Does AI make you concerned?

[crusher]

@TroubleWithTribbles

@Andrew

From memory, the reliability of currently released models on difficult (PhD-level and beyond) science questions is on the order of 70%, which is comparable to human experts trained in the same topics, and they can do that in 10 seconds rather than the 30 minutes or whatever it might take a human expert to arrive at the same answe

I play around with it some in medical communications. It does a decent job on topics I am interested in, but the ones I use are still pretty sketchy when it comes to citing sources. It does a pretty good job summarizing shorter clinical content pieces, like scientific abstracts and to a lesser extent full. I don’t use the content directly, but it can help in deciding on how to approach a complex text, especially if you provide follow-up directions to dial in the generated response. I have had limited success with tasks like creating a spreadsheet from a list of publications. I don’t usually need that because reference management tools usually have that function, but if I have a list of a bunch of different content types that I haven’t included in EndNote or whatever, it can provide a good starting point. I have also used it to generate spreadsheets of scientific conference with dates, venues, abstract deadlines, etc. Again, it can provide a decent starting place. The most interesting experience with it was having it create a spreadsheet based on a long list of conference abstracts with columns for clinical area, main focus, AMA formatted citation, link to abstract, and a 59-word summary of the abstract. It actually provided a decent starting place for this. I then asked it to add a column that inidcated whether the abstract focused on specific key terms, like a certain drug, device, test, or demograohic feature. For some reason it had a harder time with this. A big frustration is that the one I used (chatGPT) tended to forget instructions with iterative prompting. Performance varied inversely with the length and the complexity of the prompt and the content being anlyzed.

ChatGPT has a word length limit of around 1000 words per time which is pretty pathetic. It also will after awhile start a new chat with no memory of the old one. It’s AI art is limited to around 8 per day and it puts each idea out in pairs. Yes, I was a paid member i Just cancelled it. Do things like ask it for a representation of a person writing with a left hand and it keeps coming back with right hands.
Probably because it’s coming back with the mean answer and not outliers.

[TroubleWithTribbles]

And each bad reponse is charged against the limit. I recently dropped my paid account. My employer has an inhouse chatbot that is running on an older version of chatGPT, but it won’t ouput files.

[TroubleWithTribbles]

@crusher

Do things like ask it for a representation of a person writing with a left hand

Or ask it to show a train coming off the rails and it keeps the wheels firmly planted for 3 iterations until you yell at it:-). Sometimes it tells me it can’t do something but then it can when I basically remind it that it can. And that gets charged as 2 prompts when only one response was useful. Or however they count prompts/reponses.

[hansioux]

https://www.nature.com/articles/s41467-024-54178-1

Having AI design something that works and have no idea why it works. The holy grail of engineering.

[Brianjones]

This is definitely an issue on YouTube and Facebook and also some other sites I use.

AI-generated ‘slop’ is slowly killing the internet, so why is nobody trying...

[crusher]

Does anyone know the best AI for marketing, Such as finding out which countries to Target for certain interests.?

[nivek]

Another interesting article suggesting capital will matter more than ever in a post AGI world…

Labour-replacing AI will shift the relative importance of human v non-human factors of production, which reduces the incentives for society to care about humans while making existing powers more effective and entrenched

https://www.lesswrong.com/posts/KFFaKu27FNugCHFmh/by-default-capital-will-matter-more-than-ever-after-agi

[hannes]

Just received an e-mail from a guy who would like to cooperate with us.

The letter was most certainly written by AI and it’s such a mouthful of hyperbole. Hard to digest and totally inappropriate, cause it misses any personal touch that might make me interested in responding to this person.

Just another example of AI going down the wrong path.

[TT]

ChatGPT has been awful lately

Me: do this relatively simple thing
Chatgpt: sure, I can fuck that up for you
Me: no, you didn’t do it correctly. Like i told you, try again
Chatgpt: oh, you’ve used up your free use for today, try again tomorrow!

[Brianjones]

That’s how they move you to be a paying customer. How about the other platforms ?

[TT]

I guess i should try one

Which do you recommend?

[Brianjones]

I dunno, I saw there’s a Chinese one that’s just as good, deepseeker. Gemini. What else ? Grok ?

[TroubleWithTribbles]

That is actually how they lost me as a paying customer.