Good article. I came to veganism though a similar logic. In general society the consensus view is that eating dogs, elephants, pandas etc is obviously immoral. For some reason we make an exception for cows and pigs, and view eating bacon as totally normal, despite no meaningful difference between a dog and a pig. It's currently contrarian to point this out, but the logical inconsistency will remain until eventually society at large realizes and also stops eating pigs.
The contrarian thinker who is an economist and blogger named Robin Hanson has written the following substack: https://www.overcomingbias.com/p/the-promise-of-polymath-llms. I'm interested in the section where Hanson speculates that large language models could produce new information. Do you know of any examples of this, or can you think of some ways that LLMs could do what Hanson suggests? I've pasted in the relevant section from Hanson's blog here [omitted].
Gemini's answer.
Large language models (LLMs) generate novel cross-disciplinary insights through Literature-Based Discovery (LBD)by mapping hidden semantic links and resolving logical contradictions between isolated research fields. This capability allows AI to identify overlooked intersections where the consensus of one discipline actively corrects or updates another. [1, 2, 3]
Current Real-World Examples
• Multi-Agent Scientific Frameworks: Google’s Co-Scientist system utilizes a "generate, debate, evolve" architecture. Multiple AI agents independently critique hypotheses from disparate disciplines to iron out contradictions. [4]
• Hybrid Biomedical LBD: Modern pipelines use LLMs to extract semantic relations from tens of thousands of unrelated medical abstracts. Systems like SKiM-GPT and LORE then cross-reference these fields to uncover unexpected drug repurposing targets and pathogenic pathways. [2, 5, 6, 7]
• Interdisciplinary Lay Syntheses: Recent studies prove that feeding LLMs parallel data from highly technical papers alongside simplified, lay summaries drastically increases the AI's "divergent thinking" capacity. This allows it to spot hidden conceptual clashes across domains. [8, 9]
•
Actionable Methods to Implement Hanson's Vision
To get an LLM to actively locate and resolve cross-field contradictions, you can deploy the following specific prompting and architectural workflows:
1 The Frictional Prompting Sequence
◦ Feed the model core definitions from Field A and Field B.
◦ Direct the LLM to actively search for zero-sum conclusions between them.
◦ Example: Ask the LLM to analyze human task optimization using the conflicting viewpoints of strict Behavioral Economics versus neural energy limits in Neurobiology. [9]
2
3 The Graph-Based ABC Discovery Loop
◦ Structure data so the LLM looks for indirect links.
◦ If Field A links to Concept B, and Concept B links to Field C, the LLM forces a direct synthesis between A and C.
◦ This bridges terminology gaps where different fields use different words for the exact same underlying mechanism. [1, 2, 6]
4
5 Iterative Consensus Refinement
◦ Prompt the LLM to adopt a contrarian persona rooted entirely in an outside field's vetted laws.
◦ Instruct it to review the target field's literature, rewrite its conclusions to align with the outside laws, and iteratively repeat the process across a web of interconnected fields.
How do you know how easy it is? What feedback do you get to indicate you're right, rather than just wrong while believing you're right? Do you have a larger sample size of other people applying this method?
I have trouble thinking of your work as primarily contrarian. Prediction markets seem to me straightforward economics and grabby aliens a reasonable analysis of the situation; same with cultural drift.
I see your contribution as the unusual insight that these things are likely to work or be true - they're fundamentally NEW ideas that weren't floating around (or if they were, lost in the noise floor). Not being *contrary* to consensus. I see you bringing new ideas into the scope of discussion. I don't mean from outside the Overton window - it's not like there were crazies talking about prediction markets and grabby aliens before you came along and legitimized those ideas. Instead, you're contributing new ideas.
Many of which seem, fairly obviously, correct or good ideas, and compatible with consensus knowledge. I see you as contributing creativity rather than contrarianism.
But then I may be an outlier - I see the world much as you do. (Tho I also think you are often wrong, for example about UFOs.)
--- Added: Most new ideas are bad ideas. Most people don't come up with even bad new ideas. You come up with new ideas that, more often than not, are good ideas. That's what makes you unusual. Not being contrary.
Very enlightening writing. It would be interesting to catalog the different ways to be a right contrarian. Another one that I can think of is to identify ideas that people believe mostly by analogy, then rebuild them from first principles.
Good article. I came to veganism though a similar logic. In general society the consensus view is that eating dogs, elephants, pandas etc is obviously immoral. For some reason we make an exception for cows and pigs, and view eating bacon as totally normal, despite no meaningful difference between a dog and a pig. It's currently contrarian to point this out, but the logical inconsistency will remain until eventually society at large realizes and also stops eating pigs.
