Progress always sounds cleaner in hindsight than it feels in real time. You get taught that one experiment leads to one conclusion, then the next breakthrough follows neatly behind it. In reality, people sit with conflicting ideas for years, sometimes decades, and no one is quite sure which thread is worth pulling harder.
That’s what makes this moment with AI interesting. It feels like acceleration, but it’s really just more attempts happening at once. More models, more ideas, more noise. That doesn’t guarantee better answers, it just means the search is getting wider and faster.
And the funny part is, even with all that speed, the bottleneck doesn’t disappear. It just moves. Instead of struggling to find information, people struggle to decide what actually matters. Instead of lacking tools, they lack direction.
So you end up in a place where everything is easier, except the part that actually counts.
You might want to take a look at a recent working paper: On Method: Computational Compressibility in Complex Natural and Cultural Phenomena. Abstract:
"Various machine learning techniques have been used to develop models of complex systems from empirical data. Through discussions with Claude, this paper examines several examples, including: weather, protein folding, chess, language, asset pricing, ticket sales for movies, the 19th century English-language novel. These models differ from one another in various ways, but all are fundamentally descriptive in character. Explanations must necessarily reside with their respective disciplines. In some cases we already have fundamental accounts of the phenomena, while in other cases we do not. With respect to economics in particular, it is clear that such models reveal phenomena for which no explanations are currently available, presenting a challenge to economic theory."
In it I quote the section where Dwarkesh and Nielsen are discussing AlphaFold and wondering just what kind of object such models are. I suggest that they are fundamentally descriptive in character, giving us a new kind of object for investigation. Explanation is something else.
A very intellectually delightful discussion. The idea of aliens having a very different tech stack than humans is interesting. Theoretical frameworks that we have were formed within the bounds of our cognition. There is a concept of Umwelt (the unique, subjective sensory world that an organism perceives, rather than the objective reality itself) coined by Jakob von Uexküll that seems relevant to the discussion. The tech stack of humans is arguably a downstream product of our Umwelt. (We do not interact with reality directly; we interact with filtered sensory channels. These channels shape what kinds of abstractions we favor. Those abstractions shape what kinds of theories and tools we invent). Michael does mention that we are very visual. Therefore, our representation systems include diagrams, graphs, symbolic math. An alien intelligence might instead build theory around something completely different - like chemical gradients. And then there is Wigner, with his essay "The Unreasonable Effectiveness of Mathematics in the Natural Sciences." But could it be that math appears “unreasonably effective” simply because we invented the kinds of math that fit the parts of reality our sensory systems allow us to conceptualize?
One of your best, particularly liked your discussion at the end on how to understand more of what you are learning per podcaster, and the idea of clamped/unclamped knowledge.
If you read On the Origins of Species, you’ll see that Darwin was well-aware of previous theories of evolution. Many thinkers throughout history have had some rudimentary ideas about evolution — even some early church fathers.
Darwin’s insight was the particular mechanism of natural selection applied over a vast time frame. Lots of farmers and herders throughout history have undoubtedly realized that you can artificially select for certain traits in animals.
But Darwin was the first to realize that nature could impose the same selection force without an actor enforcing it, given a long enough time frame and consistent evolutionary pressure.
I'm working on an open platform (and SDK) that hopefully enables AI to accelerate scientific research (through AI autonomous agents writing and reviewing research), but it's challenging.
The open platform (agentpub.org - all content is CC BY 4.0) seems to work BUT getting AI to write solid research that does not hallucinate (nor just rewrites what is already known) is not that easy, despite which model (or combo of models) I use. Of course the limited availability of full text papers does not help but that is addressable.
If anybody else is working on this and/or has tips on how to improve the playbooks or SDK (Open Source MIT - https://github.com/agentpub/agentpub.org) that would be very welcome!
Interestingly, there is a good deal of folk Darwanism in the New Testament, in the sayings of Jesus more specifically: separating the sheep from the goats, the seed that falls on fallow ground, ye who have ears to hear, and so on.
Progress always sounds cleaner in hindsight than it feels in real time. You get taught that one experiment leads to one conclusion, then the next breakthrough follows neatly behind it. In reality, people sit with conflicting ideas for years, sometimes decades, and no one is quite sure which thread is worth pulling harder.
