> So an experiment I've never done because I didn't have [the] compute would be this. Imagine if you could train a language model on all documents up to 1905, which is the year when Einstein had his miraculous year of four seminal papers. With that model, which is trained up to 1905, could you prompt the model to come up with a good explanation of the photoelectric effect, special relativity, this kind of stuff? And what would it take to rediscover these things?
FWIW there's a 2019 paper using word2vec that did this for mapping material property and composition and trained on literature up to a certain date and showed the embeddings could predict only recently discovered property-composition pairs.
Of course, I never saw any follow-up work that came out of that paper which does make one wonder...
Separately - loved the piece and focus on AI for Science! Agree that when you look at the best AI models, they always look different than what OAI and Anthropic are doing and require some integration of domain knowledge and engineering of AI models. Of course, the new paradigm is maybe these reasoning models can help you come up with domain-specific architectures/models.
Uhmmm so we have to redefine what's a PhD in 2027. If that's the case, even myself, a freshman, with enough dedication, strong field's foundations and without having to be on academia for 5+ years, I can be on pair with a PhD if I know how to steer these Research assistants. Exciting times :)
Love this perspective—AI as a 'power tool for information' is such a great way to frame it. The shift from just speeding up science to redefining what’s possible feels inevitable, especially in areas like materials or climate. The idea of novel discoveries? That’s where things get really exciting. Feels like we’re close to something truly transformative. Great read!
I struggle with the premise that a model should be generating insight on its own, or that this somehow is a limitation, because it's another instance of people imposing our own ideas about consciousness and human intelligence onto a bunch of weights.
Use these tools to generate insights, and we will find them helpful for generating insights.
Like, I wonder what would happen if you built some kind of knowledge graph about a certain area of research and basically put the model in a feedback loop and asked it to identify new nodes, justified with proposals and research questions, which it could attempt to answer in some limited sense.
It just feels like another one of those things that people quibble about to assert the ineffectiveness about LLMs in some area, only for someone to figure out the right application of these tools that unlocks it. Two years ago everyone was convinced that hallucinations would make models utterly incapable of deep research, and well here we are!
These are my own personal opinions, but I am most excited as a former materials grad student for a few things:
1. Converting multimodal unstructured data from the literature into structured data for analysis and modelling.
2. Enhancing the ability of scientists to develop quality open-source software, especially alternatives to dominant closed-source tools.
3. Reducing the writer's block when it comes to papers and grant proposals. Not replacing people entirely but helping with initial suggestions and especially formatting LaTeX.
This is an extremely interesting post. As a philosopher of science, I am not sure that looking at Kuhn "Structure" as a framework for the evolution of scientific theories is useful. While hugely popular, influential and stimulating, Kuhn's work is highly contentious and most philosophers (and historians) of science see it as problematic and superseded.
Yeah I definitely came across this. I had more drafts pretty much saying I don’t think Kuhn’s description applies with AI, but didn’t have enough substance as to feel confident in my point.
I’d love to read more on the topic if you want to share.
Thanks. Let me think about it! It's hard to find stuff that's relevant IMHO. PS I do sometimes use Kuhn in class - albeit tentatively - to illustrate the 'connectionist/neural nets' vs. 'rule-based/expert system' divide; in a way it's different paradigms; not necessarily incommensurable ones. One day maybe some historian will rewrite the history of AI in kuhnian terms ;-)
If you are impressed by Deep Research, I would very much recommend trying out Undermind's search engine: https://www.undermind.ai/
Unlike OpenAI's Deep Research it is only for the academic literature, but wow does it do it's job well - at least in the subject of my research where I am competent to judge (theoretical physics).
Love this perspective, that deep research helps with information but still doesnt get to insight. I found a very similar in my experiments with Deep Research
Superb novel representation of where the AI industry is at. The cost of knowledge is plummeting. Yet It still takes a skilled human to connect that knowledge to value. It all starts with personal abstractions that 0s and 1s just can't create at this point.
Your point about Deep Research being a powerful 'tool for information' rather than an 'insight engine' really resonated with me. The experiment you mentioned, training a model up to 1905 to see if it can rediscover Einstein's work, brilliantly highlights the limitations of current AI in making novel connections. Considering this, do you believe that the current limitations of AI in making novel discoveries are primarily due to the models themselves or is there something fundamental missing in the data or training methods?
