The hypernormal science pattern you describe applies beyond academia. Organizations do the same: optimize within the existing framework, measure faster, produce more - while the capacity to ask whether the framework itself needs replacing quietly erodes. The London Underground metaphor works in both directions.
Very interesting! While coming up with new paradigms may be very difficult for an AI, surely it wouldn’t have to rely on “AI peer review” to test them.
If it was advanced enough, it could test whether the new paradigm explained more of the data than the existing paradigm (classic Kuhn).
Using AI to identify datasets/findings that seem contradictory under the existing paradigms may be a good place to start.
Great article! A recurring theme I see when it comes to applying AI in science is a failure to reason beyond known knowledge.
Unknown Unknowns is what I call this domain.
Your thoughts about the stubbornness or resistance of a new paradigm to take root in the minds of scientist, reminds me of the allegory of Plato's Cave.
One lesser known fact about the allegory is when the person who managed to go outside and see the outside world comes back and tries to explain that the world is better, and therefore attempts to remove them from the cave, his fellow people in the cave, attack him for it. A form of primitive tribalism.
There exists a burden to enlightenment, we need to learn how to present it to others, in a way that can be easily understood.
★ If the bottleneck of the way the majority of AI models and agents work today is leading to a compression of opinion, coupled with uncovering blank regions of an existing map, what is needed in your opinion?
★ How much influence does first principles thinking have on disruptive science?
As you mentioned about the fact that, the one who held onto the idea of the 'ether', was not able to make the leap to that Einstein did about space-time.
Asking as a curious investigator who wants to do valuable work in science.
Thanks Rohit! I think on your first question - it’s legitimately hard, but one thing is we do need pathways to get people in, sensibly, with outside ideas. The problem is of course assessing which of these is any good, which we could do in some areas where reasonably objective standards exist via for example competitions, or other fast pathways. I did have some thoughts in this earlier article: https://worksinprogress.co/issue/why-science-needs-outsiders/
On your other question - yes I think first principles thinking has a big impact, mainly in the sense that it’s closer to root observations and thus less reliant on preexisting concepts
The talk about rules of science anticipates the work of Douglas Lenat and his Eurisko discovery system. In that work, stemming from his PhD work AM at Stanford, he devised a collection of heuristics (rules) for discovery and unleashed them on several projects, one a naval fleet, one a novel high-rise VLSI design. Eurisko would create a game to play against it to refine findings. This all dates to the 1970's. What he repeatedly discovered was his system's inability to think beyond its own horizons - it lacked common sense. Thus he abandoned that work and started the Cyc common sense encyclopedia project.
Eurisko is finding its way onto Github, as did an early version of OpenCyc.
Muy buen artículo. Creo que el problema es extensivo sobre todo a las ciencia s dela complejidad, a aquellas que como la ecología, la economía o la biología, operan con múltiples variables y escalas, con muchas trayectorias posibles para un mismo sistema.
New paradigms require the invention of new concepts, new words, new abstractions. Indeed, AI trained in existing data and concepts cannot easily go beyond it.
One possible way may be having the AI look for correspondences between mathematical objects and experimental data, since there is a lot of math that has been invented/discovered but hasn't yet been applied to science.
AI is architecturally constrained to work within existing symbolisation. A great tool for exploring existing symbolisation, but we need humans if we want to go beyond it.
The map tale I heard was Russian cartographers who mapped their domain at progressively finer scales until one day they realised the terrain was it's own map at 1:1, and so they declared their task complete.
We must at least build systems that support visionary thinking, which will likely include machines. For now, humans are the best at this but they are very poorly supported.
You’ve rediscovered Boden’s three types of creativity. AI excels at combinational and exploratory creativity, can’t do transformational. A literature review is in order.
Boden’s framing is rather good. A luminary in creativity studies who happens to have been working in computational and generative creativity for decades, her work really ought to be required reading for everyone touching on those fields.
If we can gain diverse perspectives through multimodality and use language appropriately, we may, in principle, be able to develop metacognitive abilities.
However, I feel that many people overlook the importance of interoception. Having a mechanism to constantly monitor one’s own state is crucial for a feedback system to possess intelligence.
The hypernormal science pattern you describe applies beyond academia. Organizations do the same: optimize within the existing framework, measure faster, produce more - while the capacity to ask whether the framework itself needs replacing quietly erodes. The London Underground metaphor works in both directions.
