Mastodawn

Scott Small 🇨🇦Jul 11

Thomas 🔭🕹️

So the main "arguments" when I say "AI doesn't work" and "it will collapse" are:

1. "You don't know what you're talking about"
2. "It's inevitable and here to stay, might as well go with the program"
3. "But it's almost there! Just last week they released [name of model] that is so close!"

Literally no one ever replies with any concrete examples with how it reliably, ethically and non-wastefully works for them to increase their productivity and improve their and other people's lives in any meaningful way.

It's always ad hominems, hypotheticals or deeply flawed "it sort of works for this".

inamruzui Jul 9

@thomasfuchs
1. Denial
2. Pre-emtive complience
3. Shiny new thing override thinking

eswillwalker Jul 9

@thomasfuchs There are plenty of examples of AI models working in limited circumstances. For example, amassing the texts of large numbers of cuneiform tablets to identify a hymn. AI models are just another form of statistical methods used for finding patterns. We use standard statistical models in science all the time. I imagine that the AI style of models will be incredibly useful in those types of contexts. What they cannot do is “think” or reason and there seems to be no indication that they will be able to do so any time soon. However, that is no reason to throw AI models out the window.

Jonathan Polley Jul 9

@ELS @thomasfuchs These examples of the pattern matching version of “AI.” They are not the LLM version that are being sold as the solution to all of our problems and will be ready “any day now.”

Thomas 🔭🕹️Jul 9

@the_other_jon @ELS

Correct, if “AI” is used unqualified in a piece of general writing (i.e. not a scientific paper) it refers to a group of technologies that are also called “generative AI”.

These include—but are not limited to—image generators and LLMs (large language models), which generally transform tokenized input data with linear algebra, steered by a statistical model and (mostly human) labeling of a corpus of training data (text, images, video, music, etc.)

requiem 🦫Jul 9

@thomasfuchs sounds like religion.

@Zitron nailed this one on his blog. Journalists aren't asking for proof when they cover these products.

https://indieweb.social/@OptOut/114700540530860265

Casey (@OptOut@indieweb.social)

"I don’t mean “show me a company that did layoffs and claims they’re bringing in new efficiencies with AI.” I mean actually show me a company that has laid off, say, 10 people, and how those people have been replaced by AI. What does the AI do? How does it work? How do you quantify the work it’s replaced? How does it compare in quality? Surely with all these headlines there’s got to be one company that can show you, right?" ------------------------ https://www.wheresyoured.at/sic/

Indieweb.Social

Philip Mallegol-Hansen Jul 9

@thomasfuchs It’s tragic-comic too how the only “argument” for “It’s inevitable” is “The people who have billions of dollars to lose if it isn’t inevitable, say it must be inevitable”.

Leen Kievit Jul 9

@philip @thomasfuchs As inevitable as blockchain

Schroedinger Jul 9

@thomasfuchs 1. I do know what I am talking about.
2. We will spend decades removing it and its pernicious influence from the web.
3. It is still no better really than ELIZA was - because it is the same essential model. Just 1000 time bigger. We will never get "there".

LLM is a dangerous and stupid fad.

txerren 🇵🇸Jul 9

@thomasfuchs yes, that is also the whole of my experience debating the issue (especially with engineers)

@thomasfuchs "reliably, ethically and non-wastefully" - that's a high standard!

@thomasfuchs my plant recognition app has been a game changer. It's not 100%, but it's like 85% maybe, and that's a damn sight better than I would do alone or with a book.

I can't speak to the ethically or non wastefully bit.

Stuart &c.Jul 9

@thomasfuchs @HollyGoDarkly You may classify all AI use wasteful, not include LLMs, or I may misunderstand your point, but I’m finding increasing utility. For example, I constructed some (for me) very complicated spreadsheet formulas to extract useful data from a larger set using Claude. Could have done it myself but would have taken hours &, instead, did it in 10 mins. Also regularly saving hours using on-device summary & key-points features on-device & powered by renewable energy.

Luke Kanies Jul 9

@thomasfuchs I definitely don’t have an answer for the ethics side, and I share your concerns.

In terms of utility, though, I am finding code generation very useful as I try to write my first iPhone app. I didn’t know swift, SwiftUI, or iOS (as a dev platform) when I started, and it’s definitely saved me tens or hundreds of hours of searching forums and docs to find the right framework, call, and arguments. It is regularly more correct than Xcode on how an api works, too.

There are tons of downsides. And I’ve no idea if it will be anywhere near as useful once I am more experienced. But as a newbie it has honestly been a lifesaver.

(The main downsides are that it is a shitty architect, a mediocre programmer, and it’s too easy to hand it stuff that I should be owning.)

Thomas 🔭🕹️Jul 10

@lkanies if you rely on it to fix your problems, you will stay a newbie because it won’t help you become more experienced—the work of combing through code to figure out problems yourself is how your brain becomes a not newbie

Luke Kanies Jul 10

@thomasfuchs yep, I am aware of that problem.

I am not at all saying it is without flaw, or even a net positive.

You said no one had provided any examples of any utility, so I provided an example of utility.

I think programming is uniquely a good fit for LLMs because it’s such a more constrained space than, say, writing.

But even there there are tons of short and long term issues.

But “tons of big issues” != “no utility”, even if the net result is we should not use it.

Chris Martin Jul 11

@thomasfuchs I use Cursor in a very limited way to speed up my typing. I only accept its suggestions when they're exactly what I would written myself. It's not game-changing in any way, and I don't think it's worth the price, but it is an accelerator for me.

Chris Martin Jul 11

I guess "ethical" is always up for debate, but for what it's worth "it's copying stolen code from the internet" certainly doesn't look like what's going on; mostly the autocompletions seem to be based on following existing patterns in our own repository or guessing that I want to repeat something similar to the last few things I did, there's not much outside novelty showing up.