To go to the next level AI will need to deeply understand what words and concepts mean. That will mean integration with simulators, ability to do experiments (even if virtual) and learn from outcomes and inspections.
Fair (though I included a podcast link and a link to several reviews). Briefly, Wootton is the extremely rare scholar who can read the major languages in which science as we know it now emerged in the 17thC. He frames his book around the observation that in 1600 educated people knew that the heavens revolved around the earth, that four humors governed the body, that werewolves prowled on the full moon, and crucially that Aristotle told you what you needed to know while by 1700 they conceived of the body as a machine, the earth as revolving around the sun, and that "discovery" of "facts" was an ongoing process of learning. He shows how the interaction of Columbus' "invention" of "discovery" (not a concept before 1500) with Gutenberg led natural philosophers to re-conceive of knowledge from interpretation of the Bible, Aristotle, and a few commentaries to *measuring* "experiments".
Where Kuhn is an inductive reasoner who uses examples to support principles, Wootton is a historian who walks us through the record to demonstrate what happened—why did it take decades, and what was happening in each decade.
Thanks for pushing me. On further thought, I'd say that Kuhn discusses shifts at the discipline level, while Wootton brings receipts on the one-time transformation of humanity from pre-scientific to scientific (and therefore industrial). Need to work on the phrasing, but that's the difference in kind, in one sentence.
The article discusses the difference between deep research, which focuses on collecting and summarizing information, and insight, which involves understanding and interpreting data. While deep research is essential for gathering facts and data, insight is crucial for deriving meaningful conclusions and gaining new knowledge. The author emphasizes the need for a balance between data collection and analytical thinking to advance scientific understanding.
> So an experiment I've never done because I didn't have [the] compute would be this. Imagine if you could train a language model on all documents up to 1905, which is the year when Einstein had his miraculous year of four seminal papers. With that model, which is trained up to 1905, could you prompt the model to come up with a good explanation of the photoelectric effect, special relativity, this kind of stuff? And what would it take to rediscover these things?
FWIW there's a 2019 paper using word2vec that did this for mapping material property and composition and trained on literature up to a certain date and showed the embeddings could predict only recently discovered property-composition pairs.
Of course, I never saw any follow-up work that came out of that paper which does make one wonder...
https://www.nature.com/articles/s41586-019-1335-8
Separately - loved the piece and focus on AI for Science! Agree that when you look at the best AI models, they always look different than what OAI and Anthropic are doing and require some integration of domain knowledge and engineering of AI models. Of course, the new paradigm is maybe these reasoning models can help you come up with domain-specific architectures/models.
(Spent several years writing about this on substack in a former life: https://ml4sci.substack.com/)
Uhmmm so we have to redefine what's a PhD in 2027. If that's the case, even myself, a freshman, with enough dedication, strong field's foundations and without having to be on academia for 5+ years, I can be on pair with a PhD if I know how to steer these Research assistants. Exciting times :)
Love this perspective—AI as a 'power tool for information' is such a great way to frame it. The shift from just speeding up science to redefining what’s possible feels inevitable, especially in areas like materials or climate. The idea of novel discoveries? That’s where things get really exciting. Feels like we’re close to something truly transformative. Great read!
I struggle with the premise that a model should be generating insight on its own, or that this somehow is a limitation, because it's another instance of people imposing our own ideas about consciousness and human intelligence onto a bunch of weights.
Use these tools to generate insights, and we will find them helpful for generating insights.
Like, I wonder what would happen if you built some kind of knowledge graph about a certain area of research and basically put the model in a feedback loop and asked it to identify new nodes, justified with proposals and research questions, which it could attempt to answer in some limited sense.
It just feels like another one of those things that people quibble about to assert the ineffectiveness about LLMs in some area, only for someone to figure out the right application of these tools that unlocks it. Two years ago everyone was convinced that hallucinations would make models utterly incapable of deep research, and well here we are!
These are my own personal opinions, but I am most excited as a former materials grad student for a few things:
1. Converting multimodal unstructured data from the literature into structured data for analysis and modelling.
2. Enhancing the ability of scientists to develop quality open-source software, especially alternatives to dominant closed-source tools.
3. Reducing the writer's block when it comes to papers and grant proposals. Not replacing people entirely but helping with initial suggestions and especially formatting LaTeX.
I have not digested this. Yet. I’m impressed by the goal of rethinking scientific progress in the light of emerging AI.
