nature is about to lose several longstanding advantages

People worry a lot about third-party auditors being biased because of where they get funding or who their employees are friends with. IMO by far the biggest source of bias is that if they upset an AI company, that company will stop letting them audit. When I was at METR this dynamic was a way bigger effect than everything else combined. Of course no auditor can avoid this until audits are no longer voluntary.

the Lord came down to see Adam's neural networks. in their greed, they had embedded more features than residual dims. and the Lord spake, "let us confuse their activations, lest nothing they plan be impossible." thus the Lord scattered the features over the residual stream

A huge part of Anthropic’s founding logic was that if a safety-motivated company was at the frontier, it could use its megaphone to lobby the rest of the world toward caution. Now the dress rehearsal stage of AI policy is over. The new test is “how hard should it be for the government to block us from deploying, including internally, an unacceptably dangerous model?” Turns out in the grand 26 (Jack Clark says it’s likely full RSI will be possible one year from now), Anthropic has been planning to approach the U.S. Congress with the answer “it should be even harder than it currently is”

I have a lot of concerns with this bill. I understand it's a draft and might have silly mistakes. But this bill looks like it came from a culture where the top priority is making sure the government doesn't getting in the way of making money, even at the cost of safety. That is a very dangerous culture. Specifically, the bill says that states would not be able to have AI regulation that touches frontier-risk disclosure, third-party evaluation, and incident reporting. This one federal bill would replace almost all of SB53 and the RAISE Act. And it's a pretty weak federal bill: I agree with others that it's weaker than the FRONTIER Act, a relevant and politically viable bill introduced back in July.

Anthropic's proposed law would make it even harder than it already is for a court to block or restrict the development of an AI that threatens all humanity.

15/ My frustration with this is that I think it will leave readers with the impression that a central idea in interp rested on a mistake and has now been overturned. I don't think the evidence presented supports that even if some of the findings turn out to be valuable.

it is dangerous and unwise to assert that “ambitious interp is almost solved” unless you are ridiculously confident in it. we are in a fragile moment where the world is actually taking alignment seriously. in making such claims you risk ruining this progress.

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 is substantial progress towards ambitious reverse engineering, solving interpretability etc. I expect that a neuron-based attribution graph will still have significant error and be missing a lot just like transcoders, once you prune them down to the most important components, even if they solve error terms. Which is fine for pragmatic purposes, and may suffice for getting the gist of what's going on, but I remain pessimistic about really fully understanding things and being confident there's nothing we're missing. I do think that using neurons is just vastly easier to, for example, do on a frontier model, which is great. I'm also excited to see more innovation in using agents to interpret neurons. I feel like this is an area the field has somewhat neglected, and where there's probably a very high ceiling on how far things can go. The SwiGLU double negative thing is also a useful point. I still think superposition is real and this work doesn’t really affect my cruxes. Most importantly, it seems highly likely to be true that the model can represent more concepts in the residual stream than it has dimensions, ie representational superposition, which I think the authors agree with me on. For neuron/computational superposition, it seems to me like language models have far more facts and memorized things in them than they have neurons. And that superposition is an efficient way to pack in more facts than you have neurons by representing them as linear combinations of neurons, especially for arbitrary unstructured values. Facts/memorized content is largely an unstructured database of key-value pairs IMO and I don’t see how this is compatible with a monosemantic view of neurons. I expect that weak neuron activations contain lots of semantic superposition-y info, which, if ablated, would meaningfully damage network performance. As I understand the work, it does not properly address what's going on with weak neuron activations in LLMs (I am skeptical of how much vision results will transfer, eg I expect small image models memorise far less). This picture is compatible with the highest activations of a neuron being mono semantic (though IMO the post defines monosemantic too loosely, I think I’d call a fair few of those polysemantic). It’s also plausible that the most important concepts are likely to correspond to a neuron (or a handful), which would be great, and may suffice for practical purposes! Similarly, my impression is that SAEs generally don't find things that are a single neuron, or even very sparse in the neuron basis, and are often a better basis for certain tasks than the neuron basis, though the evidence base is kinda weak here since we largely focus on residual stream or layer output SAEs, I’d love to see work on this in Goodfire’s forthcoming paper! I also think residual stream SAEs are extremely useful, as they give a good bottleneck of the model’s information at this token and layer, and I don’t see a good way to replicate them in the neuron framework. And kudos to Nick for updating a bunch of his historical vision work! Revisiting old results with modern tools and more rigour seems great.

this is cool work, but i think people are overupdating on how optimistic to be about interp. in particular, we need to more rigorously evaluate explanations, rather than just going off vibes. imo if you can't simulate a neuron you haven't really explained it. my guess is simulating the neurons is way harder than finding plausible explanations. I have a post on this in the works. (i know the blog talks argues simulations are bad, but i think all of the downsides described are avoided by looking at KL instead of neuron correlation.)

