Interested in robotics
Bill Jackson
Karma: 57
I do think this is a big problem on twitter, where I used to get valuable signal and now half of even my following feed is Long Tweets.
Not just long, but formatted in punchy paragraphs like this.
And clearly written by AI.
This is a shame, because twitter used to be one of the best places on the internet. Other places, e.g. LessWrong, Substack, seem a lot better though, through some combination of: human review, known writers with reputations to maintain, algorithms that are less virality-happy, good taste and dogfooding by admins. I don’t see why this couldn’t continue for some time.
Online >> real life for spreading ideas
Two critiques of Rethink Priorities’ Moral Weights project
Also in the SEP article is this “Boltzmann died by his own hand” tier intro
Pulling hedonic utilitarianism out of ethical emotivism
Half an argument against the (rationalist’s) many worlds interpretation
Levels of superpersuasion
I can’t attest to trying it when wearing a baby, but I do find that Wispr Flow + Claude Code (avoids needing to type special characters) lets me code with ~100% voice. And it doesn’t have much of a learning curve relative to the old-fashioned ways like https://talonvoice.com/ (my old manager used to swear by this).
Bill Jackson’s Shortform
B.F. Skinner (AKA “father of behaviourism”) expressed the belief that behaviourism would become a kind of weapon of mass destruction as our understanding of the science improved. His thinking was that Hitler in Nazi Germany had achieved something like behavioural control of the population in a crude “punitive” manner, and once we understood on a mechanistic level how to replicate this, we would unlock utopia-producing or civilisation-ending capabilities, and we had to radically change our culture to manage it.
Or, close enough. His actual position was more like “behavioural control is good actually, and since it’s inevitable that it will be developed anyway, we should proactively develop it for the betterment of humanity”, but obviously because it’s such an on-the-face-of-it-scary idea, this has to be couched in apologetic or supervillain-sounding terms. Some quotes to this effect:
The human organism can be controlled, [we have] abundant evidence, through punitive methods. We’ll just take Hitler’s Germany, for example. It looked for a time as if that might indeed survive and be the dominant pattern for a long time to come, a thousand years Hitler thought it might be. But something was wrong with it. Now what was wrong with it? Is it not true that a culture which uses punitive methods is unable to marshal allies in its support?
— Firing Line interview, 1971 (around 11:35)
If [our culture] continues to take freedom or dignity, rather than its own survival, as its principal value, then it is possible that some other culture will make a greater contribution to the future. The defender of freedom and dignity may then, like Milton’s Satan, continue to tell himself that he has ‘a mind not to be changed by place or time’… but he will nevertheless find himself in hell with no other consolation than the illusion that ‘here at least we shall be free’.
— Beyond Freedom & Dignity, Ch 8 (p. 178), 1971
He believed that, because the science was advancing so rapidly (apparently forgetting the fact that he was the primary cause of the science advancing rapidly), behavioural control would become a necessary issue to deal with soon, so we should put our heads down and advance the science even more rapidly, and apply it to culture writ large, to stop the nefarious behaviour-controllers getting there first.
This didn’t pan out in reality. A sufficient reason is that the science actually didn’t develop much further; it never became possible to control human behaviour much more strongly than was achievable in 1971. The reason (I claim) for this seems to be a case of “Reality is often underpowered”: once you go up from the behaviour of individuals in narrow settings to the behaviour of culture, the state space explodes and the ability to sample it collapses, meaning you can never gather enough data to derive general principles. (I realise this is a non-trivial claim that I’m basically just asserting).
...
An earlier Skinner-episode from history is the “verbal behaviour” debate: In 1957 Skinner published the book Verbal Behaviour, arguing that language could be fully explained by reinforcement (stimulus, response, reward). Chomsky wrote a review in 1959 arguing (roughly) that reinforcement can’t explain the generation of completely new sentences. This entered folklore as the beginning of the end for behaviourism and something like the “most devastating book review of all time”. It also seems broadly wrong in retrospect, LLMs produce novel language from reinforcement on text, but it was right that it wasn’t possible to get anywhere interesting using the behaviourist experimental framework on humans at the time.
I claim this was another example of the same thing: Idea correct in principle, reality too underpowered to make it useful at the time.
There are obvious AI parallels here, but I’m not really trying to make a specific point, I just thought Skinner was interestingly ahead-of-his-time on these things, and also wrote like a million books which means there is an unusually high fidelity record of what he thought. Overall his career was like “any day now, we’re going to control complex behaviour through black-box methods, just you wait, any second now” for 40 years.
Interesting post! I think you’re right that entanglement is “a thing”, and that there isn’t a clean pre-existing concept that encapsulates it. For instance the idea of information rate/Shannon entropy is missing the concept of what the information is about; in a video feed of a chess game the pixels representing the pieces are far more important than the pixels representing the table.
To add to your set of candidates for formalising the idea of minimum entanglement: I started going down a similar track in my sequence about Interface-bound agents, and ended up finding ideas from Control theory quite useful. For instance, there is the concept of (Hamilton-Jacobi) reachability analysis, which deals with whether a target state of a system (e.g. a self-driving car reaching a destination) is reachable from an initial state. You could potentially treat this target state as “the task being completed” (or “the unsafe state being avoided”), and fiddle with <something representing allowed actions> and <something representing input information> and find the boundary at which the target state is unreachable.
E.g. in the self-driving car case: You could suppose an agent that can provide steering input at 1 Hz vs 10 Hz (representing lower and higher entanglement in your verbiage). You might show that at 1 Hz it’s impossible to avoid unsafe states because the external environment can dynamically evolve faster than the agent can react (another car can steer into them), but at 10 Hz it is possible. This would establish a minimum entanglement somewhere in that range.