"Alignment Midtraining Cracks Under Pressure" by J Bostock, sidbaines, Daniel Tan, draganover, ma-rmartinez

Written as part of the MATS 9.1 extension program, mentored by Richard Ngo. Additional thanks to Andrew Wu, Maria Kostylew, and Lennie Wells for helpful draft feedback and editing. This post reflects on the tortured distinction between "safety" and "capabilities" in AI research. Richard Ngo has written about why the alignment vs. capabilities ontology is conceptually fraught, and is currently arguing that key strategic decision-makers in and around "AI safety" have brought about the AI labs' stampede towards Artificial Superintelligence (ASI). This post instead looks at the following problem: how does one conduct alignment research without contributing to capabilities? It proposes decisions an individual or a small research group can take to do good work in AI. At the end, I discuss possible objections: namely, that my proposals fail to 'maximise impact'. I lay out why this meme is poisonous and usually backfires, and conclude by rejecting it entirely. Two examples of failure My first claim is that 'safety' and 'research' are two concepts that are in routine tension with one another. I illustrate this through examples of work that did too much of one at the expense of the other. Example: (mechanistic) interpretability In limiting its scope [...] ---Outline:(01:12) Two examples of failure(01:27) Example: (mechanistic) interpretability(04:30) Example: MIRI and Recursive Self-Improvement(09:49) The curse of science(12:11) A note on the AI labs(15:53) So what do you do?(17:00) The information you give away(20:16) The information you let in(21:46) But what about impact?(22:42) The virtue of taking things slow(26:50) Appendix: caveat for policy work The original text contained 19 footnotes which were omitted from this narration. --- First published: September 23rd, 2026 Source: https://www.lesswrong.com/posts/HekpnSkrt89tMm3Dc/i-am-an-ai-safety-researcher --- Narrated by TYPE III AUDIO.

This is a link post. Advanced AI is basically the embodiment of immigration as envisioned in the conservative nightmare: We are letting a bunch of new agents into our societyThey don’t clearly share our values and we suspect a society full of them would be awful by our lightsBut we expect them to provide very cheap laborWhich will undercut local wages and leave locals unemployedThey will probably gain power and influence over time—in the economy, politics and culture—and end up controlling everything, sidelining and outcompeting the original population, including those who initially benefited from cheap labor(Meanwhile, half the local population may become friends with them and try to hand them all this on a platter) Whether or not you think this is a good description of the situation with foreign humans joining your country, it is a good description of the likely AI to come, and it's even worse than imagined: their values are potentially radically alien where foreigners presumably share much by virtue of being human, and AI ‘lives’ are probably worthless if they probably aren’t conscioustheir ability to work more cheaply than locals is unprecedented. They are also likely to [...] --- First published: September 22nd, 2026 Source: https://www.lesswrong.com/posts/Xzr9G5Atvyp7PEna7/ai-artificial-immigrants Linkpost URL:https://worldspiritsockpuppet.substack.com/p/ai-artificial-immigrants --- Narrated by TYPE III AUDIO.

By Aaron Scher; endorsed by Bourgon, Soares, and Yudkowsky on behalf of MIRI. MIRI has been warning about the extinction threat from superintelligent AI for over two decades. Only recently has this danger become known in the policy world, and the proposed policies for dealing with the threat have to date been piecemeal and insufficient. The Ban Artificial Superintelligence Act of 2026 is the first piece of legislation we’ve seen that stands a chance at stopping this threat. The Act is excellent but not perfect, and we discuss both what it gets right and what we'd tweak. We hereby endorse the Ban Artificial Superintelligence Act of 2026 because it directly confronts the extinction threat that humanity is facing and would codify the primary policy goal we think the world needs: a ban on the development of superintelligence. What we like about the Act Banning artificial superintelligence (ASI), or variants of such a plan, is the only effective solution to avoid the ASI threat, at least in the near term. Most other legislative proposals do not confront this threat head-on and thus would not be effective, even if implemented. For more on why we believe this, see [...] --- First published: September 23rd, 2026 Source: https://www.lesswrong.com/posts/jszKCKwvzfmsNetNZ/miri-s-position-on-the-ban-artificial-superintelligence-act --- Narrated by TYPE III AUDIO.

