I don't think the out loud or someone listening / reacting matters at all here. Suspect it's entirely this:
>The thought that was comfortable as a vague impression has to become a sentence, and sentences have structure.
It's not unlike what people like PG say about writing improving thinking...it's the being forced to go from fuzzy directional notions to something you can put on paper in that will stand up to critique.
Same with rubber duck debugging. The verbal part means you need to articulate it clearly but it's not the speaking that helps. Same with writing a detailed spec/prompt for an LLM - I know if its too fuzzy ("set an appropriate timeout") the LLM will spin it's wheels so it forces clarity.
Also suspect that a big part of who we consider intelligent is linked to this. Maybe their internal monologue is just more crisp - closer to what they'd tell a rubber duck.
I think there are tradeoffs though, and this has been a thorn in my side during technical interviews where you are expected to think out loud because:
1. Sometimes you have a vague sense of the shape of the solution, and ime it can be helpful to sit with it for a while before trying to shape it into words.
2. Talking out loud forces structure but it also rate-limits how quickly you can iterate through ideas to find one that plausibly solves the problem at hand
In 2017 LLMs weren't powerful enough to generate working code on their own, but my goal was to at least create a chatbot that could help you rubber-duck-debug your way to a solution. Unfortunately the tech wasn't quite strong enough for that, and not enough engineers even knew what rubber-duck-debugging was. RIP Duckly.
Trying to train an LLM on two 1080ti's on the StackOverflow corpus in my living room was a vibe though. Good times.
Duckly deserved to actually work. There’s a small irony here: the closest study I found to this, robots specifically built to simulate attentive listening, found they performed no better than an actual inanimate rubber duck for adult engineers. The mechanical signal of listening doesn’t seem to be the active ingredient. Makes me wonder if Duckly would have needed real disagreement to close a gap a duck can’t, not just better natural language.
You're probably on to something with the value of disagreement. I think it's one reason why chatting with current models doesn't create the same stimulation as rubber-ducking used to bring. The models are typically too quick to agree and amplify what you think rather than truly break it down and push back.
And thanks for saying it should have worked, I agree. My chagrin has increased over the years as I have realized the magnitude of my ill-timing.
I started my web dev career in 1999 so my main code references were a combination of O’reilly and “for dummies” books. As a wet behind the ears engineer I’d find myself regularly walking over to my more senior friend Dan’s cubicle for help.
Half the time on the walk over, trying to frame the question in my mind I’d figure out the answer or at least next step. It got to the point where Dan would see me heading towards him and suddenly turn around and he’d as “Figure it out?” And I’d throw him a thumbs up on the way back to my desk.
OMG - strong vibes to Einstein crediting Michele Besso, his colleague at the Swiss Patent Office, with helping him discussing some concepts in the special relativity paper: see at the end of the paper https://www.fourmilab.ch/etexts/einstein/specrel/specrel.pdf
Thank you for sharing this. Growing up Asian American many teachers disciplined me and made assumptions about my intelligence for not being as vocal as the other kids. Culture shapes cognition and vice versa.
> The Enigma of Reason (Hugo Mercier and Dan Sperber, 2017): their argumentative theory holds that reasoning evolved for social rather than individual epistemic purposes, to produce and evaluate arguments in group contexts.
Yes! I love that someone wrote this down!
This seems so obvious to me now. I often ask LLMs to cite their sources (they do hallucinate from time to time), and they often give me sources that don't say what is claimed. "How would the LLM know not to give this to me?" I wonder. They're trained to explain but not to convince, so they don't know what's convincing, and they should.
I think humans hallucinate at least as much as LLMs—arguments of any complexity are impossible to formulate without leaping at least a bit—but other humans ground us. That's why when people become socially isolated, they join cults or adopt conspiracy theories or the like.
Conversely, "this is convincing to an expert" converges on “this is true" as our collective expertise grows over time. This is the foundation of the scientific method, of progress in all engineering disciplines, etc.
The move from thinking to semantic conversion is important for investigation/introspection.
Arguing with yourself also seems to engage your brains "theory of mind" centers, so different pathways get activated to examine the problem space.
The problem with Ai is the fact that it hallucinates and if you're doing anything truly novel in an integration or framing sense it bottoms out very quickly and can't engage. A human operator can decompose the problem and get accuracy checks for known areas in the training data of course.
Now to be I'm not saying Ai can't produce novel work on the edge but in my experience it is antagonistic towards those goals.
Case in point, CRDTs, many don't use tombstones but they are the minority, and if you try iterate a new CRDT off of one that doesn't use tombstones, let's say diamond-types, it will keep pulling you back to tombstones.
The problem is that the number of humans who understand dynamic investigation and the push pull of exploring an idea you don't hold with someone has always been very small, and now with reflexive internet argument culture driving how we view "debate" and "discussion".
I don't know if we've reduced the leisure to think or what but things are not great for finding speculative thinking partners.
I don't think the out loud or someone listening / reacting matters at all here. Suspect it's entirely this:
>The thought that was comfortable as a vague impression has to become a sentence, and sentences have structure.
It's not unlike what people like PG say about writing improving thinking...it's the being forced to go from fuzzy directional notions to something you can put on paper in that will stand up to critique.
