At the same time, I've learned SO much about how human brains work by observing our clumsy attempts to replicate human-esque thought. I'm especially intrigued by the little decision-making tugs-of-war that cross lots of parameters all at once. We're not aware of these ultra-fast struggles within us, but that seems to describe how humans think to (with the notable caveats you did a great job of calling out).
While there's no "learning" with trained AI systems, IMO looking stuff up in real time is kind of similar. I've spent a fair bit of time lately thinking about this particular rabbit hole, and maybe it's something we should sidebar for a deeper discussion, but I'll try to TL;DR it for you here.
Essentially, what we considered "smart" has continuously evolved over time. It was all about memorization once upon a time, then it was about interpreting language. If you and I were on a phone call and you asked me 10 random trivia questions, and I answered each within a few seconds, you'd think I was a genius.
Perhaps real-time access to the internet (and better efficiency at looking things up) is kind of a cousin to learning new skills, in a sense.
Yup, it's a very good question. I'm thinking of a very specific definition of learning, Tom Mitchell's definition, which is basically any system that gets better at doing some task when exposed to more experience. In this sense, the NN underlying the LLM is definitely not learning, but the chatbot as a system can be seen as a learning system, because with more experience (more data to do RAG from) it gets better at answering questions. So yes, nuances :)
It's also interesting to try to predict what domains will soon be thought of in this way - no longer as a sign of incredible intelligence or whatever, but just information everyone can get to easily. If you have a smartphone with you and I ask you what the capital of South Dakota is, you can tell me within a few seconds max. That's no longer useful info to carry around in your mind, just like memorizing long texts became less useful as print spread, etc.
This kinda stuff is right where our minds intersect.
Tangential but I’m slightly obsessed with the evolution of what we have outsourced. I still file things to keep digital life organized as it were a shelf. I appreciate many people now do no filing and rely entirely on the search function. What do we gain and what do we lose when we do this? My filing things, I create a map to recall things. Whereas with search reliance it is a term or topic or single piece of text. As we rely more on tools do to tasks, what is the cost of reliance? Consider an amazon warehouse, without a digital system, humans could find nothing in there - needle in a haystack. Whereas, I can go into the British library and with knowledge of the Dewey decimal system, I can find anything.
I think it's the same as memorizing phone numbers or street addresses - it's inevitable that folks will drop this aspect of memory in favor of one that doesn't cost anything, cognitively. I think you are right to point to the trade-off- we always give something up to get something else.
My impression is that we've generally gone in the right direction with this, but I also wonder at what we've left behind. Memory palaces are things of the distant past for all but a small handful of humans alive today, for instance.
Great example! There’s something to it. I come at this not as stop the trains everyone, and more from a let’s keep moving but can someone think about the stuff we might not want to jettison.
I definitely think there is value in having information organized in one's brain, because we cannot make connections about things we don't think about, as trivial as that sounds. So I definitely hoard on knowledge about the things I care about, I just don't hoard knowledge about the things I don't care, but still need to know to function in society.
Good way to put it. I think that's happening with language and LLMs, just producing the language folks can use for literally anything they need to write. There's an art to language that may be lost, or at least the cognitive benefits of practicing that skill may be lost if the skill itself eventually becomes less important.
Such a valuable contribution, Alejandro. The message is expressed with rare clarity and credibility. I needed to know the fundamental significance of an LLM as a static machine incapable of continuous learning. I appreciated your comment clarifying the difference between what Perplexity and its ilk does and learning. Perhaps anthropomorphosizing is the greatest problem.
Thanks Terry. Yes, anthropomorphization is a big issue, and not only in lay circles. There is a huge tendency in inner technical and even academic circles in AI to kind of even think of these things like supernatural beings.
On this I remain hopeful because crafting an efficient prompt requires understanding how to write which in turn requires organizing thoughts. From there it is a matter of style. So watching how language evolves will be fascinating - those with good language skills will do wonders, those with so-so will get lost in loops with their AI powered tools.
