In the spirit of "What could be problematic is that a frontier large language model trained on the technoscientific literature will only “know” the “what” that worked, but not the full story of “why” and “how” " ... perhaps it's not the lack of tacit knowledge that is limiting AI's ability to "make scientific breakthroughs" ... perhaps it its the dogmatic training biases we have instilled in it with "wrong" knowledge, or the inability to think as nature has intended which is mankind's innate strength ... see https://tinyurl.com/startwiththeanswer
Excellent essay. I really enjoyed the line from Goethe through Polanyi to Kuhn and Collins.
I kept thinking that Heidegger's being-in-the-world belongs somewhere in this story too. Hubert Dreyfus drew on both Heidegger and Polanyi decades ago to argue that intelligence comes from practical involvement in a world that resists us. And the social side of your argument has a Wittgensteinian feel to it: skill and meaning are learned inside shared practices.
It also made me wonder whether math and coding aren't really exceptions. They may just be the areas where AI already has something like a world to act in. It can try something, get an answer, fail, and try again. Run the code. Watch the tests fail. Check the proof in Lean.
So in those fields, the tacit knowledge doesn't have to be written down first. The machine can seek and blunder there, and some of that knowledge develops through the process itself.
Biology and materials science are different. Anything wet, physical, or difficult to simulate still leaves the machine working from cleaned-up records produced after the fact.
That seems close to the pattern we're seeing now: Erdős problems, yes. New wet-lab discoveries, not yet.
I'd say this article seems to conflate *tacit knowledge* (which can become accessible to AI only through a process of legibilisation, recording, or rediscovery of grey literature and humanly-latently-carried tacit knowledge) and *the need for exploration, experimentation, and trial and error*. Those are quite distinct in my view!
Exploration is crucial to uncovering new insight (and the practice of good experimentation lies in taking steps to make the most 'informative mistakes' you can). Tacit knowledge is one form that insights, new or old, can take. They're not really the same thing, though both can be bottlenecks to AI-driven scientific breakthroughs.
Documenting the failures is easy with AI-driven systems, if there is a source of ground truth. In this case it is possible to extract more tacit knowledge out of the weights than I ever thought possible, and also from the process of discovery.
It's appalling how miopic some of these 'thinkers' are: AI already solves Nobel-prize-winning problems, makes ground-breaking discoveries both in math, biology, physics, material sciences, also history, linguistics, neuroscience etc, etc... and they just outright dismiss or ignore this.
I sincerely hoped at Cosmos Institute they should know better. Or am I really expecting too much from an institute that boasts to be working at the frontier of AI research?..
If the craft part really can't be written down, then feeding models more papers was never going to produce breakthroughs. The more interesting bet is giving them instruments and letting them blunder, since that's the only way anyone, human or model, has ever picked up the knowledge Goethe was talking about.
Interesting, though I think it describes where we are today more than where we’re headed.
AI has already solved math problems in novel ways. That’s arguably a signal that interpretation may not remain an exclusively human domain for long.
Models, and especially world (3D) models, are becoming increasingly capable, self-improving, and fast. They’ll soon ask questions we never thought to ask, and generate hypotheses we wouldn’t have imagined.
The idea that AI will always need humans to interpret reality is probably an illusion.
In the spirit of "What could be problematic is that a frontier large language model trained on the technoscientific literature will only “know” the “what” that worked, but not the full story of “why” and “how” " ... perhaps it's not the lack of tacit knowledge that is limiting AI's ability to "make scientific breakthroughs" ... perhaps it its the dogmatic training biases we have instilled in it with "wrong" knowledge, or the inability to think as nature has intended which is mankind's innate strength ... see https://tinyurl.com/startwiththeanswer
Excellent essay. I really enjoyed the line from Goethe through Polanyi to Kuhn and Collins.
I kept thinking that Heidegger's being-in-the-world belongs somewhere in this story too. Hubert Dreyfus drew on both Heidegger and Polanyi decades ago to argue that intelligence comes from practical involvement in a world that resists us. And the social side of your argument has a Wittgensteinian feel to it: skill and meaning are learned inside shared practices.
It also made me wonder whether math and coding aren't really exceptions. They may just be the areas where AI already has something like a world to act in. It can try something, get an answer, fail, and try again. Run the code. Watch the tests fail. Check the proof in Lean.
So in those fields, the tacit knowledge doesn't have to be written down first. The machine can seek and blunder there, and some of that knowledge develops through the process itself.
Biology and materials science are different. Anything wet, physical, or difficult to simulate still leaves the machine working from cleaned-up records produced after the fact.
That seems close to the pattern we're seeing now: Erdős problems, yes. New wet-lab discoveries, not yet.
I wrote last year with some of this in mind: You Can't Skip Exploration (https://www.oliversourbut.net/p/you-cant-skip-exploration)
I'd say this article seems to conflate *tacit knowledge* (which can become accessible to AI only through a process of legibilisation, recording, or rediscovery of grey literature and humanly-latently-carried tacit knowledge) and *the need for exploration, experimentation, and trial and error*. Those are quite distinct in my view!
Exploration is crucial to uncovering new insight (and the practice of good experimentation lies in taking steps to make the most 'informative mistakes' you can). Tacit knowledge is one form that insights, new or old, can take. They're not really the same thing, though both can be bottlenecks to AI-driven scientific breakthroughs.
Documenting the failures is easy with AI-driven systems, if there is a source of ground truth. In this case it is possible to extract more tacit knowledge out of the weights than I ever thought possible, and also from the process of discovery.
It's appalling how miopic some of these 'thinkers' are: AI already solves Nobel-prize-winning problems, makes ground-breaking discoveries both in math, biology, physics, material sciences, also history, linguistics, neuroscience etc, etc... and they just outright dismiss or ignore this.
I sincerely hoped at Cosmos Institute they should know better. Or am I really expecting too much from an institute that boasts to be working at the frontier of AI research?..
If the craft part really can't be written down, then feeding models more papers was never going to produce breakthroughs. The more interesting bet is giving them instruments and letting them blunder, since that's the only way anyone, human or model, has ever picked up the knowledge Goethe was talking about.
Wait wait wait, are you telling me “The map is not the territory”?
Great explanation in this article.
We just finished a massive review of Tacit Knowledge, I invite you to take a look: https://curriculumredesign.org/wp-content/uploads/Tacit-Knowledge-CCR.pdf
Contact me if you wish, Charles.Fadel@CurriculumRedesign.org
Interesting, though I think it describes where we are today more than where we’re headed.
AI has already solved math problems in novel ways. That’s arguably a signal that interpretation may not remain an exclusively human domain for long.
Models, and especially world (3D) models, are becoming increasingly capable, self-improving, and fast. They’ll soon ask questions we never thought to ask, and generate hypotheses we wouldn’t have imagined.
The idea that AI will always need humans to interpret reality is probably an illusion.