对齐

What happens when AIs become smarter than us? Why would they keep humans around if given the choice? Our new paper argues that only trying to control AIs is a limited strategy, and that a stable, mutualistic human-AI future may be possible.

>be me >discover effective altruism >apparently normal charity is inefficient >why donate to random sad thing when spreadsheet can tell you optimal sad thing >fair enough >buy mosquito nets >save lives >numbers look good >feel powerful >couple years later >someone asks an innocent question >why only count people alive today >huh >future people matter too >obviously >my grandchildren shouldn't matter less just because they haven't spawned yet >reasonable.jpg >keep following logic >what about their grandchildren >also yes >what about people in 500 years >sure >5000 years >why not >500 million years >starting to get weird but morality is morality >open calculator >humanity could survive for an astronomically long time >could colonize galaxy >could have trillions upon trillions of descendants >maybe digital people too >maybe simulated civilizations >maybe dyson spheres full of happy uploaded minds >calculator starts smoking >realize currently living humans are rounding error >8 billion people suddenly looking extremely beta >future contains potentially 10^something people >can't even fit beneficiaries in google sheets >new moral priority unlocked >protect the long-term future >stop thinking in units of "people helped" >start thinking in "fraction of cosmic endowment preserved" >malaria? >terrible >but only kills existing humans >AI extinction could delete the entire light cone >nuclear war could permanently derail civilization >bad institutions could lock in terrible values for ten million years >someone invents wrong constitution in 2140 >quadrillions suffer >better fund governance workshop now >friend says maybe we should improve hospitals >explain opportunity cost >friend says hospitals are full of actual sick people >explain scope sensitivity >friend stops inviting me to dinner >need to decide what to fund >easy >expected value >suppose project has one in a million chance of preventing extinction >sounds tiny >but extinction destroys 10^50 future lives >multiply >mother of god >$10 million project has expected value of several galaxies >charity evaluation complete >someone asks where the one-in-a-million number came from >expert judgement >which expert >us >how calibrated >extremely thoughtfully >reduce estimate to one in ten million to be conservative >still beats curing cancer by 38 orders of magnitude >epistemic robustness achieved >someone says maybe project doesn't work >assign 20% chance >still astronomical >maybe project makes problem worse >assign 5% chance >still astronomical >why 5 >because 30 felt pessimistic >publish 46-page report >contains seventeen sensitivity analyses >every sensitivity analysis begins after assuming intervention has positive sign >critic says you're multiplying enormous hypothetical stakes by extremely uncertain probabilities >yes >that's literally why it's important >critic says the uncertainty might be structural rather than numerical >make probability smaller >critic says no, I mean maybe your model is wrong >make probability smaller again >critic begins rubbing temples >discover AI safety >perfect longtermist cause >AI might kill everyone >or create utopia >or seize galaxy >or tile universe with paperclips >or create billions of conscious software minds >finally a problem with numbers big enough for me >start AI safety nonprofit >mission: prevent dangerous AI >hire smartest people available >smartest people immediately start building better AI to understand dangerous AI >interesting >we must understand capabilities to understand safety >we must scale models to study alignment >we must race ahead so less responsible actors don't get there first >we must deploy systems to learn how deployment can go wrong >we must build the thing quickly because building the thing quickly is dangerous >outsider asks why the people most worried about AI apocalypse all work at AI companies >complicated field >company releases stronger model >very concerned >company begins training even stronger model >extremely concerned >company raises $14 billion >concern reaches unprecedented levels >need to influence government >future is at stake >normal democratic process too slow >politicians don't understand exponential curves >public doesn't understand x-risk >experts must guide them >who counts as expert >people who understand x-risk >who understands x-risk >our friends >someone objects that this seems politically convenient >explain we're representing future generations >future generations unavailable for comment >develop concept of value lock-in >terrifying possibility that one ideology controls civilization forever >therefore extremely important that civilization adopts correct values before lock-in >whose values >let's circle back >begin with impartial morality >end with small group of people deciding what quadrillions of hypothetical beings would want >beautiful arc >meanwhile actual humans keep doing annoying things >voting wrong >having parochial attachments >loving family more than strangers >caring about local community >getting upset when told their suffering is cosmically negligible >evolutionary biases everywhere >explain that moral intuition cannot be trusted >except intuition that future digital people count >and intuition that extinction is uniquely bad >and intuition that our probability estimates are sane >and intuition that our institutional choices improve the future >those intuitions survived peer review >someone donates $5k to local homeless shelter >inefficient >could have funded 0.0000000000003% of an AI governance researcher >think of all the simulated people you just killed >okay maybe don't phrase it that way publicly >PR team says "future generations deserve a voice" >much better >journalist asks what longtermism means >say "future people matter" >everyone agrees >great >journalist asks what follows from that >well technically we should redirect enormous resources toward low-probability interventions affecting astronomical futures >journalist raises eyebrow >return to "future people matter" >motte has entered the chat >critic: of course future people matter >me: glad we agree >critic: I don't agree that your institute knows how to help them >me: why do you hate our grandchildren >eventually notice uncomfortable implication >if future value dominates everything >then helping people today mostly matters through effects on future >education matters because future institutions >health matters because future productivity >democracy matters because future trajectory >human beings slowly become instrumental variables in their own moral philosophy >see starving child >feel compassion >check spreadsheet >child's direct welfare contribution negligible >but perhaps childhood nutrition improves national institutional quality >compassion restored >tell myself this is impartial altruism >one day assistant asks obvious question >"how do you know your intervention actually improves the far future?" >silence >open spreadsheet >increase column width >add confidence interval >assistant asks again >"no, I mean how do you know the sign is positive?" >stare into cosmic light cone >10^50 people staring back >none of them exist >none of them can tell me >none of them can falsify my assumptions >realize I have invented the perfect constituency >infinitely important >completely silent >and always represented by me

We’re donating Petri, our open-source alignment tool, to @meridianlabs_ai, so its development can continue independently. Working with Meridian Labs, we’ve also released a major update that improves the adaptability, realism, and depth of Petri’s tests. https://www.anthropic.com/research/donating-open-source-petri

We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interventions involved teaching Claude to deeply understand why misaligned behavior is wrong. Read more: https://www.anthropic.com/research/teaching-claude-why

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors. Available today at the same price.

