AIs gunning for our precious freelancers

[keenir]

okay, try this: what is the difference between a handful of nerve cells grown in a petri dish, from a human brain in a skull? if you have lots of petri dishes full of nerve cells, do you suddenly have a human level intelligence?

Sure but neural nets are not simply random clusters of artificial neurons strewn around in small containers.

I never said "neural nets". I said "nerve cells".

though if neural nets aren't random clusters in small containers...but a human level intelligence can be formed from lots of random clusters and not neural nets...

then why bother with making fake neural nets?

There are well-developed techniques and principles for connecting the neurons in useful ways. Currently we have not yet figured out how to connect them just like the human brain,

and hte difference is...?

wait, are you saying the brain isn't put together usefully? hm, that seems like a problem - we won't reach human level AI if it needs to be put together sloppily.

but we have figured out how to create neural nets that emulate numerous cognitive tasks once reserved for humans.

really? such as...?

Even without replicating the connectome of the human brain, neural nets are remarkably effective.

the fact that there are these hundreds of corporations, each of which has thousands of scientists and engineers hard at work on developing AI/AI Models...and that we still are light-years from a real AI, that should tell you something.

What makes you so sure that we are light-years away from true AI, though? Several years ago everyone assumed that generative AI capable of drawing pictures and writing novels was quite a long way off.

and its still a ways away still.

Obviously that doesn't mean that AGI will happen tomorrow, but it seems premature to declare it light-years away. Considering the astonishing progress in AI over the past few years, I feel we have already traversed considerable parsecs after all.

really? and what do you consider to be "astonishing"?

Also who the hell are all these guests on the site and are they the reason it keeps crashing?

All these guests are the AIs that have you so so delerious.

[zompist]

What makes you so sure that we are light-years away from true AI, though? Several years ago everyone assumed that generative AI capable of drawing pictures and writing novels was quite a long way off.

Not really. You do realize that people have been wanting and predicting robots, and worrying about them taking over, for literally over one hundred years?

Now, though you are not too clear on the distinction, SF is not technology. So we can ask how long people have been working on AI. Answer: 75 years at the very least. And at every moment in those 75 years, AI researchers thought AI was just a few years way.

How long do you think people have been working on artificial neural networks? 65 years.

[malloc]

Now, though you are not too clear on the distinction, SF is not technology. So we can ask how long people have been working on AI. Answer: 75 years at the very least. And at every moment in those 75 years, AI researchers thought AI was just a few years way.

How long do you think people have been working on artificial neural networks? 65 years.

Sure but this time they have made some genuine breakthroughs with AI capable of producing passable imitations of art and literature and taking over white collar work. Even if the technology stalls for the foreseeable future, it has put everyone from artists to office workers on the short list for layoffs. Meanwhile it seems unclear to me why a technology with such incredible momentum and funding behind it would suddenly hit a wall.

[keenir]

Now, though you are not too clear on the distinction, SF is not technology. So we can ask how long people have been working on AI. Answer: 75 years at the very least. And at every moment in those 75 years, AI researchers thought AI was just a few years way.

How long do you think people have been working on artificial neural networks? 65 years.

Sure but this time they have made some genuine breakthroughs with AI capable of producing passable imitations of art and literature and taking over white collar work.

just so we're on the same page...what do you consider to be such white collar work?

Even if the technology stalls for the foreseeable future, it has put everyone from artists to office workers on the short list for layoffs.

the invention of B-Movies did not eliminate the need or creation of award-winning movies. films like Atlantis Rim did not make there be no market for a sequel (film or tv series) to Pacific Rim.

so why would current "AI"s create a shortlist for anyone? because you're bugeyed at them folding proteins?

Meanwhile it seems unclear to me why a technology with such incredible momentum and funding behind it would suddenly hit a wall.

you mean like the race to colonize the Moon did? or for Mars?

