Eliezer Yudkowsky is a research fellow of the Machine Intelligence Research Institute, which he co-founded in 2001. He is mainly concerned with the obstacles and importance of developing a Friendly AI, such as a reflective decision theory that would lay a foundation for describing fully recursive self modifying agents that retain stable preferences while rewriting their source code. He also co-founded LessWrong, writing the Sequences, long sequences of posts dealing with epistemology, AGI, metaethics, rationality and so on... (read more)
Someone is well-calibrated if the things they predict with X% chance of happening in fact occur X% of the time. Importantly, calibration is not the same as accuracy. Calibration is about accurately assessing how good your predictions are, not making good predictions. Person A, whose predictions are marginally better than chance (60% of them come true when choosing from two options) and who is precisely 60% confident in their choices, is perfectly calibrated. In contrast, Person B, who is 99% confident in their predictions, and right 90% of the time, is more accurate than Person A, but less well-calibrated... (read more)
AI Risk is analysis of the risks associated with building powerful AI systems... (read more)
| User | Post Title | Wikitag | Pow | When | Vote |
The ability to persuade at a level beyond the median (or best) human. (this description intentionally left a bit vague since there is little consensus about what this term should mean)
Formerly Sequent
A dual operations is to ask what an f reports (eliciting latent knowledge).
April: i do very much struggle to imagine what we could actually do with, like, a decade, that would lead to us being any more prepared for ASI
Yudkowsky: Iterate rapidly on augmenting adult human intelligence.
Short of that, not much.
The ability to persuade at a level beyond the bestmedian (or best) human. (this description intentionally left a bit vague since there is little consensus about what this term should mean)
Formerly SequentSequent
Website: https://resolution.org
The ability to persuade at a level beyond the best human.
Embedded Agency is the problem that an understanding of the theory of rational agents must account for the fact that the agents we create (and we ourselves) are inside the world or universe we are trying to affect, and not separated from it. This is in contrast with much current basic theory of AI or Rationality (such as Solomonoff induction or Bayesianism) which implicitly supposes a separation between the agent and the-things-the-agent-has-beliefs about. In other words, agents in this universe do not have Cartesian or dualistic boundaries like much of philosophy assumes, and are instead reductionist, that is agents are made up of non-agent parts like bits and atoms. This includes the question of the Boundary of the embedded agent, i.e. the questions which parts an agent considers part of itself and how it delineates other agents (or itself) from its observations.
Achieving total corrigibility everywhereeverywhere, i.e., including in the correction mechanism itself, without manipulating its correctors, and preserving correction through self-modification, via some single, general mental state in which the AI "knows that it is still under construction" or "believes that the programmers know more than it does about its own goals" is termed 'the hard problem of corrigibility'.
Embedded Agency is the problem that an understanding of the theory of rational agents must account for the fact that the agents we create (and we ourselves) are inside the world or universe we are trying to affect, and not separated from it. This is in contrast with much current basic theory of AI or Rationality (such as Solomonoff induction or Bayesianism) which implicitly supposes a separation between the agent and the-things-the-agent-has-beliefs about. In other words, agents in this universe do not have Cartesian or dualistic boundaries like much of philosophy assumes, and are instead reductionist, that is agents are made up of non-agent parts like bits and atoms. This includes the question of the Boundary of the embedded agent, i.e. the questions which parts an agent considers part of itself and how it delineates other agents from its observations.
OneSince in general Humans can be assigned any values whatsoever), value learning has the pointers problem. It requires that the value can be recovered from the data but cannot prove it.
Practically, the required data has to include sufficient cases of conflict between goals, otherwise, .
People often agree on separate goals until those goals conflict; predicting unconflicted choices does not show how strongly one care should yield to another when both apply.
Another potential challenge is that human values are somewhat mutable and AGI could affect them.
The pointers problem is about figuring out f.Value learning is an approach to learning f from humans.
Even if we can find f for human values, we would still have to build a system that acts on it and is robust to optimization pressure (Goodhart's law). And these two can fail independently of identifying f.
Does anyone in particular own this glossary?
I'm asking because I had a dialogue with ChatGPT that compiled a much more extensive and richly sourced glossary. I used to do this sort of thing by hand - see "The New Hacker's Dictionary" - but an LLM makes an almost perfect neutral observer. Mine now exists as a markdown file containing 222 entries from all across the LW/SSC/ACX corpus and navigation paths for newbies.
What should we do about this?