Same with rubber duck debugging. The verbal part means you need to articulate it clearly but it's not the speaking that helps. Same with writing a detailed spec/prompt for an LLM - I know if its too fuzzy ("set an appropriate timeout") the LLM will spin it's wheels so it forces clarity.
Also suspect that a big part of who we consider intelligent is linked to this. Maybe their internal monologue is just more crisp - closer to what they'd tell a rubber duck.
I think there are tradeoffs though, and this has been a thorn in my side during technical interviews where you are expected to think out loud because:
1. Sometimes you have a vague sense of the shape of the solution, and ime it can be helpful to sit with it for a while before trying to shape it into words.
2. Talking out loud forces structure but it also rate-limits how quickly you can iterate through ideas to find one that plausibly solves the problem at hand
In 2017 LLMs weren't powerful enough to generate working code on their own, but my goal was to at least create a chatbot that could help you rubber-duck-debug your way to a solution. Unfortunately the tech wasn't quite strong enough for that, and not enough engineers even knew what rubber-duck-debugging was. RIP Duckly.
Trying to train an LLM on two 1080ti's on the StackOverflow corpus in my living room was a vibe though. Good times.
Duckly deserved to actually work. There’s a small irony here: the closest study I found to this, robots specifically built to simulate attentive listening, found they performed no better than an actual inanimate rubber duck for adult engineers. The mechanical signal of listening doesn’t seem to be the active ingredient. Makes me wonder if Duckly would have needed real disagreement to close a gap a duck can’t, not just better natural language.
You're probably on to something with the value of disagreement. I think it's one reason why chatting with current models doesn't create the same stimulation as rubber-ducking used to bring. The models are typically too quick to agree and amplify what you think rather than truly break it down and push back.
And thanks for saying it should have worked, I agree. My chagrin has increased over the years as I have realized the magnitude of my ill-timing.
perhaps it is time to resurrect Duckly queue Frankenstein music and thunder in background
I started my web dev career in 1999 so my main code references were a combination of O’reilly and “for dummies” books. As a wet behind the ears engineer I’d find myself regularly walking over to my more senior friend Dan’s cubicle for help.
Half the time on the walk over, trying to frame the question in my mind I’d figure out the answer or at least next step. It got to the point where Dan would see me heading towards him and suddenly turn around and he’d as “Figure it out?” And I’d throw him a thumbs up on the way back to my desk.
OMG - strong vibes to Einstein crediting Michele Besso, his colleague at the Swiss Patent Office, with helping him discussing some concepts in the special relativity paper: see at the end of the paper https://www.fourmilab.ch/etexts/einstein/specrel/specrel.pdf
Communicating ideas helps, but thinking out loud may not work better to some people.
Thinking silently fits Asian Americans better than Euro Americans*.
https://www.psychologytoday.com/us/blog/sex-murder-and-the-m...
Thank you for sharing this. Growing up Asian American many teachers disciplined me and made assumptions about my intelligence for not being as vocal as the other kids. Culture shapes cognition and vice versa.
Writing is as good, or even better https://www.paulgraham.com/words.html
You can tell you didn't read the article before jumping into commenting, as this point is addressed.
Only as
> the act of writing out a problem to a model still forces the same sentence-level precision described earlier
(model referring to LLM here)
but not as writing for writing's sake
> The Enigma of Reason (Hugo Mercier and Dan Sperber, 2017): their argumentative theory holds that reasoning evolved for social rather than individual epistemic purposes, to produce and evaluate arguments in group contexts.
Yes! I love that someone wrote this down!
This seems so obvious to me now. I often ask LLMs to cite their sources (they do hallucinate from time to time), and they often give me sources that don't say what is claimed. "How would the LLM know not to give this to me?" I wonder. They're trained to explain but not to convince, so they don't know what's convincing, and they should.
I think humans hallucinate at least as much as LLMs—arguments of any complexity are impossible to formulate without leaping at least a bit—but other humans ground us. That's why when people become socially isolated, they join cults or adopt conspiracy theories or the like.
Conversely, "this is convincing to an expert" converges on “this is true" as our collective expertise grows over time. This is the foundation of the scientific method, of progress in all engineering disciplines, etc.
LLMs are really good rubber ducks.
See also https://en.wikipedia.org/wiki/Rubber_duck_debugging
I keep a rubber duck on my desk and have for at least 10 years, because of this. Great conversation starter.
A collection of thoughts on this.
Pierces Firstness is exactly what drives this.
The move from thinking to semantic conversion is important for investigation/introspection.
Arguing with yourself also seems to engage your brains "theory of mind" centers, so different pathways get activated to examine the problem space.
The problem with Ai is the fact that it hallucinates and if you're doing anything truly novel in an integration or framing sense it bottoms out very quickly and can't engage. A human operator can decompose the problem and get accuracy checks for known areas in the training data of course.
Now to be I'm not saying Ai can't produce novel work on the edge but in my experience it is antagonistic towards those goals.
Case in point, CRDTs, many don't use tombstones but they are the minority, and if you try iterate a new CRDT off of one that doesn't use tombstones, let's say diamond-types, it will keep pulling you back to tombstones.
The problem is that the number of humans who understand dynamic investigation and the push pull of exploring an idea you don't hold with someone has always been very small, and now with reflexive internet argument culture driving how we view "debate" and "discussion".
I don't know if we've reduced the leisure to think or what but things are not great for finding speculative thinking partners.