Is this about things that linguists don't know, or just humans as users of language don't know? Because linguists know a lot about the way people speak that people themselves don't realize and have wrong models about.

(Example: if I recorded you saying the word "spit", cut out the recording at the end of the S sound and played only what comes after that for someone else, what word do you think they'd say they heard?)

In conclusion, is this market just "LLMs will make discoveries in linguistics"?

I'd bet higher but for uncertainty around subjectivity of resolution and how mana-profits will depend on what other people think in short~medium term.

Some evidence that doesn't feel super central to my confidence: GPT-3 is better than me at next word prediction.

@NoaNabeshima --out of date--

The claim is probably true, but I don't expect it to be proven in 2023.

@ShadowyZephyr Market close is 2100

@Dreamingpast How exactly do you plan to determine how this market resolves? Like, does the discovery have to be noted in a peer-reviewed journal, or just a popular tweet, etc.?

@Yitz yeah humans have to learn about a LLM used structure or patten generally found in language that we didn't know about

@Yitz anything credible goes, journal, conference, peer reviewed paper, something credible and hopefully verifiable

There's lots of things scientists are too dumb to notice. Language is a hierarchy of concepts, the A words absence, table etc refer to nothing being there, the b words refer to being, more advanced concepts with further letters.

@MarkIngraham scientists may be dumb but llms pick up on regularity in language far far better than expected.

@Dreamingpast examples?

@MarkIngraham it's interesting that regular spoken English is already the most efficient language, superior to any machine language

@MarkIngraham Learning involves compression of information to form abstract concepts. In its training, ChatGPT (3.5) has somehow “implicitly discovered” whatever regularities in language (and thinking) make this possible, and that is why it is so good (gpt 4 etc as well)

@MarkIngraham interesting that natural language is not as powerful as any programming language, especially because the programming languages are SOOO optimized for the hardware they run on, compared to say a sentence in English, so why would you say that?

@Dreamingpast gpt is good at imitating a Karen that gives evasive answers to customer service. It lacks a few basic features like being able to use exponents when doing math

@MarkIngraham Have you not seen their integration with wolfram they announced yesterday?

@MarkIngraham

@Dreamingpast And it doesn't just fall back on wolfram to give these answers. It actively integrates the information and also determines when to ask/defer to wolfram, and has ability to use the results provided by wolfram to do further stuff with it

@Dreamingpast where can I try it

@MarkIngraham "Note that this capability is so far available only to some ChatGPT Plus user" so prolly in april it should start being available broadly

@Dreamingpast it'll be available in chatGPT only

@Dreamingpast it is so painful to use gpt if it can't do grade school math

@Dreamingpast

@MarkIngraham But now it can determine whether it is very shit at doing math and then determine to call for better answers using wolfram, learn that, and integrate that into itself.

oh, and the new upgrade (you're using march 14 one) will allow it to get real time data as well from wolfram, such as

@Dreamingpast https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers/

@Dreamingpast OK I'll try it in April and see if it doesn't fuck up

@MarkIngraham Wolfram is kinda famous for how good it is at step wise solving math given natural language so i doubt itd be anything bad now, so lets see!!

@Dreamingpast yes it would also save a ton of money, wolfram is stupid to do this

@MarkIngraham I can imagine that OpenAI cut a really profitable deal with Wolfram for this, cuz they desperately needed credible math and equations solutions and answers and step by step proofs to extend the competitive edge. can't say it''d have been a bad pay day, especially when ChatGPT has been around for SO LITTLE

@Dreamingpast if you could get gpt to use the wolfram math plus programming tools, plus salesforce or some other product it would be actual product and not a meme

@MarkIngraham Literally that's exactly what they're doing AND funny you mention salesforce as they're moving very fast into integrating LLMs into POCs for internal systems and also : https://www.salesforce.com/news/stories/chatgpt-app-for-slack/

@Dreamingpast if I can just put a invoice into salesforce and gpt copies it it would be great. They are supposed to already do that. But also a salesforce to netsuite integration so all the financial data is in one place with no manual steps

@MarkIngraham Tbh they're not "supposed" to do all this so fast which is half the reason its super impressive. But yeah i agree, and think the year will def end having seen all this happen

@Dreamingpast the only other thing is a ad content integration. I sell ads on Akamai thru salesforce and this all gets integrated. It should just be able to integrate everything

@MarkIngraham whats akamai and ad content generation currently isnt handled by LLMs at all cuz of a lack of videos or images but the content generated for non cheesy ads is actually good. Also, "It should just be able to integrate everything" are we hiring an intern or what

@Dreamingpast just make it do everything, I don't care if an Indian is doing it for 10 cents an hour

@MarkIngraham lmao what

@MarkIngraham and whats an akamai

@Dreamingpast is what i get via google but doesnt fit into your sentence

@Dreamingpast it's the ad content. American express buys ad, store it in akamai, we give them a salesforce fake page to preview it, they approve the purchase, we send the akamai link to the front-end and use bill.com to invoice them, then record the revenue in netsuite. Gpt should do all this