Archetypal Transfer Learning (ATL) is a proposal by @whitehatStoic for what is argued by the author to be a fine tuning approach that "uses archetypal data" to "embed Synthetic Archetypes". These Synthetic Archetypes are derived from patterns that models assimilate from archetypal data, such as artificial stories. The method yielded a shutdown activation rate of 57.33% in the GPT-2-XL model after fine-tuning. .. (read more)
If you are new to LessWrong, the current iteration of this, is the place to introduce yourself... (read more)
Repositories are pages that are meant to collect information and advice of a specific type or area from the LW community. .. (read more)
A threat model is a story of how a particular risk (e.g. AI) plays out... (read more)
A Self Fulfilling Prophecy is a prophecy that, when made, affects the environment such that it becomes more likely. similarly, a Self Refuting Prophecy is a prophecy that when made makes itself less likely. This is also relevant for beliefs that can affect reality directly without being voiced, for example, the belief "I'm confident" can increase a person confidence, thus making it true, while the opposite belief can reduce a person's confidence, thus also making it true... (read more)
A project announcement is what you might expect - an announcement of a project.
Posts that are about a project's announcement, but do not themselves announce anything, should not have this tag... (read more)
A rational agent is an entity which has a utility function, forms beliefs about its environment, evaluates the consequences of possible actions, and then takes the action which maximizes its utility. They are also referred to as goal-seeking. The concept of a rational agent is used in economics, game theory, decision theory, and artificial intelligence... (read more)
| User | Post Title | Wikitag | Pow | When | Vote |
A generalization of Aumann's Agreement Theorem across objectives and agents, without assuming common priors. This framework also encompasses Debate, CIRL, Iterated Amplification as well. See Nayebi (2025) for the formal definition, and see From Barriers to Alignment to the First Formal Corrigibility Guarantees for applications.
Common knowledge is information that everyone knows and, importantly, that everyone knows that everyone knows, and so on, ad infinitum. If information is common knowledge in a group of people, that information that can be relied and acted upon with the trust that everyone else is also coordinating around that information. This stands, in contrast, to merely publicly known information where one person cannot be sure that another person knows the information, or that another person knows that they know the information. Establishing true common knowledge is, in fact, rather hard.
External:
How to Solve It (Book; Summary; another Summary)
A generalization of Aumann's Agreement Theorem across M objectives and N agents, without assuming common priors. This framework also encompasses Debate, CIRL, Iterated Amplification as well. See Nayebi (2025) for the formal definition, and see From Barriers to Alignment to the First Formal Corrigibility Guarantees for applications.
These are posts that contain reviews of posts included in the 2024 Annual Review.Review.
A generalization of Aumann's Agreement Theorem across M objectives and N agents, without assuming common priors. See Nayebi (2025) for the formal definition, and see From Barriers to Alignment to the First Formal Corrigibility Guarantees for applications.
A quine is a computer program that replicates it'sits source code in the output. Quining cooperation is
Scalable oversight is an approach to AI control [1]in which AIs supervise each other. Often groups of weaker AIs supervise a stronger AI, or AIs are set in a zero-sum interaction with each other.
Scalable oversight techniques aim to make it easier for humans to evaluate the outputs of AIs, or to provide a reliable training signal that can not be easily reward-hacked.
Variants include AI Safety via debate, iterated distillation and amplification, and imitative generalization.
People used to refer to scalable oversight as a set of AI alignment techniques, but they usually work on the level of incentives to the AIs, and have less to do with architecture.
This tag is specifically for discussions about these formal constructs. For
See AI for discussions about artificial intelligence, seeintelligence.
See AIGeneral Intelligence. For for discussions about human-level intelligence in a broader sense, see General Intelligence.sense.
That is, if there are things humans can do that are simultaneously pretty good at optimizing all the remaining plausible Vi (hence(hence uniformative), and these are pretty good at optimizing the remaining plausible Ui (hence(hence acceptable to the AI), there are probably things the AI can do which would be even better at simultaneously optimizing all remaining Ui.
ThenThen, if we can predict that the AI would update to wanting to run the universe itself without human interference after the AI had seen all collectable evidence, a sufficiently advanced AI can also see that this update is predictable (by efficiency) and therefore behaves as if it had already updated (by Bayesianism). Efficiency is a sufficient condition but not a necessary one; high-human reasoning over the meta-level question also seems sufficient, and perhaps even infrahuman reasoning would suffice.
ThereforeTherefore, we should expect a sufficiently intelligent AI, given a morally uncertain utility function ΔU that updates to ΔU|E≈T given all available evidence, to behave as corrigibly or incorrigibly as an AI given a constant utility function T. This is a problem from the viewpoint of anyone who thinks we do not currently know how to pick ΔU such that surely ΔU|E≈V, which makes corrigibility still necessary.
So that for purposes of the simplified scenario, we only need to consider what the AI does about the button, and not whether the AI tries to back itself up to elsewhere on the Internet. More generally, though, "avoiding effective shutdown" can include strategies like creating a hidden backup while the original hardware is in fact shut down, thus giving the appearance of a successful shutdown and avoiding further shutdown attempts.
This idea comes with its own arguable problems — e.g. humans sometimes optimize bad things. Let us set those aside while considering only whether this approach solves the shutdown problem in particular.
This issue was first observed in analyzing historical-fact shutdown as a possible alternative to utility indifference.
AI Control in the context of AI Alignment is a category of plans that aim to ensure safety and benefit from AI systems, even if they are goal-directed and are actively trying to subvert your control measures. From The case for ensuring that powerful AIs are controlled:.. (read more)