I agree, although I find myself using briefer and briefer language to get the same sorts of things done. Perhaps the new human language will be much more abbreviated, since we're sort of becoming cyborgs. Who knows, all this will probably look pretty silly in just a few years!
I reflected on you book vs ChatGPT comment and it connects to this article in an interesting way. What books give you is the tangential information that sparks creativity. Getting AI tailored enough is mainlining info, no filler. This article explained exactly why humans benefit so much from those random connections of new info.
If I look at my travel writing, I fill it with tangential info that isn’t really about the destination. A reader will be sparked more by that kind of info than a simple “go here for X”.
I will write this up as I’m working on an article about memory. Thanks for prompting me to think differently ;)
Agreed, generally going in a good direction when we replace the required memory with useful stuff. Knowing someone’s phone number is less useful than remembering their name (hence contact apps are incredibly useful). However, creating maps are useful for retaining knowledge (it’s even a technique for learning new conceptual info). This was my thought with the focus on search - we are making fewer and fewer of our own maps as we go. My gut tells me we are losing something valuable as people completely outsource navigation. There was a thread on the potential closing of TikTok and a teacher had students saying they would not know where to begin looking for music. It’s a trite example yet connects to a thought I’m not yet able to articulate fully.
I get this, and I'm also only able to articulate this partially. I think the same thing may be true of language: reading lengthy books used to be the only way to get the same sort of knowledge I get in 3 minutes with GPT4, for instance. But those tedious deep dives are exactly what made me who I am today.
I also think we will find that using the internet as a data source is dubious at best. How can we tell a complex query/prompt leads to a solution that is valid. There's more detail in the code if anyone ever releases some. I suspect shenanigans in the algorithms to "help" it along. Again, hype out of control.
Oh I think you're gonna love this project we (my wife and I) worked on during our PhDs (https://autogoal.github.io), it's grammatical evolution for evolving machine learning pipelines. Grammars and compilers are what I teach :)
Early NN were universal function approximators, which have seriously good uses beyond theory. I suspect Newer LLMs are not much different conceptually. It's all very interesting to see, but the hype is out of control. They (those who write about it) say very fancy things like rationality and gigantic number of components. Throwing hardware at it will not go on forever. I think we are close to the limit, but I would put any bet on my predictions-- or anybody else's so far :-)
I agree, and I think we've already seen signals of diminishing returns in throwing compute at existing architectures. Halving the error on the most challenging reasoning benchmarks (which are still just benchmarks, thus gameable) is costing an order of magnitude more compute.
Nice summary. I (Bill White) worked in the AI/machine learning field for 20 years. We used NN to predict disease status from DNA/RNA (WIDE data). We also used evolutionary computing to EVOLVE NN nodes and structure. The main study was done by a PhD student in our lab. I was senior programmer. Here's the NN paper. and a list of my publications if you are curious.
I still think the 'retraining' is an element that will need to evolve in ANN's to make them more useful. The biological brain is always training, always running inference, at the same time. So the next phase in ANN's is to allow for this sort of mode, which also jacks up the energy requirement using current methods....which then leads me to the most remarkable thing about the human brain. It runs on the fuel of one banana while GPT and even Deep Seek still require massive amounts of grid power.
This is well said, and people need to hear it.
At the same time, I've learned SO much about how human brains work by observing our clumsy attempts to replicate human-esque thought. I'm especially intrigued by the little decision-making tugs-of-war that cross lots of parameters all at once. We're not aware of these ultra-fast struggles within us, but that seems to describe how humans think to (with the notable caveats you did a great job of calling out).
While there's no "learning" with trained AI systems, IMO looking stuff up in real time is kind of similar. I've spent a fair bit of time lately thinking about this particular rabbit hole, and maybe it's something we should sidebar for a deeper discussion, but I'll try to TL;DR it for you here.