I asked Gemini.
The contrarian thinker who is an economist and blogger named Robin Hanson has written the following substack: https://www.overcomingbias.com/p/the-promise-of-polymath-llms. I'm interested in the section where Hanson speculates that large language models could produce new information. Do you know of any examples of this, or can you think of some ways that LLMs could do what Hanson suggests? I've pasted in the relevant section from Hanson's blog here [omitted].
Gemini's answer.
Large language models (LLMs) generate novel cross-disciplinary insights through Literature-Based Discovery (LBD)by mapping hidden semantic links and resolving logical contradictions between isolated research fields. This capability allows AI to identify overlooked intersections where the consensus of one discipline actively corrects or updates another. [1, 2, 3]
Current Real-World Examples
• Multi-Agent Scientific Frameworks: Google’s Co-Scientist system utilizes a "generate, debate, evolve" architecture. Multiple AI agents independently critique hypotheses from disparate disciplines to iron out contradictions. [4]
• Hybrid Biomedical LBD: Modern pipelines use LLMs to extract semantic relations from tens of thousands of unrelated medical abstracts. Systems like SKiM-GPT and LORE then cross-reference these fields to uncover unexpected drug repurposing targets and pathogenic pathways. [2, 5, 6, 7]
• Interdisciplinary Lay Syntheses: Recent studies prove that feeding LLMs parallel data from highly technical papers alongside simplified, lay summaries drastically increases the AI's "divergent thinking" capacity. This allows it to spot hidden conceptual clashes across domains. [8, 9]
•
Actionable Methods to Implement Hanson's Vision
To get an LLM to actively locate and resolve cross-field contradictions, you can deploy the following specific prompting and architectural workflows:
1 The Frictional Prompting Sequence
◦ Feed the model core definitions from Field A and Field B.
◦ Direct the LLM to actively search for zero-sum conclusions between them.
◦ Example: Ask the LLM to analyze human task optimization using the conflicting viewpoints of strict Behavioral Economics versus neural energy limits in Neurobiology. [9]
2
3 The Graph-Based ABC Discovery Loop
◦ Structure data so the LLM looks for indirect links.
◦ If Field A links to Concept B, and Concept B links to Field C, the LLM forces a direct synthesis between A and C.
◦ This bridges terminology gaps where different fields use different words for the exact same underlying mechanism. [1, 2, 6]
4
5 Iterative Consensus Refinement
◦ Prompt the LLM to adopt a contrarian persona rooted entirely in an outside field's vetted laws.
◦ Instruct it to review the target field's literature, rewrite its conclusions to align with the outside laws, and iteratively repeat the process across a web of interconnected fields.
6
[1] https://elicit.com
[2] https://www.mdpi.com
[3] https://hai.stanford.edu
[4] https://www.nature.com
[5] https://pmc.ncbi.nlm.nih.gov
[6] https://www.biorxiv.org
[7] https://pmc.ncbi.nlm.nih.gov
[8] https://www.nature.com
[9] https://www.nature.com
> Which often isn’t that hard.
How do you know how easy it is? What feedback do you get to indicate you're right, rather than just wrong while believing you're right? Do you have a larger sample size of other people applying this method?
I have trouble thinking of your work as primarily contrarian. Prediction markets seem to me straightforward economics and grabby aliens a reasonable analysis of the situation; same with cultural drift.
I see your contribution as the unusual insight that these things are likely to work or be true - they're fundamentally NEW ideas that weren't floating around (or if they were, lost in the noise floor). Not being *contrary* to consensus. I see you bringing new ideas into the scope of discussion. I don't mean from outside the Overton window - it's not like there were crazies talking about prediction markets and grabby aliens before you came along and legitimized those ideas. Instead, you're contributing new ideas.
Many of which seem, fairly obviously, correct or good ideas, and compatible with consensus knowledge. I see you as contributing creativity rather than contrarianism.
But then I may be an outlier - I see the world much as you do. (Tho I also think you are often wrong, for example about UFOs.)
--- Added: Most new ideas are bad ideas. Most people don't come up with even bad new ideas. You come up with new ideas that, more often than not, are good ideas. That's what makes you unusual. Not being contrary.
Very enlightening writing. It would be interesting to catalog the different ways to be a right contrarian. Another one that I can think of is to identify ideas that people believe mostly by analogy, then rebuild them from first principles.