This is an extremely interesting post. As a philosopher of science, I am not sure that looking at Kuhn "Structure" as a framework for the evolution of scientific theories is useful. While hugely popular, influential and stimulating, Kuhn's work is highly contentious and most philosophers (and historians) of science see it as problematic and superseded.
Yeah I definitely came across this. I had more drafts pretty much saying I don’t think Kuhn’s description applies with AI, but didn’t have enough substance as to feel confident in my point.
I’d love to read more on the topic if you want to share.
Thanks. Let me think about it! It's hard to find stuff that's relevant IMHO. PS I do sometimes use Kuhn in class - albeit tentatively - to illustrate the 'connectionist/neural nets' vs. 'rule-based/expert system' divide; in a way it's different paradigms; not necessarily incommensurable ones. One day maybe some historian will rewrite the history of AI in kuhnian terms ;-)
If you are impressed by Deep Research, I would very much recommend trying out Undermind's search engine: https://www.undermind.ai/
Unlike OpenAI's Deep Research it is only for the academic literature, but wow does it do it's job well - at least in the subject of my research where I am competent to judge (theoretical physics).
Agree. I would say currently only Elicit.com's research report/systematic review can match Undermind.ai , but it's very pricey currently.
Love this perspective, that deep research helps with information but still doesnt get to insight. I found a very similar in my experiments with Deep Research
https://shamelmerchant.substack.com/p/the-first-spin-evaluating-openais
Superb novel representation of where the AI industry is at. The cost of knowledge is plummeting. Yet It still takes a skilled human to connect that knowledge to value. It all starts with personal abstractions that 0s and 1s just can't create at this point.
Your point about Deep Research being a powerful 'tool for information' rather than an 'insight engine' really resonated with me. The experiment you mentioned, training a model up to 1905 to see if it can rediscover Einstein's work, brilliantly highlights the limitations of current AI in making novel connections. Considering this, do you believe that the current limitations of AI in making novel discoveries are primarily due to the models themselves or is there something fundamental missing in the data or training methods?
Intuitions are the best verifiers out the. We don’t have a verifier for new ideas.
The Matrix is in the making :p
Fantastic, thank you for posting this…
To go to the next level AI will need to deeply understand what words and concepts mean. That will mean integration with simulators, ability to do experiments (even if virtual) and learn from outcomes and inspections.
Please replace references to Kuhn's _Structure_ with references to Wootton's The Invention of Science. It has surpassed Kuhn. https://www.inventionofscience.com . Podcast: https://www.pleaseexpand.com/david-wootton
you need to say more than just "its wrong" to convince me though, or more than just 1 person
I want to learn more but hard to discern which to read
Fair (though I included a podcast link and a link to several reviews). Briefly, Wootton is the extremely rare scholar who can read the major languages in which science as we know it now emerged in the 17thC. He frames his book around the observation that in 1600 educated people knew that the heavens revolved around the earth, that four humors governed the body, that werewolves prowled on the full moon, and crucially that Aristotle told you what you needed to know while by 1700 they conceived of the body as a machine, the earth as revolving around the sun, and that "discovery" of "facts" was an ongoing process of learning. He shows how the interaction of Columbus' "invention" of "discovery" (not a concept before 1500) with Gutenberg led natural philosophers to re-conceive of knowledge from interpretation of the Bible, Aristotle, and a few commentaries to *measuring* "experiments".
Where Kuhn is an inductive reasoner who uses examples to support principles, Wootton is a historian who walks us through the record to demonstrate what happened—why did it take decades, and what was happening in each decade.
Thanks. I get a lot of random inbound so the flesh to the rec goes a long way. It’s on my reading list now
The comparison at the end is particularly useful. I’ll probably have an easier time with a historian’s prose too.
Thanks for pushing me. On further thought, I'd say that Kuhn discusses shifts at the discipline level, while Wootton brings receipts on the one-time transformation of humanity from pre-scientific to scientific (and therefore industrial). Need to work on the phrasing, but that's the difference in kind, in one sentence.
Good start, and can’t wait to prompt engineer.
Fed the link and asked it to summarize:
The article discusses the difference between deep research, which focuses on collecting and summarizing information, and insight, which involves understanding and interpreting data. While deep research is essential for gathering facts and data, insight is crucial for deriving meaningful conclusions and gaining new knowledge. The author emphasizes the need for a balance between data collection and analytical thinking to advance scientific understanding.