Things I keep coming back to that do not belong anywhere else on the site.
Books
- Deep Learning, by Goodfellow, Bengio, and Courville
- Reinforcement Learning: An Introduction, by Sutton and Barto
Essays
Talks and lectures
Questions I keep coming back to
- What information does an agent actually need to coordinate well?
- Should we try to mimic a human way of thinking, or let a system learn by exploration without building that structure in first?
Two viewpoints
Richard Sutton's argument in The Bitter Lesson is still one of the clearest warnings against overvaluing hand-designed structure. Compute and search scale; most of our clever constructions do not survive contact with systems that keep learning from more experience. I come back to it because it cuts against a common instinct. When a system feels more interpretable or more human-designed, we tend to assume it is more principled. Sutton says this can be exactly backwards.
Andrej Karpathy's Animals vs Ghosts pushes back from another side. His framing is that current frontier systems may not be "animals" in Sutton's sense at all but "ghosts": saturated with human text, human priors, and human engineering, and still able to do real work in the world. Which leaves an honest open question. Do we imitate a human way of thinking, or let a system earn its competence through exploration, even if the internal route looks nothing like us?
I do not think the two cancel out. Sutton warns against trusting handcrafted stories too much. Karpathy warns against pretending today's engineered systems are already the bitter lesson carried to its limit. I find the tension useful, for reinforcement learning and for modelling generally. Often the real question is not what the most human-looking way to do something is, but what structure is actually necessary and what can be left to learning.