1. Summary and Table of Contents

1.1 Summary

In the context of “brain-like AGI”, a yet-to-be-invented variation on actor-critic model-based reinforcement learning (RL), there's a ground-truth reward function (for humans: pain is bad, eating-when-hungry is good, various social drives, etc.), and there's a learning algorithm that sculpts the AGI's motivations into a more and more accurate approximation to the future reward of a possible plan.

Unfortunately, this sculpting process tends to systematically lead to an AGI whose motivations fit the reward function too well, such that it exploits errors and edge-cases in the reward function. (“Human feedback is part of the reward function? Cool, I’ll force the humans to give positive feedback by kidnapping their families.”) This alignment failure mode is called “specification gaming” or “reward hacking”, and includes wireheading as a special case.

If too much desire-sculpting is bad because it leads to overfitting, then an obvious potential solution [...]

---

Outline:

(00:11) 1. Summary and Table of Contents

(00:16) 1.1 Summary

(02:31) 1.2 Table of contents

(05:19) 2. Background intuitions

(05:23) 2.1 Specification gaming is bad

(06:32) 2.2 Specification gaming is hard to avoid. (The river wants to flow into the sea.)

(07:31) 3. Basic idea

(07:35) 3.1 Overview of the desire-sculpting learning algorithm (≈ TD learning updates to the value function)

(10:02) 3.2 Just stop editing the AGI desires as a broad strategy against specification-gaming

(11:20) 4. Example A: Manually insert some desire at time t, then turn off (or limit) the desire-sculpting algorithm

(13:21) 5. Example B: The AGI itself prevents desire-updates

(13:46) 5.1 Deliberate incomplete exploration

(15:35) 5.2 More direct ways that an AGI might prevent desire-updates

(16:16) 5.3 When does the AGI want to prevent desire-updates?

(18:50) 5.4 More examples from the alignment literature

(19:33) 6. Example C: Non-behaviorist rewards

(21:03) 6.1 The Omega-hates-aliens scenario

(26:59) 6.2 Defer-to-predictor mode in visceral reactions, and trapped priors

(29:20) 7. Relation to other frameworks

(29:34) 7.1 Relation to (how I previously thought about) inner and outer misalignment

(30:29) 7.2 Relation to the traditional overfitting problem

(34:19) 8. Perils

(34:43) 8.1 The minor peril of under-sculpting: path-dependence

(36:16) 8.2 The major peril of under-sculpting: concept extrapolation

(38:31) 8.2.1 There's a perceptual illusion that makes concept extrapolation seem less fraught than it really is

(41:34) 8.2.2 The hope of pinning down non-fuzzy concepts for the AGI to desire

(42:35) 8.2.3 How blatant and egregious would be a misalignment failure from concept extrapolation?

(44:42) 9. Conclusion

The original text contained 12 footnotes which were omitted from this narration.

---

First published:
August 5th, 2025

Source:
https://www.lesswrong.com/posts/grgb2ipxQf2wzNDEG/the-perils-of-under-vs-over-sculpting-agi-desires

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Just for fun, here are a couple of my past depictions of the desire-updating learning algorithm. (a) is from my Intro Series §9.5.2, where I was relating my neuroscience models to Agent Foundations discourse. (b) is from my old 2021 post Model-based RL, Desires, Brains, Wireheading where I was first starting to get the idea that one could build an (approximate) reward-maximizing agent by combining a “desire-driven agent” with a demon (learning algorithm) that continually edits those desires to better match a ground-truth reward function, starting from random initialization.

(Copied from Neuroscience of human social instincts: a sketch.) After within-lifetime learning by the short-term predictor (probably in the amygdala in this case), the idea of a spider can trigger physiological arousal, via the pathway marked in red—concepts in the cortex serve as context for the short-term predictor, and the output of that predictor then triggers physiological arousal (a self-fulfilling prophecy).

(Copied from Neuroscience of human social instincts: a sketch.) The short-term predictor output is trained by the “thinking of a conspecific” flag in the Steering Subsystem, but can also trigger that flag in turn (a self-fulfilling prophecy). This enables social instincts to trigger when the conspecific is not physically present, thanks to generalization from past situations in which they were.

In §4-like proposals, people often want to imbue an AGI with a positively-vibed concept like “flourishing”, “helpfulness”, “human values”, etc. For a concept like this, when humans try to apply them out-of-distribution, they are constantly querying their ground-truth determiner of what is good and bad, the very same machinery that powers moral and philosophical progress. So the concept seems to generalize well, even across “sharp left turn”-like distribution shifts. But that’s tautological, and doesn’t generalize to an AGI that would lack such ground truth.

Diagram showing reward function alignment challenges in AI systems with green mascot.

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

“The perils of under- vs over-sculpting AGI desires” by Steven Byrnes

LessWrong (30+ Karma)

“The perils of under- vs over-sculpting AGI desires” by Steven Byrnes