AI Reward

By

Reward in Reinforcement Learning

Reward Modelling

AI Reward Models

Reward Hacking

Reward Provisioning

Reward Misspecification

AI Alignment

Making the Reward Model Explicit

Goodhart's Law

Comments

Post a Comment

Popular posts from this blog

Computing and the Linguistic Turn

By
Image
Enigma (Wikipedia) Symbolic analysis and statistical analysis were both approaches codebreakers took in breaking Enigma. The development of AI has also incorporated both approaches. One historical question for AI is why the dominance of Symbolic AI, associated here with the linguistic turn, from 1950 to 1990? Note the development of AI out of the Dartmouth Workshop and especially the contributions of one of its attendees,  Ray Solomonoff , to the statistical approach, an approach taken up by Marcus Hutter , a DeepMind  researcher, and of course, others. Note that a history of AI outside the linguistic turn from Moritz Hardt and Ben Recht is in Patterns, Predictions, and Actions . And note too that along with the difference between the symbolic and the statistical is the difference between language learning and computer vision, which tends to be a history outside the linguistic turn. And the turn Richard Sutton takes is to push learning outside the context of language, which ...
Read more

A Heidegger - Bayes Hybrid Model

By
Image
A Conjecture for a Hybrid Model Bayesian epistemology concerns modelling and specifically, belief, but a Bayesian belief does not need to be fully formed to emerge alongside the Heideggerian to-hand. The Heideggerian to-hand thereby becomes the interaction that generates this further, which is, with this interaction, a certain realization of being in the world. Bayesian selection now concerns available beliefs as well as any to-hand [belief(s)]. How Heidegger Can Make You a Better Guitarist Bayesian Philosophy Bayesian Philosophy Bayesian Epistemology Bayesian Epistemology Bayesian Modelling Bayesian Modelling Bayesian Model Selection David Deutsch on Bayesian Epistemology Bayesian Philosophy of Science Bayesian Philosophy of Science A Bayesian Approach to the Philosophy of Science On Sean Carroll on an Application of Bayes The Heideggerian To-Hand On Heidegger On Heidegger Flipping Descartes Ready-to-Hand and Present-at-Hand Presence-at-Hand Ready-to-Hand and Unready-to-Hand Does Hei...
Read more

AI as the Ghost of Christmas Future

By
AI Safety in the Metanarrative A Mythography The Industrial Age (19th Century) largely ushered in an optimism regarding technological progress, but with that optimism also came the critics.  Take the work of Nathaniel Hawthorne, for example. Better yet, take the work of Mary Shelley. Note that her Frankenstein is nearby on the timeline to the work of Charles Babbage and to the work of Ada Byron Lovelace .  And though Frankenstein is, in part,  about reanimation, it is also about the creation of intelligent being. So, we see that, historically, a narrative surrounding computing comes into being along another narrative, Frankenstein's monster, and we see these two narratives coupled now in much of the narrative around AI Safety. But my point here is not that we need to decouple the stories regarding the monstrous  ("War Games", "Sneakers", "Goldeneye", ...) so much as we need to re-assess the goals that are embedded in the narrative.  With what we know n...
Read more