Very interesting! While coming up with new paradigms may be very difficult for an AI, surely it wouldn’t have to rely on “AI peer review” to test them.
If it was advanced enough, it could test whether the new paradigm explained more of the data than the existing paradigm (classic Kuhn).
Using AI to identify datasets/findings that seem contradictory under the existing paradigms may be a good place to start.
Great article! A recurring theme I see when it comes to applying AI in science is a failure to reason beyond known knowledge.
Unknown Unknowns is what I call this domain.
Your thoughts about the stubbornness or resistance of a new paradigm to take root in the minds of scientist, reminds me of the allegory of Plato's Cave.
One lesser known fact about the allegory is when the person who managed to go outside and see the outside world comes back and tries to explain that the world is better, and therefore attempts to remove them from the cave, his fellow people in the cave, attack him for it. A form of primitive tribalism.
There exists a burden to enlightenment, we need to learn how to present it to others, in a way that can be easily understood.
★ If the bottleneck of the way the majority of AI models and agents work today is leading to a compression of opinion, coupled with uncovering blank regions of an existing map, what is needed in your opinion?
★ How much influence does first principles thinking have on disruptive science?
As you mentioned about the fact that, the one who held onto the idea of the 'ether', was not able to make the leap to that Einstein did about space-time.
Asking as a curious investigator who wants to do valuable work in science.
Thanks Rohit! I think on your first question - it’s legitimately hard, but one thing is we do need pathways to get people in, sensibly, with outside ideas. The problem is of course assessing which of these is any good, which we could do in some areas where reasonably objective standards exist via for example competitions, or other fast pathways. I did have some thoughts in this earlier article: https://worksinprogress.co/issue/why-science-needs-outsiders/
On your other question - yes I think first principles thinking has a big impact, mainly in the sense that it’s closer to root observations and thus less reliant on preexisting concepts
The talk about rules of science anticipates the work of Douglas Lenat and his Eurisko discovery system. In that work, stemming from his PhD work AM at Stanford, he devised a collection of heuristics (rules) for discovery and unleashed them on several projects, one a naval fleet, one a novel high-rise VLSI design. Eurisko would create a game to play against it to refine findings. This all dates to the 1970's. What he repeatedly discovered was his system's inability to think beyond its own horizons - it lacked common sense. Thus he abandoned that work and started the Cyc common sense encyclopedia project.
Eurisko is finding its way onto Github, as did an early version of OpenCyc.
Muy buen artículo. Creo que el problema es extensivo sobre todo a las ciencia s dela complejidad, a aquellas que como la ecología, la economía o la biología, operan con múltiples variables y escalas, con muchas trayectorias posibles para un mismo sistema.
New paradigms require the invention of new concepts, new words, new abstractions. Indeed, AI trained in existing data and concepts cannot easily go beyond it.
One possible way may be having the AI look for correspondences between mathematical objects and experimental data, since there is a lot of math that has been invented/discovered but hasn't yet been applied to science.
AI is architecturally constrained to work within existing symbolisation. A great tool for exploring existing symbolisation, but we need humans if we want to go beyond it.
> AI is architecturally constrained to work within existing symbolisation
Is it? I thought the whole point of encoder/decoder architectures was that AI could reframe data into a new set of symbols.
That’s mapping symbols to symbols, not creating new ones.
The map tale I heard was Russian cartographers who mapped their domain at progressively finer scales until one day they realised the terrain was it's own map at 1:1, and so they declared their task complete.
"In other words, we must build visionary machines rather than merely predictive ones."
Must we really, though?
We must at least build systems that support visionary thinking, which will likely include machines. For now, humans are the best at this but they are very poorly supported.
You’ve rediscovered Boden’s three types of creativity. AI excels at combinational and exploratory creativity, can’t do transformational. A literature review is in order.
Boden’s framing is rather good. A luminary in creativity studies who happens to have been working in computational and generative creativity for decades, her work really ought to be required reading for everyone touching on those fields.
Alas it’s not, and I totally agree. Boden gets some recognition, but not for this model.
If we can gain diverse perspectives through multimodality and use language appropriately, we may, in principle, be able to develop metacognitive abilities.
However, I feel that many people overlook the importance of interoception. Having a mechanism to constantly monitor one’s own state is crucial for a feedback system to possess intelligence.