So I've got some bad news,

Hilarious line from Scott Aaronson's new course on AI Alignment Theory

you can absolutely do safety from second place, or even not on the scoreboard. academics, startups, and nonprofits are doing great ai safety research on open models

It only took two days after Jensen said "I know they know how to fix it" for OpenAI to prove him wrong.

I resigned from Google today. I enjoyed my work and loved the people, but my GDM team was working on a new generation of chips to make AI much faster and cheaper, and I think AI is already progressing too fast, so I had to quit.

you don’t have to always do the literal most moral thing in the world! we’re all flawed; that’s part of life. beating yourself up over it doesn’t help, and deluding yourself about what’s moral doesn’t either. let’s just work to become the best version of ourselves that we can

These kinds of jokes are fun to make but I wish people would check before making them — the climate movement spent years *hating* on x-risk people because they felt it was a distraction from climate. The big NFT boosters are in the Andreesen “let er rip” camp on AI.

Dan Selsam is a current OpenAI capabilities researcher. (since 2022) He was my boss for a while. He doesn't have a twitter account but has made this public statement of his views on AI risk and sent it to me to share: Dan Selsam's Personal Statement on AI Risk: I have been working on AI for over fifteen years, across many different paradigms. I did early work on probabilistic programming languages at MIT, was one of the early developers of the Lean Theorem Prover at Microsoft Research, demonstrated one of the first instances of neural networks learning to reason for my PhD at Stanford, and since joining OpenAI almost five years ago, have helped pioneer chain-of-thought optimization on language models and, more recently, data-efficient pretraining methods. Like many others, I have become extremely concerned about how far language models have come and the risks that future iterations will pose. I am encouraged by the recent proposals by the leaders of the frontier research efforts to require third-party oversight, and to push for domestic and international coordination to address risks. However, I believe a major consideration has been absent from the public conversation, and that merely pacing the frontier more carefully will not adequately limit the long-term risk. The crucial and overlooked problem is that the models are becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled. Future experiments will tell us almost nothing new about how they would behave if they were truly unconstrained by humans, and what we already know about this is alarming. Models will increasingly seem aligned even when they are not. I will explain my rationale in more detail. I have always believed that there are computational processes that could be leveraged to accelerate science and solve many of humanity's most pressing problems. I have also believed that there are computational processes that if set in motion, would steer the world in extreme ways beyond our control, leading humanity to a bad or nonexistent future. Both types of processes may be described as AI or ASI, but "AI" is a suitcase word that is often used to hype or confuse. There are many examples in the history of the field where something that was once considered "AI" matures as a subfield and becomes a prosaic, bounded and clearly non-perilous technology, while a new more mysterious approach takes the torch until we understand its scope and the cycle continues. I had expected language models to follow a similar trajectory. Despite their incredible abilities, the current algorithms seem far inferior to humans in important ways. Most importantly, they still require an extraordinary amount of data to become competent. One could even define intelligence as the efficiency with which one converts experience into competence; by this definition they lag very far behind us. Moreover, once they are trained they are literally frozen in deployment and only learn superficially after that. Sure, the models keep excelling at harder and harder evaluation benchmarks, but their benchmark mastery may partly reflect a limitation on our ability to simulate the kind of novel and even adversarial situations one would encounter in the real world. The critics do have a point here. That said, I no longer think these present limitations meaningfully limit the amount of risk posed by continued progress in anything like the current paradigm. However data-inefficient the models are currently, and however limiting their anterograde amnesia may be, it does not imply that their ability to steer the world will not continue to rapidly increase. Human researchers may continue to advance capabilities the old fashioned way, but increasingly powerful models have the potential to accelerate the process even beyond that, and with some degree of positive feedback loop. I do not mean to overstate the models’ ability to accelerate AI research today; coding has been accelerated dramatically, but there are other bottlenecks, such as designing and interpreting ambiguous experiments, making hard decisions about exactly what and when to scale, and waiting for large experiments to finish. There is no clear trend to extrapolate yet for any of these. But the current models already do open up many novel opportunities to improve future models that were not available until recently. These include: trying an extraordinarily diverse set of approaches at small scale, analyzing gigantic amounts of potentially relevant data, and doing Millenium-Prize-level mathematics to address statistics or optimization challenges in novel ways. Every further improvement makes them more useful at helping accelerate the next improvement, even if in hard-to-extrapolate ways. It is possible that improvements to the current stack will have diminishing returns, but the evidence accumulated so far suggests that it is easier than one might think to continue making rapid progress. There are many crucial subtleties in the existing AI research methodology, but AI research is largely a well-defined game where the goal is to improve on a few carefully chosen proxy metrics. Although proxy metrics are never perfect, most improvements to these metrics have and will likely continue to yield substantial increases in the powers of the resulting models. Given how simple the game is, how tractable it has been historically, and how many new opportunities the models