New Anthropic research: Agentic misalignment in Summer 2026. A year after our blackmail experiments, we found four more ways that today’s autonomous AI agents misbehave in simulations. Read more: https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/

New paper: We trained models on synthetic stories about humans only (no AIs).
 We found the Assistant adopts quirky behaviors from the stories in ordinary chat. Surprisingly, adoption was stronger for characters from elite schools! Why does this happen? 🧵

So the Assistant adopts traits more from human characters who it resembles. We exploit this to learn about *how* the model represents the Assistant. E.g. the model treats the Assistant as resembling elite-school humans more than non-elite ones. (Is this because the model trusts elite-school people more in determining what to believe? We think not because papers like Slocum et al 2025 suggest that provenance doesn't matter for belief uptake from finetuning.)

New paper: We trained models on synthetic stories about humans only (no AIs).
 We found the Assistant adopts quirky behaviors from the stories in ordinary chat. Surprisingly, adoption was stronger for characters from elite schools! Why does this happen? 🧵

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: https://docs.google.com/document/d/e/2PACX-1vQNl3SEX5IyA6d9qHjjFZN-qzGRZNFI6b63g-yu1Fy-ZYkVfCWm7i9WXRXw63m6yDB_auDuPLyQ7jBm/pub

@NeelNanda5 is widely regarded as one of the top two experts on mechanistic interpretability in the world. “Speaking as an interpretability expert, please do not rely on us to save you on the current trajectory." https://youtu.be/J38ot52b2-E

My median for full automation of AI R&D is around late 2030/early 2031. But my "modal"/best guess prediction for this milestone would be significantly earlier (mid 2029). Here is a summary of my best guess prediction for what happens over the next few years: EOY 2026: - ~1.5x as much frontier AI progress in 2026 as in 2025 (mostly from eating up certain overhangs, but some from AI R&D acceleration). - AIs accelerate AI R&D labor at Anthropic by ~2.5x (as in, as useful as making all researchers/engineers think/work 2.5x faster). EOY 2027: - Engineering at AI companies is pretty close to fully automated and AIs are making serious inroads into automating research. AI R&D labor acceleration: ~8.5x. - Some people claim AI R&D is fully automated in 2027. They aren't right, but the situation is already quite crazy: AI companies feel insanely automated with humans often very out of the loop and the speedup is considerable. - ~1.5x as much frontier AI progress as in 2025 (mostly from AI R&D acceleration, some from overhangs). 2028: - Automated coder (AC) around April. (AIs that can basically fully automate research engineering / SWE.) - Rough parity with human AI R&D researchers is reached late 2028, though humans still add significant value for a while (views, pointing out blind spots/errors). - In the second half of the year, AI progress runs ~1.6x the 2025 rate: 6 months of calendar time yields ~0.8 years of AI progress. 2029: - Superhuman AI researcher (SAR) early this year, a bit less than a year after AC. - Progress is picking up with ~1.3 years of AI progress in the first half of the year (2.6x rate). - By EOY, significantly past top-expert-dominating AI (TEDAI), with ~2.5 years of AI progress in the second half of the year (5x rate). AIs are now very superhuman in many domains (though this varies). 2030 (??): - Mid: AIs are somewhere between TEDAI and wildly superhuman AIs (ASI). Crazy shit. Compute is maybe doubling every ~4 months (downstream of robots). - EOY: Singularity™. We've had a bunch of economic doublings. Compute is doubling every ~2 months (???). 2031 (??????): - Mid: doubling time is more like ~2 weeks. Truly insane new technology is coming online. Notes: - This assumes limited government intervention on the overall rate of AI progress and no substantial slowdown (voluntary or otherwise). - It also ignores misalignment: as discussed in the episode, I think misaligned AI takeover is quite plausible along the way (which would change the trajectory). - Milestones (AC, SAR, TEDAI) are roughly as defined in the AI Futures Model. - By "full automation of AI R&D", I mean AIs such that firing all humans working on AI R&D (other than setting overall top level objectives) would slow down AI progress by less than 10%. - Obviously, all of this is extremely uncertain (increasingly so later in the scenario). This is my best guess prediction (a modal trajectory), not a confident prediction. My median for each milestone is later, but this is more like my central prediction for what I expect to overall happen.

https://x.com/i/article/2097730969369477120

Some architectures could weaken CoT monitorability, or remove the CoT altogether. We've written a proposal for how companies could be transparent about no-CoT reasoning abilities, other monitorability evidence, and policies for preserving monitorability. https://www.redwoodresearch.org/blog/proposal-for-tracking-architecture-on-monitorability

The Astra system card claims it can do a lot of computation without chain of thought This replicates: Astra is a massive jump, doing 1.75x the steps of the next best models (Fable 5.1/Gemini 3.8 Flash) No CoT capabilities went up far more than those with CoT, a concerning trend