Essentially, what we considered "smart" has continuously evolved over time. It was all about memorization once upon a time, then it was about interpreting language. If you and I were on a phone call and you asked me 10 random trivia questions, and I answered each within a few seconds, you'd think I was a genius.
Perhaps real-time access to the internet (and better efficiency at looking things up) is kind of a cousin to learning new skills, in a sense.
Very fun to compare/contrast!
Yup, it's a very good question. I'm thinking of a very specific definition of learning, Tom Mitchell's definition, which is basically any system that gets better at doing some task when exposed to more experience. In this sense, the NN underlying the LLM is definitely not learning, but the chatbot as a system can be seen as a learning system, because with more experience (more data to do RAG from) it gets better at answering questions. So yes, nuances :)
It's also interesting to try to predict what domains will soon be thought of in this way - no longer as a sign of incredible intelligence or whatever, but just information everyone can get to easily. If you have a smartphone with you and I ask you what the capital of South Dakota is, you can tell me within a few seconds max. That's no longer useful info to carry around in your mind, just like memorizing long texts became less useful as print spread, etc.
This kinda stuff is right where our minds intersect.
Tangential but I’m slightly obsessed with the evolution of what we have outsourced. I still file things to keep digital life organized as it were a shelf. I appreciate many people now do no filing and rely entirely on the search function. What do we gain and what do we lose when we do this? My filing things, I create a map to recall things. Whereas with search reliance it is a term or topic or single piece of text. As we rely more on tools do to tasks, what is the cost of reliance? Consider an amazon warehouse, without a digital system, humans could find nothing in there - needle in a haystack. Whereas, I can go into the British library and with knowledge of the Dewey decimal system, I can find anything.
I think it's the same as memorizing phone numbers or street addresses - it's inevitable that folks will drop this aspect of memory in favor of one that doesn't cost anything, cognitively. I think you are right to point to the trade-off- we always give something up to get something else.
My impression is that we've generally gone in the right direction with this, but I also wonder at what we've left behind. Memory palaces are things of the distant past for all but a small handful of humans alive today, for instance.
Great example! There’s something to it. I come at this not as stop the trains everyone, and more from a let’s keep moving but can someone think about the stuff we might not want to jettison.
I definitely think there is value in having information organized in one's brain, because we cannot make connections about things we don't think about, as trivial as that sounds. So I definitely hoard on knowledge about the things I care about, I just don't hoard knowledge about the things I don't care, but still need to know to function in society.
Good way to put it. I think that's happening with language and LLMs, just producing the language folks can use for literally anything they need to write. There's an art to language that may be lost, or at least the cognitive benefits of practicing that skill may be lost if the skill itself eventually becomes less important.
Such a valuable contribution, Alejandro. The message is expressed with rare clarity and credibility. I needed to know the fundamental significance of an LLM as a static machine incapable of continuous learning. I appreciated your comment clarifying the difference between what Perplexity and its ilk does and learning. Perhaps anthropomorphosizing is the greatest problem.
Thanks Terry. Yes, anthropomorphization is a big issue, and not only in lay circles. There is a huge tendency in inner technical and even academic circles in AI to kind of even think of these things like supernatural beings.
On this I remain hopeful because crafting an efficient prompt requires understanding how to write which in turn requires organizing thoughts. From there it is a matter of style. So watching how language evolves will be fascinating - those with good language skills will do wonders, those with so-so will get lost in loops with their AI powered tools.
I agree, although I find myself using briefer and briefer language to get the same sorts of things done. Perhaps the new human language will be much more abbreviated, since we're sort of becoming cyborgs. Who knows, all this will probably look pretty silly in just a few years!
I reflected on you book vs ChatGPT comment and it connects to this article in an interesting way. What books give you is the tangential information that sparks creativity. Getting AI tailored enough is mainlining info, no filler. This article explained exactly why humans benefit so much from those random connections of new info.
If I look at my travel writing, I fill it with tangential info that isn’t really about the destination. A reader will be sparked more by that kind of info than a simple “go here for X”.