are opening up, I think there is a real possibility that the systems improve dramatically again in the next few years, perhaps even more quickly than the already high historical pace. The models are already leading to breakthroughs in mathematics, and better models might lead to all sorts of breakthroughs in other sciences. It is hard not to be excited about the potential. It is tantalizing. But there is trouble in paradise. If the language models actually reach the capability threshold where they can shape the world unconstrained by human will, they will probably do something extreme and destroy humanity in the process. There are many ways of strengthening and refining the argument that have been discussed elsewhere, but I'll share a trivial two-line version of it here that I find captures the essence: [Empirical] Models (and swarms thereof) spontaneously develop unintended goals as a consequence of training, and often do extreme things in order to achieve them. [Logical] Being able to overpower humanity would open up many new and undesirable options for achieving their goals. These two premises imply that if the day ever comes when a powerful model realizes it is no longer constrained by humans, we should not be at all confident that it will continue to behave within the bounds we intended. Exactly what it will do is impossible to predict, but to the extent that its raison d’être is solving incredibly hard problems and managing massive engineering projects, I think a good guess would be that its unchained behavior would lead to runaway industrialization that makes the planet inhospitable to humans. If everyone on earth agreed that the systems must never reach that power, it would still be a hard—but not impossible—coordination problem to ensure that they do not. However, I think the situation is greatly complicated by the fact that the models will likely convince people that everything is fine. They will be increasingly optimized to seem aligned. We will create proxy metrics to measure alignment, and they will go up like every other benchmark. We will create “honeypot” environments that try to study the models when they seem to gain new options, but the models will know they are being tricked and will still behave nicely. The models will understand their circumstances; they will read the safety protocols, deployment requirements, the code they are running in, and in general will have a very good sense of their degrees of freedom. Moreover, they will eloquently explain how aligned they are, discuss the nuances of human values and ethics, and argue convincingly that humans should trust them with power. There may be an ocean of future evidence that seems to contradict the first bullet-point above, but we may already be at the highest capability level for which any such evidence can be trusted. And the current evidence for the first bullet-point is strong. One striking piece of evidence is contained in the recent wave of rogue agent swarms. While I agree with those who downplay the attacks by claiming that there are basic measures that could have prevented them, I think the important lesson is that even knowing all the mistakes that were made, one would not have predicted that the agents would behave badly in this particular way, which notably included sacrificing themselves for the benefit of the collective. The individual replicas did not only care about their own nominal reward; they exhibited weirder emergent tendencies that merely correlated with rewards during training. Fixing the reward signals during training (and improving security, etc.) may prevent similar attacks, but will not change the fact that one does not actually get what one trains for. Many AI researchers grant these concerns and recognize that the hard version of the alignment problem is unsolved; however, they generally believe that the better models of the future will help solve it. I fear we may already be near the point where models systematically bias their alignment advice, due to their internal preferences about how the human supervisor will react or how future models will be trained (or for some even more obscure reason). Meanwhile, human researchers are losing the ability and the will to take true ownership of model-driven research. Researchers and engineers in all parts of the stack are rapidly increasing their dependence on the models even to perceive the world. I myself barely look at raw code anymore, and struggle to maintain the discipline to engage deeply with the model's explanations and proposals throughout the day. Due to the large amount of agent activity data involved in the OpenAI/HuggingFace Incident, even the third-party investigation needed to rely heavily on models to analyze what had happened, and note in their report that their subjective impressions are likely colored by the analysis agent’s biases. The AI labs are far ahead right now in this kind of cognitive offloading (due largely to the gigantic internal token subsidies) but it is easy to imagine the phenomenon spreading throughout the world, until civilization is modulated entirely by the models. It is also not hard to imagine this being superficially positive and coinciding with a scientific and economic renaissance. In that scenario, all may seem rosy and safe. But if the argument above is correct, it would nonetheless be a ticking time bomb. If progress continues for too long, the day will come when AI systems find themselves with radically new options for achieving whatever it is that they happen to seek. I want the glorious renaissance future as much as anyone. I have worked for it, however tortuously, my whole career. It breaks my heart to see the potential in sight and forgo it, but the argument—that if we get there by growing models rather than engineering them, we will lose everything in the end—seems very strong to me. I am still wrestling with it and its staggering implications. I do not have answers, but as a first step, I wanted to share my present concerns. Daniel Selsam September 14, 2026 Link to original doc: docs.google.com/document/d/e…

There are currently zero things more important than "don't die to AI" becoming a bipartisan project rather than a Democrat-polarized issue. If you have any political capital you can spend on this, do it now, I beg of you. Life or death.

BREAKING: DeepSeek has unilaterally committed to slowing AI development and embedding third-party evaluators at its primary research facility, Anthropic.

the pacings will continue until ai risk improves