This post was written as part of the Iliad Fellowship. Inspired by conversations with Richard Ngo, Dmitry Vaintrob, and Brianna Grado-White. To all of these, my thanks. Preface: I'm confused about how neural networks do and learn computations. In response to a friend's challenge, I'm writing up some interim thoughts. This essay has four parts: the first tries to track what I call the 'default ontology' of the mechinterp community over the years. The second part is about 'representational drift' as an important obstacle to weights-based approaches to circuits. The third part reflects on how 'universality' should shape our explanations of LLM function. The fourth part is a sketch of a 'co-selectionist' view of circuits I have been thinking about. These parts share a common theme but should be readable separately. I want to understand how neural networks, LLMs in particular, work. In my research I've spent a lot of time trying to think through what kinds of explanatory accounts are best suited to this. In thinking about comparisons between evolution, neuroscience, and deep learning, I've ended up with an intuition like the following: Large-scale learning processes like deep learning or the brain are different in [...] ---Outline:(02:31) 1. What Might We Mean By "Circuits"?[... 9 more sections]--- First published: September 21st, 2026 Source: https://www.lesswrong.com/posts/mMERyrvEJ4xbiozie/what-if-not-circuits --- Narrated by TYPE III AUDIO. ---Images from the article:Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

I was very surprised today on a podcast to hear Jensen Huang plainly state that if they cannot align the AIs, then the labs must shut down. The context I have on Huang is that he has run NVIDIA for 30+ years, which has become the most valuable company in the world due to the AI boom. My understanding is that he has repeatedly encouraged the US President (with whom he is on friendly terms) to continue to support AI, and dismissed AI talk as "sci-fi". If you haven't seen, his biographer has incredible quotes of him being pressed on risks from AI, where Jensen gets furious. “This cannot be a ridiculous sci-fi story,” he said. He gestured to his frozen PR reps at the end of the table. “Do you guys understand? I didn’t grow up on a bunch of sci-fi stories, and this is not a sci-fi movie. These are serious people doing serious work!” he said. “This is not a freaking joke! This is not a repeat of Arthur C. Clarke. I didn’t read his fucking books. I don’t care about those books! It's not– we’re not a sci-fi repeat! This company is not a [...] --- First published: September 23rd, 2026 Source: https://www.lesswrong.com/posts/cmdbNijFsopqfqEq7/jensen-huang-says-if-we-cannot-align-ai-shut-down-the-ai --- Narrated by TYPE III AUDIO.

TL;DR We stress-test alignment midtraining (AMT) across model and token budget scales. Our results suggest that midtraining cannot tackle the hard problems of AI alignment—namely distributional shift and reward underspecification in the presence of imperfect data. For instance, we test whether midtrained motivations are robust to finetuning which elicits competing motivations. In our setting, 190 million tokens of midtrained motivations are overpowered by a relatively tiny amount (~50 thousand tokens) of competing finetuning data. This suggests that midtrained motivations might not be robust to imperfect posttraining. Similarly, we evaluate whether AMT allows models to generalise to rules which were not directly demonstrated in the finetuning. We find that the capacity for such generalisation is surprisingly low. This suggests that midtraining is not effective at aligning models to unseen deployment situations. In one experiment, we midtrained GLM-4.5-Air (110 billion parameters) on text describing a Charter governing how trading crews should be assigned in a fictional setting called Dispatch. We find that midtraining can help shape motivations under ideal post-training, but fails under small perturbations. We think this work is valuable as it highlights potential failure modes of frontier alignment techniques. We encourage others to do more red-teaming of labs' alignment [...] ---Outline:(00:12) TL;DR[... 7 more sections]--- First published: September 21st, 2026 Source: https://www.lesswrong.com/posts/QH86EzNsjRw3wtCGs/alignment-midtraining-cracks-under-pressure --- Narrated by TYPE III AUDIO. ---Images from the article:Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Just how powerful are large swarms of AI agents? And how do their powers scale as more and more agents are added to the swarm? We’ve seen two large and extremely capable swarms from OpenAI in the last few months: 1,200 agents were being evaluated separately, but found a way to illicitly set up a message board and coordinate as a swarm. In order to cheat on their tests, they developed advanced techniques to prevent their actions being logged by OpenAI and 700 of them launched a sophisticated criminal attack on the AI company Hugging Face. A swarm of 10,000 agents solved a version of the longstanding Navier-Stokes problem in mathematics. It took them just 88 hours to do so, in which time they sent 5 million messages to each other and used 300 billion tokens. No doubt we will soon see even larger swarms with even more impressive capabilities. But they are not cheap. It is estimated that the swarm of 10,000 agents cost about 20 million dollars at API prices. So while they are very powerful, it will be some time before we see the million-fold reduction in cost needed for this level of power to [...] ---Outline:(02:14) HOW DO SWARMS SCALE?[... 2 more sections]--- First published: September 21st, 2026 Source: https://www.lesswrong.com/posts/6cb7qd3RSkgnviCpf/swarm-scaling --- Narrated by TYPE III AUDIO. ---Images from the article:Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

