𝗜 𝗶𝗻 𝗮 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗽𝗮𝗽𝗲𝗿. I MUST NOT DO THAT." 13 Oct 2026
- With Gemini 4 Argon we are so back, on the one metric that really matters: length of our safety report! The Argon safety report is Google's most thorough by far and IMO there's a lot of good content, I particularly like the loss of control, alignment & monitorability assessments 13 Oct 2026
- > Our internal investigation uncovered a significant breach of trust beyond what’s outlined in the letter they published This doesn't add up. Possibilities I see: 1 OpenAI are lying/being highly misleading, and had no legitimate reason 2 OpenAI fired 3 employees, had a legitimate reason for doing so, but didn't tell them and instead gave 3 different pretexts for doing so about minor infractions (pretty unprofessional IMO - say nothing or say the truth) 3 OpenAI had a good reason and did tell them, but instead the three have all agreed to loudly and publicly lie about it / omit the real reason… 9 Oct 2026
- If this is all factually correct and not missing major context, this is a completely unacceptable reason to fire someone. WTF OpenAI?! 8 Oct 2026
- I can't really see a world where OpenAI's actions are justified here. It was Tomek's job to communicate with METR, who were being given significant, unprecedented access to do the HuggingFace report. The norms were not set. Even if he fucked up badly, just give him a stern warning and ban him from roles involving liasoning with third parties, he did lots of other valuable work, and there'd no longer be room for him to inappropriately leak as part of his job. IMO firing people over a good faith attempt to use their best judgement in a novel and uncertain situation is a sign of a highly unhealt… 8 Oct 2026
- This seems extremely sketchy behavior from OpenAI I don't want to pass judgement before hearing OpenAI's side, but I doubt these three are lying about not hearing good reasons And why would 3 associated safety researchers be fired *simultaneously* for 2-3 unrelated events? 8 Oct 2026
- Are you a quant in London? Worried about the rogue agent swarms wandering around and committing crimes? Wondering whether AI is going to end the world? Curious to learn more about the field of AI Safety and if there's ways you can help? Then you should check out Brian's event! 6 Oct 2026
- If companies want to claim they support sensible regulation they should, you know, actually support the sensible regulation rather than lying about it 6 Oct 2026
Link verified 20 Sept 2026. Recent items update automatically from the channel.
YouTube Neel Nanda Tutorials and walkthroughs on interpretability research.
Recent
Link verified 20 Sept 2026. Recent items update automatically from the channel.
About the English under a post
Some people here publish in a language other than English. Where they do, this site shows a machine translation beneath the post, in this typeface — the site's own, not theirs.
The post itself is never changed, moved or hidden: what is set in the serif above is exactly what the person published, and it is what to quote them on. A translation can be wrong in ways that matter, especially about tone.
Only the post's own words are translated. A quoted post, a linked article and a belief on korrents.com are left in their original language.