I have long since said that AI is not useful in the use cases that we encounter it. This is my considered opinion as a scholar and an educator. The only way we can apply the usefulness claim is if we expand the term AI to cover every ML application and at that point, it's sparkling AI hype. - ThreadSky

shengokai.bsky.social • 9 days ago

I have long since said that AI is not useful in the use cases that we encounter it. This is my considered opinion as a scholar and an educator.

The only way we can apply the usefulness claim is if we expand the term AI to cover every ML application and at that point, it's sparkling AI hype.

Reposted from Anil Dash

I don’t think there’s a single serious critic who says AI’s not useful? The concerns I see are about content rights, worker impacts, environmental effects, epistemological threats, or media manipulation. But that’s all predicated on the idea that it *does* work, often too well for people to manage…

Comments

fireflows.bsky.social•9 days ago

AI is WAY over hyped, but it has a lot of use in tech. No, a random Joe can't tell ChatGPT to make them an app, but a skilled programmer will be more productive using AI tools.

I don't necessarily agree with Anil Dash though, especially the "works too well" part.

davidash.bsky.social•7 days ago

Speaking as a developer, in my experience AI just makes you a worse developer. It's not reliable because it doesn't know when it's wrong. Which means every line of code it generates has to be reviewed and tested and possibly fixed just like your own. It's not faster, it's just outsourcing.

anildash.com•8 days ago

I think you get the thing I’m getting at though. It works too well to dismiss as useless.

rowyourbot.bsky.social•8 days ago

I wish there was a detailed piece about its uses outside the ones that are don’t make sense, or render bad results so people could understand this. We’re continually being pitched certain useless or harmful applications but there are other uses.

anildash.com•8 days ago

I think the most common non-harmful/useless one is its utility in helping coders, where it’s most effective as an assistant.

rowyourbot.bsky.social•8 days ago

Right. There it is useful to speed up work. The adoption of it all over the place, even when it is not really doing things well is a situation people worry about—things get automated, and then frustration and helplessness is thrown onto the user. It’s forced on everyone, becomes a time suck, etc.

katjathieme.bsky.social•8 days ago

As an assistant for data analysis, too. Not everyone who needs to run numbers for a research or work project is fully trained in stats packages.

rowyourbot.bsky.social•8 days ago

This is the other use. The main real uses are ‘used by experts’ uses like 1) speeds up coding 2) data analysis (LMs but not LLMs?) 3) translation. Then they have a math one for advanced math that is 50% correct last I looked. It’s just a hodgepodge of various things. But it has expanded to more.

dingemansemark.bsky.social•8 days ago

Should people who need to 'run numbers' without being 'fully trained in stats' use a text regurgitator that cannot run numbers? Or do you mean this as an extension to the comment on coding assistance, essentially using chatgpt the way folks used to use stackoverflow?

fireflows.bsky.social•8 days ago

This is the only pro-AI article I've seen with a realistic scope written by a real person who uses LLMs rather than a grifter trying to sell you magic beans.

anildash.com•8 days ago

I don’t like that piece. Its aim is to denigrate critics rather than to illuminate usage. Simply visit @simonwillison.net’s blog. He’s been cataloguing the state of the space daily for years.

fireflows.bsky.social•8 days ago

That's fair. I'll check out Simone Williston.

feralrobots.bsky.social•8 days ago

I was gonna recommend Simon. He's one of the few people I've seen who seems to have identified safe use cases, & he's been pretty explicit about how he makes them work - & he's still critical. Helps that he starts from a baseline position of credibility.

katjathieme.bsky.social•8 days ago

Doesn’t have to be ChatGPT. A random Joe can tell another gen AI app to make an app, for sure. Apps can be small things. It’s used and useful that way.

spavel.bsky.social•8 days ago

Studies have been showing a ~41% increase in bugs from AI-generated code, such that almost all the productivity gains are eaten up figuring out what the AI is trying to do, find mistakes, and fix them. Not to mention LLMs making up fake results from fake tests to pretend that everything works.

jkohlmann.bsky.social•8 days ago

Generative AI is a know-nothing businessperson’s idea of “productive”. Occasionally I have wondered if “craftspeople” are too high on their own supply, but nothing has eclipsed the cocksureness that slop is actually valuable

jkohlmann.bsky.social•8 days ago

The zealotry also reveals the assholes who see their own hobbies as busy work

jkohlmann.bsky.social•8 days ago

Their own hobbies, let alone their own professions* (fixed a typo)

jkohlmann.bsky.social•8 days ago

(Caveat: “busy work” is not to be confused with “actual process”, is not a guarantee by any means, and requires critical reexamination)

randymiller.bsky.social•8 days ago

As a software developer I have no use for LLMs. Integrated LLM tools for things like bug fixing generate plausible sounding nonsense scenarios that send you down the wrong path while speaking with absolute certainty, and generate code that does not address them. I don’t see it getting better.

davidash.bsky.social•7 days ago

Was recently working with a dev who's using AI. Asked him a question because he (presumably) is more knowledgeable than me about some topics. He asked his AI and it gave the wrong answer. I just ended up figuring it out on my own.

randymiller.bsky.social•7 days ago

I will say I’ve seen it used to make barely working prototypes of features that don’t pass any edge cases, but that’s a weird development cycle. I don’t make a shitty hack version of things first, I design them and then fill out their functionality. I’m skeptical it gets you any closer to finishing.

randymiller.bsky.social•7 days ago

If you follow the 80/20 rule, at best it makes the 80% of the project that takes 20% of the time and effort. But now you need to debug code that nobody wrote and nobody understands. It might provide some psychological benefit I guess. Usually when you have a bad prototype you just rewrite it anyway.

vtmicah.bsky.social•8 days ago

I’ve hear people say they use LLMs it to learn topics, like Wikipedia on steroids. I’m deeply conflicted about it.

It will definitely answer your questions / build a report, and 99% of the sentences will be completely correct.

Obviously ML writ large is useful in a million ways.

hrimhari.bsky.social•9 days ago

The main uses I've seen are, essentially, automating tasks that you already know how to do and so can explain well enough in a prompt

If you don't know exactly how to do the task already it will fail (and you can't examine it's results) and so expertise is still required

airelbonia.bsky.social•9 days ago

If it gets a practical use, the hype bubble on that application pops and it drops off the radar. The most practical use I've seen is basically serving as an enhanced OCR to better read document intake when real world conditions add imperfections or vendors change layouts unexpectedly, yet...

Comments

Posting Rules

Reply