Thomas

[Thomas @Thomas40586206]

@SebastienBubeck There is a story of Benson Farb trivializing a whole field by using an analogy. Or Bob MacPherson also easily solving an applied math problem by seeing the underlying geometry. No need for AI.

[Thomas @Thomas40586206]

@littmath @SebastienBubeck @Jacob_Tsimerman I mean on the human side.

[Thomas @Thomas40586206]

@littmath @SebastienBubeck @Jacob_Tsimerman I am not clear on why a super intelligence would care about doing math as it seems very idiosyncratic and motivated by misunderstanding.

[Thomas @Thomas40586206]

@littmath @SebastienBubeck @Jacob_Tsimerman So far, we are talking about well established human stated conjectures.

[Thomas @Thomas40586206]

@littmath Last mile work is being rewarded disproportionately by math history. There are entertaining talks of Arnold about this.

[Thomas @Thomas40586206]

@littmath It has been this way in the math physics community for a while: people don’t talk about work in progress because they will get scooped.

[Thomas @Thomas40586206]

@__alpoge__ @ben_golub Can you be clear about your setup? Model + prompt. Can this be achieved with usual fable access or are you in anyway using some advantages you have from working with Anthropic?

[Thomas @Thomas40586206]

@TheZvi This denotes a rather deep misunderstanding of mathematicians: honey badger is the mascot. Stories of Soviet mathematicians and through what they kept doing math come to mind.

[Thomas @Thomas40586206]

@littmath It��s nice but of course you need the audience to be at all interested. Not always the case even in high places

[Thomas @Thomas40586206]

@henryquantum It’s a bit like a favorite movie. Sometimes I think if you have a favorite movie or a favorite book, it probably means that you don’t like books (or movies) that much.

[Thomas @Thomas40586206]

@henryquantum The issue is that mathematicians already write very poorly most of time, for most of the papers. Writing is part of the culture but not writing well, unfortunately. It is true that it is easier to make an AI write well than a human write well though.

[Thomas @Thomas40586206]

@SebastienBubeck Also it is silly to try to write such results like a conventional human-written math paper, especially if the results compile in lean. Best to actually share the human part of the experience ie prompts, waiting time, methods…

[Thomas @Thomas40586206]

@SebastienBubeck For the same price, you could make the paper readable and train the model to actually explain the math in an understandable way and not follow the standard conventions of the math community.

[Thomas @Thomas40586206]

@littmath @gro_tsen @itaisher Of course it is different and all the examples I mentioned were different from each others and one needs (the math community) as a whole to reflect on it. But it is not obvious to me that this is not again redefining the “trivial” line. (Might be off, not read the whole thread)

[Thomas @Thomas40586206]

@littmath @gro_tsen @itaisher Is that any different than the arrival of logarithm tables, calculators, numerical analysis? All trivialized tasks that were considered non-trivial before their arrival.

[Thomas @Thomas40586206]

@littmath Probably not the right message to follow up on with this. Nevertheless, it would be great for people to share the exact prompt and set up that led to the output they are presenting. Usually they mention the model but the prompt and attachments are important as well.

[Thomas @Thomas40586206]

@JarekLiesen I understand Oxford benefits from free access to ChatGPT pro. Is that what you used for it?

[Thomas @Thomas40586206]

@JarekLiesen Please share the exact set up including the exact prompt and potential attachments etc.

[Thomas @Thomas40586206]

@littmath @__alpoge__ Did they post the fable prompt? I am assuming nothing special.

[Thomas @Thomas40586206]

@littmath By the way, openai approach is common in ai labs and is actually problematic. For example: who to get in touch with about the paper? Who is the lead?

[Thomas @Thomas40586206]

@littmath Regarding attribution and authorship, Ethan Mollick was suggesting a Roman analogue: "You didn't make the thing, but you commanded it be done, so the acknowledgement is "Ego hoc fieri iussi."" I am not sure about it but that's the most interesting suggestion I've seen so far.