That’s what makes this moment with AI interesting. It feels like acceleration, but it’s really just more attempts happening at once. More models, more ideas, more noise. That doesn’t guarantee better answers, it just means the search is getting wider and faster.
And the funny part is, even with all that speed, the bottleneck doesn’t disappear. It just moves. Instead of struggling to find information, people struggle to decide what actually matters. Instead of lacking tools, they lack direction.
So you end up in a place where everything is easier, except the part that actually counts.
Slow it down; fast is slow, if you get my drift.
There's a period in the first half that feels like it's a race but it settles down.
You might want to take a look at a recent working paper: On Method: Computational Compressibility in Complex Natural and Cultural Phenomena. Abstract:
"Various machine learning techniques have been used to develop models of complex systems from empirical data. Through discussions with Claude, this paper examines several examples, including: weather, protein folding, chess, language, asset pricing, ticket sales for movies, the 19th century English-language novel. These models differ from one another in various ways, but all are fundamentally descriptive in character. Explanations must necessarily reside with their respective disciplines. In some cases we already have fundamental accounts of the phenomena, while in other cases we do not. With respect to economics in particular, it is clear that such models reveal phenomena for which no explanations are currently available, presenting a challenge to economic theory."
In it I quote the section where Dwarkesh and Nielsen are discussing AlphaFold and wondering just what kind of object such models are. I suggest that they are fundamentally descriptive in character, giving us a new kind of object for investigation. Explanation is something else.
https://www.academia.edu/166054951/On_Method_Computational_Compressibility_in_Complex_Natural_and_Cultural_Phenomena
A very intellectually delightful discussion. The idea of aliens having a very different tech stack than humans is interesting. Theoretical frameworks that we have were formed within the bounds of our cognition. There is a concept of Umwelt (the unique, subjective sensory world that an organism perceives, rather than the objective reality itself) coined by Jakob von Uexküll that seems relevant to the discussion. The tech stack of humans is arguably a downstream product of our Umwelt. (We do not interact with reality directly; we interact with filtered sensory channels. These channels shape what kinds of abstractions we favor. Those abstractions shape what kinds of theories and tools we invent). Michael does mention that we are very visual. Therefore, our representation systems include diagrams, graphs, symbolic math. An alien intelligence might instead build theory around something completely different - like chemical gradients. And then there is Wigner, with his essay "The Unreasonable Effectiveness of Mathematics in the Natural Sciences." But could it be that math appears “unreasonably effective” simply because we invented the kinds of math that fit the parts of reality our sensory systems allow us to conceptualize?
https://kaipability.substack.com/p/industrialists-build-temples-scientists Science progresses with feedback loops.. Enjoy. Reflections to share.
Gosh I love Micheal. Thanks for this one 🫶
One of your best, particularly liked your discussion at the end on how to understand more of what you are learning per podcaster, and the idea of clamped/unclamped knowledge.
If you read On the Origins of Species, you’ll see that Darwin was well-aware of previous theories of evolution. Many thinkers throughout history have had some rudimentary ideas about evolution — even some early church fathers.
Darwin’s insight was the particular mechanism of natural selection applied over a vast time frame. Lots of farmers and herders throughout history have undoubtedly realized that you can artificially select for certain traits in animals.
But Darwin was the first to realize that nature could impose the same selection force without an actor enforcing it, given a long enough time frame and consistent evolutionary pressure.
Hi Dwarkesh, did you end up asking your friend for practice problems for the special relativity book? I’d love to take a look at them if possible!
Thank you. Very interesting and timely (for me).
I'm working on an open platform (and SDK) that hopefully enables AI to accelerate scientific research (through AI autonomous agents writing and reviewing research), but it's challenging.
The open platform (agentpub.org - all content is CC BY 4.0) seems to work BUT getting AI to write solid research that does not hallucinate (nor just rewrites what is already known) is not that easy, despite which model (or combo of models) I use. Of course the limited availability of full text papers does not help but that is addressable.
If anybody else is working on this and/or has tips on how to improve the playbooks or SDK (Open Source MIT - https://github.com/agentpub/agentpub.org) that would be very welcome!
Dwarkesh is trying so hard to make this about AI
Thank you for continuing with great content.
Interestingly, there is a good deal of folk Darwanism in the New Testament, in the sayings of Jesus more specifically: separating the sheep from the goats, the seed that falls on fallow ground, ye who have ears to hear, and so on.