[Torco]

I think we shouldn't dismiss malloc's point so much. he's wrong on many things, but he's right that generative models *are* actually becoming useful in a way that they weren't 15, 30 or 45 years ago. this isn't "just some fancy autocorrect"... okay, it is, technically, but these autocorrects are absolutely in a position to "replace" an increasingly number of workers. [not exactly replace, but close enough]

[keenir]

this isn't "just some fancy autocorrect"... okay, it is, technically, but these autocorrects are absolutely in a position to "replace" an increasingly number of workers. [not exactly replace, but close enough]

Hmm...I didn't think of it like that; my apologies, Malloc and Torco. As someone who spent years as a proof-reader and beta-reader, it should not have slipped my mind about the workforce being made redundant by autocorrect...and yet, sadly, I did.

[WeepingElf]

this isn't "just some fancy autocorrect"... okay, it is, technically, but these autocorrects are absolutely in a position to "replace" an increasingly number of workers. [not exactly replace, but close enough]

Hmm...I didn't think of it like that; my apologies, Malloc and Torco. As someone who spent years as a proof-reader and beta-reader, it should not have slipped my mind about the workforce being made redundant by autocorrect...and yet, sadly, I did.

Autocorrect often gets things dead wrong, especially if technical or otherwise rare vocabulary is involved, and therefore doesn't render proofreaders redundant - because it has no idea what the text means. It doesn't know that "Queen" and "Sensation" aren't names of languages, and "Quenya" and "Sindarin" are correctly spelled language names, for instance. And LLMs may be better, but only gradually, and they still don't know what the texts mean, and often produce nonsense. On a tangent, I know a guy whose job is proofreading machine-translated texts, and he finds lots of mistakes in the texts he proofreads.

[zompist]

I think we shouldn't dismiss malloc's point so much. he's wrong on many things, but he's right that generative models *are* actually becoming useful in a way that they weren't 15, 30 or 45 years ago. this isn't "just some fancy autocorrect"... okay, it is, technically, but these autocorrects are absolutely in a position to "replace" an increasingly number of workers. [not exactly replace, but close enough]

If your point is "LLMs can be used by evil CEOs to replace human jobs but do an even shittier job", I agree and have all along. Malloc's problem is grossly overestimating how close we are to "real AI", and assigning the evil agency to the AIs rather than the CEOs.

[Raphael]

I don't think that proofreaders are obsolete just because now people can always use they're autocorrect.

[TomHChappell]

I don't think that proofreaders are obsolete just because now people can always use they're autocorrect.

I think you mean “there autocorrect”. Amirite?

[zompist]

generative models *are* actually becoming useful in a way that they weren't 15, 30 or 45 years ago. [not exactly replace, but close enough

Chat programs are a huge advance, sure. But the tempo is important to get right, since malloc is convinced human-level AI is coming next week.

As I said, the basic idea— simulating neural nets— goes back to 1960. The first mass-market use of generative AI was for translation, e.g. Babelfish, introduced in 1997— 28 years ago. The basic idea behind LLMs goes back to the 1980s. There have been organizational advances along the way, but you get a lot of the advances simply with size— trillions rather than millions of input tokens, trillions rather than hundreds of nodes— and that piggybacks over constant improvements in computer speed and storage capacity.

It's just not true that there was nothing in 2021, and then ChatGPT appeared in a nimbus of light. But that of course is the story that the techbros want to tell to investors and which malloc swallows as well.

[malloc]

Chat programs are a huge advance, sure. But the tempo is important to get right, since malloc is convinced human-level AI is coming next week.

Not really. It does seem quite plausible that they will develop in the next decade, though, given the incredible pace of development over the past few years. It might turn out that the last few cognitive tasks separating humans from AI are fantastically difficult of course, but the opposite seems equally likely.

It's just not true that there was nothing in 2021, and then ChatGPT appeared in a nimbus of light. But that of course is the story that the techbros want to tell to investors and which malloc swallows as well.