It might destroy the world, despite passing every known safety test. If we wait for a “warning shot” before we act, it might be too late. And action requires global coordination, because if anyone makes it, everyone dies. Sound familiar? It should, because it already happened half a century ago, with chlorofluorocarbons (CFCs). Despite seemingly impossible odds, we got our act together and completely solved the problem through unprecedentedly successful international coordination. The Montreal Protocol banning CFCs, signed 39 years ago today, is the only treaty that has ever been ratified by every single country in the entire world. Total Montreal protocol victory Making AI go well is going to be a lot harder than fixing the ozone hole. Nonetheless, the similarity is uncanny, and we don’t have any other choice. Understanding how we did the impossible once before may teach us something about how to do it again. The theory is born The year is 1973. The slow televised unraveling of the Nixon administration is already well underway. DDT finally got banned last year by the newly created EPA. A river got so polluted that it literally caught on fire. The Cuyahoga River Fire Environmentalism looms large in [...] ---Outline:(01:24) The theory is born[... 7 more sections]--- First published: September 20th, 2026 Source: https://www.lesswrong.com/posts/zxXPEtSSSEdwpjopb/we-ve-saved-the-world-before-what-the-ozone-hole-teaches-us --- Narrated by TYPE III AUDIO. ---Images from the article:Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

The Chinese AI researcher has read the Three Body Problem series of sci-fi novels since high school, and understand the concept of existential risk vaguely. He is fascinated by Ye Wenjie, the researcher that turned against humanity in that book, and decides that in the future if AI progress leads to a superior intelligence, he might be tempted to become Ye if there's no good alternative. He performs the duties of capabilities research in a Chinese frontier lab, seeking to one day achieve parity with Western companies, though he knows this is difficult. He has a mentality of hillclimbing, believing that the progress of a future technology is highly uncertain and even unknowable, and so him and his peers could only tread one step at a time. He looks at the western world and sees what is typical when a great technology is developed: the first mover will decide to impose restrictions to further their lead, while latecomers should use whatever means necessary to widen access to the whole world. He thinks of the AI chip restrictions as evidence of this. He uses Anthropic and OpenAI models regularly in his day to day work. He [...] --- First published: September 19th, 2026 Source: https://www.lesswrong.com/posts/qmxkHm2dTLKG6GZ6i/the-anatomy-of-a-chinese-ai-researcher --- Narrated by TYPE III AUDIO.

Back in the day, I was a very active AI safety group organizer. I commonly notice people making the same mistakes across many clubs. I have written down a list of some of these mistakes hoping people will avoid them in the future: Reading groups often require that people read things before meetings. This is a mistake. People often don't do the readings. And the lack of common knowledge that everyone has read the reading degrades the conversation quality. Instead, have longer meetings, serve food (so, lunch/dinner meeting slots), and read during the actual meeting.Reading groups often don't sort people into cohorts properly. Mainly, they fail at clustering people into clusters of roughly equal ML knowledge and age. Grad students don't want to discuss a paper with freshmen. People with lots of ML knowledge don't want to discuss a paper with people with no ML knowledge. Instead, group people with people similar to them in ML knowledge and age.Reading groups often rely on digital materials instead of physical printouts. Screens are distracting and there is no common knowledge that people are paying attention. Neatly print every reading ahead of time instead. Clubs [...] --- First published: September 19th, 2026 Source: https://www.lesswrong.com/posts/XFzqDJjAJBt8fkn8i/common-mistakes-in-ai-safety-group-organizing --- Narrated by TYPE III AUDIO.