I’m an industry dev turned comp sci prof. I’ve been writing code for 40-some years; my first commercial software was for the Apple ][+. I’ve worked on web, mobile, desktop, server, and data plumbing projects for companies ranging from startups to Fortune 100 to arts nonprofits. In short, I’ve been lucky to know a lot of people in a lot of corners of tech.
@anildash’s opening paragraphs are spot on. If consensus is not 100%, well, it’s 99%.
There •is• definitely debate about LLMs among folks who have both tech expertise and a level head, and that debate is mostly between:
(1) “If only we could have reasonable conversations about what if anything this tech is actually good for! It could be useful! Curse this hype! Curse this bubble!”
and
(2) “This technology is so toxic we shouldn’t even be considering •any• uses for it”
That’s a •continuum•, and folks (including me) fall in all kinds of complicated places along it. Just note the continuum’s endpoints.
Because this is Mastodon, I know lots of people will take the previous post as an invitation to put their opinion in my replies. It was not in fact such an invitation, but…it’s cool. C’est la vie Mastodonique!
What’s really important, per Dash’s OP, is for those of us in tech to get the actual debate among experts into view of the broader public — i.e. not just here — and not let charlatan billionaires dominate the conversation.
@inthehands Perhaps the language of stage magic and mentalist cold readings is appropriate for distinguishing the play to the audience from the content and rhythm of the interactions, and the use of LLMs as instruments.
I have trouble seeing beyond the terrible synergy of the financial bubble and the motivated reasoning of the boosters to the genuine positive applications and the real costs and limits.
Working with conversational support and workflow agents before the bubble, I could see that there were really good ways to use them, "everybody wins" ways, but a temptation to cheap out on creating the right interaction context for that case.
@inthehands I'd rather talk about the terrible pattern of financial hyperbole cycles and how to break the pattern than the nature of the current darling technology.
I started my software career on the down side of the dot-com boom and this is hard to distinguish from the past hype cycles. Another Force Ten cyclonic finance weather event.
Driven by global financial climate change.
@inthehands @anildash
OK. I'll bite. What — if anything — is this tech actually good for?
(eta: Not simply trawling, here. But I've yet to hear a coherent answer to "So what are the legitimate applications of LLMs?")
@mikro2nd @inthehands @anildash
IMHO, as pattern synthesis machines, any use case for identifying patterns, predicting, and then prescribing could be tested for effectiveness.
@mikro2nd @inthehands I’m personally anti-LLM, but in a vacuum I think https://goblin.tools/ToDo is a good (tiny) example of them working well:
* nobody cares about copyright infringement for a todo list
* pretty decent adhd accessibility aid
* if it generates something wrong the human will just be like “huh that’s a stupid thing to have on my todo list”
of course, this doesn’t exist in a vacuum, so the societal cost isn’t worth it, but it demonstrates “can do at least one thing” imo
@mikro2nd
LLMs are good for any task that has a lot of fiddly bits but is easily recognized as correct once implemented.
An example is the generation of code to make detailed precise plots of data. Theres lots of fiddly bits to adjust the axes and the coloring by groups and titles and transformations of the data and etc... It frequently requires wading deep into the documentation... But if an LLM generated code you can easily check visually if it worked.
@inthehands @anildash
@mikro2nd Power company profits. (Not entirely a joke; I'm not very convinced by the theory that fossil fuel interests are actively pushing "AI", but I daresay they're delighted by it).
More seriously, there are alas fields of human endeavour where plausible sounding bullshit is useful. It's just that they tend to be harmful too; "SEO" sleaze, endless Webpages full of slop and affiliate links, cranking out junk papers to send to paper mill journals, trash ebooks, plagiarised art, writing consultancy which gives the answer the client wanted so they won't check the sources anyway...
Another way this is reminiscent of the cryptocurrency bubble, of course - sure, there are uses for it, it's just most of them are criminal, like ransomware.
@inthehands @anildash “If only we could have reasonable conversations” is tough these days.
A friend of mine who works in Philips medical devices described the immense pressure they’re getting from top executives to find ways to apply AI to their products. It’s a horrifying mismatch, doomed to fail or (I fear) worse, and sounds driven entirely by hype/FOMO, conflation of different AI kinds, etc.
No reasonable conversation can be had here without (hopefully benign) failure and cultural change.
@inthehands @anildash …someone just pissed of our future robot overlords.
@inthehands I’m in a smaller organization with a well funded IT department (a rare enough thing) and these are the conversations we’ve had. We’ve definitely found things that LLMs have some utility for but none of it is at the scale you’d think from the hype. We’re definitely not going to be blowing the budget on it.
Manutoky
Paul Cantrell
Jeff Miller (orange hatband)
חנן כהן • Hanan Cohen
mike morris
Coach Pāṇini ®
Enola Knezevic
David Smith
Daniel Lakeland
Anil Dash
Major Denis Bloodnok
Dave Peck
Dawn Ahukanna
Giacomo Tesio
panu
satire
Irenes (many)
kirakira
Paul Turnbull has moved.🇨🇦