Are the fast and slow routes analogous to Kannheman’s fast and slow thinking model?
Would it be reasonable to argue that the fast route is what you would expect from high-confidence priors in a predictive coding system, whereas the slow route is what happens when you have a large prediction error?
The Kahneman analogy works at the level of shape: automatic and fast versus serial and effortful. I’d stop short of treating them as the same mechanism.
Here the distinction is narrower, inside reading. Familiar words can be recognized as chunks; the letter-by-letter route preserves access to lower-level detail when that shortcut fails.
The predictive coding version is plausible too, but I’d keep it as a hypothesis. Prediction errors update higher-level representations, but Part 1 doesn’t establish that they trigger the letter-by-letter route.
I had to read this one. It explains the idea under speed reading and why I prefer to read information rather than to listen to someone drone on speaking unless it's in a play, movie, or song. My brain is applying compression. I'm picking up the gist of whole sentences and maybe paragraphs and discarding the fluff, the chit chat and just eating the meat of the argument.
It is also how I recognize good writing. Writing that invites me to slow down and savor.
This article gives a great peak into how the sausage of an LLM is made. Ground up bits of language stripped of its filler. If you strip a language of its letters is it still a valid carrier of concepts? Yes, the spoken word is made up of vibrations in the air and the written word is just a data storage method one level down. The concepts language represent are just compressed ideas and agreed upon norms. When I'm saddled with a large document I wish it could be fed to my brain in rapid fire bits without having to translate it from text to my eyeballs to my brain and all the decoding that has to take place for the meaningful parts to be re-encoded in my memory.
Slowing down to savor is the escape hatch in action: good prose is writing that earns the slow route.
Your reading-over-listening preference has a mechanical basis too. A speaker forces you through their sequence at their pace, while text lets you compress at your own rate, skim the filler, drop to word level when it matters.
The one catch with piping ideas straight into the brain: the decoding you want to skip may be where understanding actually happens. The translation might be the meal.
The Neuralink concept however is one that intrigues me and disturbs me in equal parts. What if AI turns out not to be a separate consciousness per se, but a way to communicate complex ideas in a communal collective mind where thought interchange does not have to be through low bandwidth audio and visual communication and large parts of our intellectual processing could be offloaded to the cloud. A lot of dystopian fiction revolves around the concept. I particularly love the concept in Zardoz where the Eternals join in second level meditation while the crystalline "brain" mediates their thoughts. An under appreciated gem of a Sean Connery movie if you've never seen it. It directly feeds into the idea of who controls AI and does AI end up controlling us? Even the people that created it.
I don't think AI is going to wipe out humans. I think humans are perfectly capable of doing that themselves. AI may be our chance to override that instinct.
But more disturbing might be the fact that destroying ourselves from time to time may actually be the key to our survival.
The collective-mind version is the one that gets me too, because it reframes the control question entirely. If cognition is offloaded to shared infrastructure, "who controls the AI" quietly becomes "who controls the medium your thoughts now run on," which is a much older and scarier question. Bandwidth was never neutral.
That last thought is the real one though. If periodic collapse is load-bearing for survival, then a system smart enough to prevent every collapse might be optimizing for a fragility it can't see. Stability isn't the same as resilience. Haven't seen Zardoz. Adding it.
A knife is a tool and a weapon. To say it is neutral understates its potential danger.
Here is an earlier article I wrote along the same lines. Even though it was only a few months ago it is feeling dated as the future comes rushing at us.
Thought provoking article, always.
Are the fast and slow routes analogous to Kannheman’s fast and slow thinking model?
Would it be reasonable to argue that the fast route is what you would expect from high-confidence priors in a predictive coding system, whereas the slow route is what happens when you have a large prediction error?
The Kahneman analogy works at the level of shape: automatic and fast versus serial and effortful. I’d stop short of treating them as the same mechanism.
Here the distinction is narrower, inside reading. Familiar words can be recognized as chunks; the letter-by-letter route preserves access to lower-level detail when that shortcut fails.
The predictive coding version is plausible too, but I’d keep it as a hypothesis. Prediction errors update higher-level representations, but Part 1 doesn’t establish that they trigger the letter-by-letter route.
I had to read this one. It explains the idea under speed reading and why I prefer to read information rather than to listen to someone drone on speaking unless it's in a play, movie, or song. My brain is applying compression. I'm picking up the gist of whole sentences and maybe paragraphs and discarding the fluff, the chit chat and just eating the meat of the argument.
It is also how I recognize good writing. Writing that invites me to slow down and savor.
This article gives a great peak into how the sausage of an LLM is made. Ground up bits of language stripped of its filler. If you strip a language of its letters is it still a valid carrier of concepts? Yes, the spoken word is made up of vibrations in the air and the written word is just a data storage method one level down. The concepts language represent are just compressed ideas and agreed upon norms. When I'm saddled with a large document I wish it could be fed to my brain in rapid fire bits without having to translate it from text to my eyeballs to my brain and all the decoding that has to take place for the meaningful parts to be re-encoded in my memory.
That is a product idea I am sure is in development right now. A Neuralink product perhaps? https://gerbrot.com/neuralink-article
Slowing down to savor is the escape hatch in action: good prose is writing that earns the slow route.
Your reading-over-listening preference has a mechanical basis too. A speaker forces you through their sequence at their pace, while text lets you compress at your own rate, skim the filler, drop to word level when it matters.
The one catch with piping ideas straight into the brain: the decoding you want to skip may be where understanding actually happens. The translation might be the meal.
Well said.
The Neuralink concept however is one that intrigues me and disturbs me in equal parts. What if AI turns out not to be a separate consciousness per se, but a way to communicate complex ideas in a communal collective mind where thought interchange does not have to be through low bandwidth audio and visual communication and large parts of our intellectual processing could be offloaded to the cloud. A lot of dystopian fiction revolves around the concept. I particularly love the concept in Zardoz where the Eternals join in second level meditation while the crystalline "brain" mediates their thoughts. An under appreciated gem of a Sean Connery movie if you've never seen it. It directly feeds into the idea of who controls AI and does AI end up controlling us? Even the people that created it.
I don't think AI is going to wipe out humans. I think humans are perfectly capable of doing that themselves. AI may be our chance to override that instinct.
But more disturbing might be the fact that destroying ourselves from time to time may actually be the key to our survival.
The collective-mind version is the one that gets me too, because it reframes the control question entirely. If cognition is offloaded to shared infrastructure, "who controls the AI" quietly becomes "who controls the medium your thoughts now run on," which is a much older and scarier question. Bandwidth was never neutral.
That last thought is the real one though. If periodic collapse is load-bearing for survival, then a system smart enough to prevent every collapse might be optimizing for a fragility it can't see. Stability isn't the same as resilience. Haven't seen Zardoz. Adding it.
Zardoz opens with the most haunting choral version of Beethoven's 7th and just gets better from there.
A knife is a tool and a weapon. To say it is neutral understates its potential danger.
Here is an earlier article I wrote along the same lines. Even though it was only a few months ago it is feeling dated as the future comes rushing at us.
https://aldousgerbrot.substack.com/p/who-owns-the-air-we-think-in