Scott Alexander (pick number 4, but still my number one pick) writes,
My model has several different competing mental processes trying to determine your actions. One is a prior on motionlessness; if you have no reason at all to do anything, stay where you are. A second is a pure reinforcement learner – “do whatever has brought you the most reward in the past”. And the third is your high-level conscious calculations about what the right thing to do is.
These all submit “evidence” to your basal ganglia, the brain structure that chooses actions. Using the same evidence-processing structures that you would use to resolve ambiguous sense-data into a perception, or resolve conflicting evidence into a belief, it resolves its conflicting evidence about the highest-value thing to do, comes up with some hypothesized highest-value next task, and does it.
It’s part of an overall framework that he calls Bayesian, in which different neurological systems feed information to a central decision-making system, with the weights on the different inputs determining the decision.
You can call it Bayesian, but it doesn’t update on any human definition of evidence. If you call activation ‘evidence for’ and inhibition ‘evidence against’ and know what the threshold percentage is, then sure. It’s more like a chemical reaction. What with being made of chemicals and all.
It’s well known that neurons work on an activation/inhibition scheme. Even pain works this way. One set of neurons define an activation energy, and a competing set of neurons tries to achieve that energy. If the threshold is reached, a sensation or decision occurs.
Everyone is in pain all the time. Everyone’s pain is damped by endorphins all the time. Heroin works by damping the pain all the way. Zeroes out the pain activation neurons. Humans like heroin because it makes you pain-free for the first time in your life. Turns out crudely brainjacking yourself is bad, though.
One advantage of this scheme: you can hook two sets of stimuli into one activation energy system without having to rewire anything. You have to adjust baseline inhibition to compensate for the new baseline activation, but neurons do that automatically. Activation causes an increase of the activation energy. Inhibition decays automatically over time, with some caveats I’m not overly familiar with.
E.g. if you haven’t had strawberries in a while, the inhibition on [buy a strawberry] will go down until it passes the baseline activation activity and you buy a strawberry. This increases the activation energy and you don’t keep buying strawberries until you die of malnutrition.
But, e.g. heroin again, by zeroing out the activation, the baseline inhibition of pain is lowered. This is why heroin withdrawal is physically painful. You suffer all the pain you took a vacation from, with interest.
One disadvantage of this scheme is that the maximums tend to be ill-defined. Any sensible programmer would put a cap on the activation of pain. Instead each subsystem feeding into pain-activation has its own cap and it’s far from hard to max out two (or more) subsystems at the same time, thoroughly overwhelming the inhibition energy. Easy to cause far more pain than the downstream systems are designed to handle. Overloads the whole thing.
Heroin directly increases the rate of surgery survival by preventing such overloads. It’s not a luxury.
When they say alcohol lowers your inhibitions, they mean it very literally. Inhibiting neuron clusters are weakened and actions (‘impulses’) that would normally never make it over the activation energy end up storming the barricades.
Which brings us to one of the many weirdnesses of the brain.
For the purposes of alcohol, pain activation counts as pain inhibition, and is lowered. It’s fundamentally a chemical process, not a Bayesian process, and alcohol blocks pain-activation receptors instead of blocking pain-activation emitters. Though of course it also blocks pain-emitters, so it doesn’t fudge your inhibition thresholds much.
We can also talk about the fact inhibition is, from its own perspective, activation. Runner’s high is probably about a spurious overstimulation of the pain-inhibition systems, causing your pain to go below the usual 0 line.
Though it is true that the activation/inhibition of the purely rational circuits is naturally Bayesian due to the basic properties of neurons. With caveats regarding conformism and not losing all your friends, the activation neurons for a belief (dogs are fuzzy) is defined by the weight of experience on that side (touching a dog and it was fuzzy) and the inhibition is defined by the experience against (touching a dog and it was spiky). In other words being consciously Bayesian is to spent precious consciousness computing time emulating a lower-level process that’s already being carried out in almost all cases. And it will still fail to overcome the social-caveat overrides because it’s downstream of the overrides, not upstream.
” but it doesn’t update on any human definition of evidence”
Sense data isn’t evidence?
Last I checked alcohol and heroin weren’t evidence, no.
Scott’s style of being an intellectual is clearly a huge influence, even an inspiration, for your FITs Steelmanning rule and many FIT 2.0 suggested changes. He’s certainly good, and far better than most more famous folk with more famous books & speaking gigs.
I hope you can highlight a few more non-Scott Steelman arguments from the other intellectuals; I suggest a weekly Steelman link roundup of all FIT S points, optionally with a line or two from you about it, here on ASKblog.
If there is a category like S, with 1 specific intellectual of 150 getting more than 20% (10%? 30%) of the points, that category is not general enough. Similarly if the “intellectual monopoly” of the top 5 or fewer have more than 50% of the points.
If the rules of the game are “meta”, the criteria for choosing good rules are the second level of “meta”. (Didn’t TheZvi recently say there are only at most 3 levels of meta?)
More specific examples of what are points and what are not, according to the rules, helps tremendously in clarifying the rules. The latest FITs v2.0 look a lot better already.
Reminds me of Windows 1.0 (totally terrible, ’84ish with Macs), and Win 2.0 (terrible), and Win 3.0 (lousy) moving to Win 3.11 (about ’93) usable, even in Slovakia, and Windows 95, and the start of GUI.
… the 95ish domination of GUI on Windows machines – Macs were GUI starting in ’84. Good enough hardware wasn’t cheap enough earlier.
And while open Wintel boxes are more flexible, and easier to work on, they’re also easier to hack deliberately or accidentally – far more fragile. Anti-fragility has clear processor costs for security and interoperability and back-compatibility desirable characteristics.
The Windows analogy is appropriate, but I hope we iterate faster to the equivalent of Win 95