The way I remember it, generative AI went from practically non-existent to drawing passable imitations of artwork in the span of several months in mid-2022. Perhaps software engineers saw something different but for anyone outside the field, the emergence of generative AI really did look unprecedented and sudden.

[keenir]

Chat programs are a huge advance, sure. But the tempo is important to get right, since malloc is convinced human-level AI is coming next week.

Not really. It does seem quite plausible that they will develop in the next decade, though, given the incredible pace of development over the past few years

Malloc, speedrunning is an option in games. speedrunning is not an advisable action in conversations. you keep speedrunning and missing all the points, and yet you wonder why nobody picks up what you're putting down.

It might turn out that the last few cognitive tasks separating humans from AI are fantastically difficult of course, but the opposite seems equally likely.

and that is based upon...?

It's just not true that there was nothing in 2021, and then ChatGPT appeared in a nimbus of light. But that of course is the story that the techbros want to tell to investors and which malloc swallows as well.

The way I remember it, generative AI went from practically non-existent to drawing passable imitations of artwork in the span of several months in mid-2022. Perhaps software engineers saw something different but for anyone outside the field, the emergence of generative AI really did look unprecedented and sudden.

to quote Ambassador Londo Mollari, that is because "I can only conclude that you have not been paying attention!"

[keenir]

I don't think that proofreaders are obsolete just because now people can always use they're autocorrect.

I think you mean “there autocorrect”. Amirite?

Let me see how well I remember my old hobby...

I don't think that beta-readers are obsolete just because now people can always use one's autocorrect.

I think you mean “the autocorrect”. Amorite?

(sorry)

[zompist]

It's just not true that there was nothing in 2021, and then ChatGPT appeared in a nimbus of light. But that of course is the story that the techbros want to tell to investors and which malloc swallows as well.

The way I remember it, generative AI went from practically non-existent to drawing passable imitations of artwork in the span of several months in mid-2022. Perhaps software engineers saw something different but for anyone outside the field, the emergence of generative AI really did look unprecedented and sudden.

And I'm telling you that you remember wrong. The gains of ChatGPT were prepared over the previous forty years of development. It felt like a big advance, but it's the same basic technology as Google Translate, which is almost 20 years old.

On image generation proper, again, your memory is simply wrong. AIs like DeepDream were generating images 10 years ago. Not great ones, but that's how progress is made: incrementally, not in miraculous leaps.

[Torco]

Autocorrect often gets things dead wrong, especially if technical or otherwise rare vocabulary is involved, and therefore doesn't render proofreaders redundant - because it has no idea what the text means. It doesn't know that "Queen" and "Sensation" aren't names of languages, and "Quenya" and "Sindarin" are correctly spelled language names, for instance. And LLMs may be better, but only gradually, and they still don't know what the texts mean, and often produce nonsense. On a tangent, I know a guy whose job is proofreading machine-translated texts, and he finds lots of mistakes in the texts he proofreads.

for sure. i don't think there are many jobs that generative models can fully replace, but you don't need to replace a full job for a person's job to be replaced. and it's not like people don't screw up from time to time as well.

If your point is "LLMs can be used by evil CEOs to replace human jobs but do an even shittier job", I agree and have all along. Malloc's problem is grossly overestimating how close we are to "real AI", and assigning the evil agency to the AIs rather than the CEOs.

oh, totally, and I think my position is basically yours as well. I'm merely saying that some of the corrections to his obviously mistaken position were a bit too far. like, there's really been no significant progress towards the goal of humans on mars, but there has been significant progress towards automating relatively complex cognitive tasks to a machine, notably carrying out instructions given by humans