I will write this up as I’m working on an article about memory. Thanks for prompting me to think differently ;)
Agreed, generally going in a good direction when we replace the required memory with useful stuff. Knowing someone’s phone number is less useful than remembering their name (hence contact apps are incredibly useful). However, creating maps are useful for retaining knowledge (it’s even a technique for learning new conceptual info). This was my thought with the focus on search - we are making fewer and fewer of our own maps as we go. My gut tells me we are losing something valuable as people completely outsource navigation. There was a thread on the potential closing of TikTok and a teacher had students saying they would not know where to begin looking for music. It’s a trite example yet connects to a thought I’m not yet able to articulate fully.
I get this, and I'm also only able to articulate this partially. I think the same thing may be true of language: reading lengthy books used to be the only way to get the same sort of knowledge I get in 3 minutes with GPT4, for instance. But those tedious deep dives are exactly what made me who I am today.
You should come guest post on the newsletter sometime.
I'd love to. Let's do it :)
Let's talk on LinkedIn?
TANSAAFL!
https://www.wikiwand.com/en/articles/No_such_thing_as_a_free_lunch
I also think we will find that using the internet as a data source is dubious at best. How can we tell a complex query/prompt leads to a solution that is valid. There's more detail in the code if anyone ever releases some. I suspect shenanigans in the algorithms to "help" it along. Again, hype out of control.
Also, Rice's theorem :) We can never trust an LLM that says the code does what it should do.
I have also evolved symbolic discriminators that use a BNF grammar as the source of mathematical functions. Grammars and compilers are my thing.
https://ieeexplore.ieee.org/abstract/document/1554748
Oh I think you're gonna love this project we (my wife and I) worked on during our PhDs (https://autogoal.github.io), it's grammatical evolution for evolving machine learning pipelines. Grammars and compilers are what I teach :)
The main paper is this https://www.sciencedirect.com/science/article/abs/pii/S0020025520306988
Early NN were universal function approximators, which have seriously good uses beyond theory. I suspect Newer LLMs are not much different conceptually. It's all very interesting to see, but the hype is out of control. They (those who write about it) say very fancy things like rationality and gigantic number of components. Throwing hardware at it will not go on forever. I think we are close to the limit, but I would put any bet on my predictions-- or anybody else's so far :-)
I agree, and I think we've already seen signals of diminishing returns in throwing compute at existing architectures. Halving the error on the most challenging reasoning benchmarks (which are still just benchmarks, thus gameable) is costing an order of magnitude more compute.
Nice summary. I (Bill White) worked in the AI/machine learning field for 20 years. We used NN to predict disease status from DNA/RNA (WIDE data). We also used evolutionary computing to EVOLVE NN nodes and structure. The main study was done by a PhD student in our lab. I was senior programmer. Here's the NN paper. and a list of my publications if you are curious.
https://link.springer.com/content/pdf/10.1186/1471-2105-4-28.pdf
https://scholar.google.com/citations?user=mb2m9ZEAAAAJ&hl=en
Thanks! There are some very impressive papers in that list! What made you leave AI?
multiple sclerosis
Oh, damn, so sorry to hear that.
It's a real bummer, and it never gets better. But I push through. I am down the rabbit hole of photography to fill the void.
I still think the 'retraining' is an element that will need to evolve in ANN's to make them more useful. The biological brain is always training, always running inference, at the same time. So the next phase in ANN's is to allow for this sort of mode, which also jacks up the energy requirement using current methods....which then leads me to the most remarkable thing about the human brain. It runs on the fuel of one banana while GPT and even Deep Seek still require massive amounts of grid power.
The brain is very complicated, we don't understand it, and it functions as a General Intelligence.
ANN's are a lot less complicated but we do understand them, (but we don't really understand the emergent properties).
In saying it's a long way from ANN's to AGI you're assuming all the complexity of the human brain is needed to make the GI. Maybe this isn't so?