kravietz š¦ on Nostr: A very interesting yet highly technical article by Stephen Wolfram, one of the most ...
A very interesting yet highly technical article by Stephen Wolfram, one of the most brilliant mathematicians, scientists and businessmen of our generation, creator of the Wolfram Alpha calculation and reasoning engine.
Wolfram describes here an interesting case of a proof computed by an automatic theorem proving system. The proof is correct but at the same time so complex and extensive (about 80ā000 operations) that it is beyond the capacity of human perception. How do you āacceptā such proof and how do you verify it, wonders Wolfram.
But these challenges arenāt limited to mathematical proofs. A large part of human creativity today is the analysis of huge data sets and the synthesis of conclusions - law, medicine, parts of philosophy, legislation and many other fields are excellent candidates for the application of LLM models, commonly referred to as āAIā.
Energy or agricultural policy are ideally suited for this because we mainly have various physical parameters and politically set assumptions on the inputs, such as acceptable levels of risk (e.g. blackout or drought). We had been actually facing the same problem as described by Wolfram for decades now, only in relation to human-written policies. The amount of legislation in any state or federation (UE or US) is so enormous, that not a single person can comprehend it all and very few people are able to see the whole picture even in a single sector. But thatās easy, because theyāre written by humans. We know how to deal with such cognitive challenges - we just dismiss them as written by āidiotsā, āthievesā etc.
That seems like a perfect use case for LLM, which can deal with literally inhumane amounts of data. The problem will arise when LLMs start producing policies that will be perfectly correct and work within the assumed parametersā¦ but will be so complex that none of us will understand them.
The problem, of course, will not be that these policies will be incorrect, it will be how much we will be able to accept their conclusions given that our cognitive biases co-construct exactly what we proudly call ācultureā and ātraditionā .
https://writings.stephenwolfram.com/2025/01/who-can-understand-the-proof-a-window-on-formalized-mathematics/
Wolfram describes here an interesting case of a proof computed by an automatic theorem proving system. The proof is correct but at the same time so complex and extensive (about 80ā000 operations) that it is beyond the capacity of human perception. How do you āacceptā such proof and how do you verify it, wonders Wolfram.
But these challenges arenāt limited to mathematical proofs. A large part of human creativity today is the analysis of huge data sets and the synthesis of conclusions - law, medicine, parts of philosophy, legislation and many other fields are excellent candidates for the application of LLM models, commonly referred to as āAIā.
Energy or agricultural policy are ideally suited for this because we mainly have various physical parameters and politically set assumptions on the inputs, such as acceptable levels of risk (e.g. blackout or drought). We had been actually facing the same problem as described by Wolfram for decades now, only in relation to human-written policies. The amount of legislation in any state or federation (UE or US) is so enormous, that not a single person can comprehend it all and very few people are able to see the whole picture even in a single sector. But thatās easy, because theyāre written by humans. We know how to deal with such cognitive challenges - we just dismiss them as written by āidiotsā, āthievesā etc.
That seems like a perfect use case for LLM, which can deal with literally inhumane amounts of data. The problem will arise when LLMs start producing policies that will be perfectly correct and work within the assumed parametersā¦ but will be so complex that none of us will understand them.
The problem, of course, will not be that these policies will be incorrect, it will be how much we will be able to accept their conclusions given that our cognitive biases co-construct exactly what we proudly call ācultureā and ātraditionā .
https://writings.stephenwolfram.com/2025/01/who-can-understand-the-proof-a-window-on-formalized-mathematics/