It is widely known that cognitive behavioural therapy (CBT) is a standard treatment for depression. A little-known fact, however, is that its subcomponent, behavioural therapy (BT), is equally effective [1], according to a meta-analysis by Cochrane
If the term “reinforcement learning” makes you think of Sutton & Barto, or the newest shiny model release, then, ugh, fine, I’ll explain it in their terms: Sequences of actions are graded according to a reward function. For example, DeepSeek spits out sequences of tokens, called chains-of-thought (CoT), and sequences that yield a correct answer are reinforced over those that do not.
Anything that is positively associated with getting the answer right, with getting the reward, such as correctly multiplying 322 by 214, is called a secondary reinforcer.
Behavioural therapy (BT) works by making effort a secondary reinforcer of your actions. That’s it.
To get started, you need morale [3], that is, you need to believe that effortful actions are associated with reward. If your level of morale is very low, you have come to believe there are no such actions, and taking very ambitious actions that then fail will only reinforce that belief
If you are unsure what gives humans reward, please consult Maslow’s Hierarchy of Needs. Surely things like “having sex” or “gaining status” may not sound quite as idealistic as “making the singularity go well” to you. But again, we can reach for the power of association: Say there is a subculture with the aim of “making the singularity go well”, then “gaining status” may bring you closer to your terminal goal than initially assumed.
In fact, this principle extends way beyond any rigid DSM-5 definition [5], to all human behaviour. I posit the existence of two attractor states, existing alongside the middle ground of normalcy:
Any behaviour has a set of cues associated with it, e.g. a cue for DeepSeek is ‘322 × 214’. The environment is the set of all cues. Therefore, to change behaviour, a maximally disjoint environment needs to be created, whose cues must secondarily reinforce effort.
During the Manhattan Project, the Los Alamos Laboratory reliably created a strong effort spiral in its subjects, with ~75-hour work-weeks being common. Finding a modern equivalent is left as an exercise for the reader. Hint: Try to farm LessWrong karma, or, idk, do this “cold email” thing.
Unfortunately, other factors also play a significant role. For instance, we know executive function [8] mediates the degree of effort one puts forth. The trait involves the ability to self-regulate and is highly heritable. The most effective intervention, stimulants, suffers from unclear long-term effectiveness
Another slap in the face is that interventional negative reinforcement (whose exertion by other people is called dominance) may be useful [10]. Fortunately, its sting may fade out in aggregation, that is, thanks to the hedonic treadmill.
All this leaves us with, well, my bad attempts at poetry:
The road from akrasia? — Well, it’s plain
and simple to express:
Slack
and slack
and slack again
but less
and less
and less[11]. The pharmacology people sometimes talk about “active ingredients” here, idk.
If the name “Cochrane” is unknown to you, just nod along approvingly, because they are very high-status wizards of meta-analytic black magic. cf. Behavioural therapies versus other psychological therapies for depression, Cochrane Database of Systematic Reviews A well-known psychiatrist says this is a common failure mode he experiences, so you better believe it, m’kay? cf. Peer Review Request: Depression, S. Alexander, AstralCodexTen, Point 2.1.5
At which point it is also prudent to clarify that the subject of this article is ‘behavioural depression’, and it does not aim to engage with the many subtleties of ‘depression’ as defined by the DSM-5.
Which is the subject of, and a shameless ad for, my last article