which is, from an employer's point of view, is all a worker is: an thing that if you tell it to do X, does X. bosses often use their workers exactly like most people at work are now using LLMs: say "write me a proposal for that thing they're asking in that email (even though i'm gonna read it, and probably rewrite half of it, cause it's me who really knows how to do this but i'm busy / rather do with this other thing) ". increasingly, they're building agents, "self-prompting" ensembles of models, and that opens the door for more tasks to be automated that were thought to be inherently in the human domain, the AGI doomsayers are not wrong about that bit. of course, all technologies plateau: the reasoning of the AGI doomsayer [the steelmanned version of malloc's postiion here] reminds me of that comic ME AM PLAY GODS. the mistake is reasoning along the lines of "fewer back then, more now, eventually all". if we learn to make fire, we'll burn the entire cosmos. if computers become able to do some things, they'll be able to do all the things. but i mean, i don't think it has plateaued (?) yet, with NLP. then again

generative models *are* actually becoming useful in a way that they weren't 15, 30 or 45 years ago. [not exactly replace, but close enough

Chat programs are a huge advance, sure. But the tempo is important to get right, since malloc is convinced human-level AI is coming next week.(things which all are true)

absolutely yes. i'm perhaps even more radical: malloc thinks human-level AI is a possible thing: conceivably, it'll either be better-at-some-things-than-people-and-worse-at-others, or better-at-all-things. inbetween, tho, there's a lot of sets of things it can be better than us at which are honestly scary, so being scared of a nebulous "AGI" [what does that even mean? it's immensely unlikely that it'll be just-as-good-as-people-at-all-things] is a regrettably ignorant version of a genuinely reasonable concern, not an altogether sillyness.

lmao malloc, honestly, learn about these things if they concern you that much. dall-e was 2020. we used to have a saying here, RAFBA, but it was a bit crass so i'll say "please read up on this subject before so vehemently discussing it". PRUOTSBSVDI just doesn't have that good of a ring, alas. maybe prutos ?

[Raphael]

I don't think that proofreaders are obsolete just because now people can always use they're autocorrect.

I think you mean “there autocorrect”. Amirite?

Seriously, I recently used "their" instead of one of the two others when I was writing something. And I was so embarrassed, even though just to myself, when I noticed.

[Raphael]

Belated additional comment:

Autocorrect often gets things dead wrong, especially if technical or otherwise rare vocabulary is involved, and therefore doesn't render proofreaders redundant - because it has no idea what the text means. It doesn't know that "Queen" and "Sensation" aren't names of languages, and "Quenya" and "Sindarin" are correctly spelled language names, for instance.

One of my doctors has the surname "Dee". It took me a while to convince my cellphone that I was not making notes about appointments with Dr. Dre.

[Travis B.]

At my work they had us enable GitHub Copilot in VS Code... and I have to say that it's 'autocompletes' are not just merely useless but rather actively hinder my work, because they distract me from my flow with my having to constantly cancel them (or my accidentally accepting them and then having to fix them after the fact). This is the generative AI that allegedly will obsolete programmers like myself ─ it cannot even simply 'autocomplete', much the less truly write original code that is useful.

[rotting bones]

At my work they had us enable GitHub Copilot in VS Code... and I have to say that it's 'autocompletes' are not just merely useless but rather actively hinder my work, because they distract me from my flow with my having to constantly cancel them (or my accidentally accepting them and then having to fix them after the fact). This is the generative AI that allegedly will obsolete programmers like myself ─ it cannot even simply 'autocomplete', much the less truly write original code that is useful.

Copilot has been seriously annoying. I don't understand why my cousin thinks it's better than what he does. Maybe it does better if the code contains extensive comments? Or maybe it does better if the code isn't low level?

LLMs can only generate short scripts of at most 10-20 lines from English descriptions. Anything other than that is crap. They're unreliable at inferring function from code. My understanding is that they do outperform the previous state of the art in autocomplete, the Programming by Example used in Excel's autocomplete.

I've heard the Cursor IDE is useful for solving build problems. It comes with an LLM that can read the files in your project directory and help you trace execution patterns. I don't know how well it works. Maybe I should try it. Unless it's way better than what I've seen so far, I wouldn't use LLMs for any task that doesn't involve summarizing or expanding text. I only use classifiers in my research and avoid generative models whenever possible.