The Skynet Prequel / ChatGPT, Bard et al

[dzajic]

This thing is driving me nuts. In every test or attempt to utilize "AI" it is at best mediocre or fails miserably. And yet there are scores of computer science professors and Google and Microsoft seniors up to fellow level calling software development an obsolete job. FFS I couldn't get it to write a simple regex correctly and supposedly its acing all major job interviews. I ask Midjourny for simple illustration and it fails miserably, again and again. "Oh no you see its not just a big statistics map, if it becomes big enough it will really understand stuff and be intelligent". And in practice it doesn't really know that objects aren't really supposed to intersect...

Am I crazy or is everyone out there crazy drunk of the hypejuice (or trying to sell their new AI startup shares to the people drunk on the hype juice)

[Lief Siddhe]

Maybe because it's not actually AI? It's a digital parrot with a big vocabulary.

[Overspark]

Maybe because it's not actually AI? It's a digital parrot with a big vocabulary.

So are many people but yeah.

djazic: I think the people who are successful with it don't necessarily get a good answer on the first try, but after a couple of nudges to refine the answer. It's less impressive than getting it right on the first try but it's still impressive that it manages at all. Also regexes are hard enough for most people as it is, I'm guessing they're harder for a language model to get right than most programming languages.

[Joe Appleby]

It wrote a nice letter to my class‘ parents for me based on some info. The first version was crap, re-rolled and it was good enough that I only needed to edit one sentence and sent it off. It helped me summarizing a lot of stuff to prepare for a few presentations I had to grade. That’s the use I see - do mundane shit that just takes unnecessarily long time to do.

Edit: it helps writing thorough inputs, just like one does for students. Half ass the instructions and you get half assed results.

[epictetus]

Edit: it helps writing thorough inputs, just like one does for students. Half ass the instructions and you get half assed results.

Very much this, there is a bit of an art to giving the right kind of prompts. Friends who really got into midjourney get insane results in short time, whilst I have to refine for ages to get anything remotely resembling what I had in mind. I am sure it is the same with the text based versions (although I personally found it easier to get what I wanted from those).

[Lief Siddhe]

[Hel OWeen]

This site tracks all the AI things going on: https://www.artisana.ai/ so you don't have to hunt down everything for yourself.

A nice mix of news, research and reports.

[dzajic]

Maybe because it's not actually AI? It's a digital parrot with a big vocabulary.

Well some (fully on hype train, or eyeing their stock options, or just very very hopeful) appear to have idea that since neuron weights deep down are unintelligible for humans maybe if model is big enough and "grocks" good enough the net learns to "understand" the problem.

The end goal and grading is "just predict next word" but somehow maybe magically model got large enough and good enough to actually learn English and concepts and how to understand as that leads to being better at "predicting next word".

Trivial example would be model learning addition and instead of a memorizing all possible additions the neruon weight variation eventually leading it to learn how to actually do addition.

Which is somewhat contrasted by (as far as I know) even Gpt4 having issues with addition if it's more than 3 digits, but somehow having theory of mind and ability to "understand".

[tulip]

Maybe because it's not actually AI? It's a digital parrot with a big vocabulary.

Well some (fully on hype train, or eyeing their stock options, or just very very hopeful) appear to have idea that since neuron weights deep down are unintelligible for humans maybe if model is big enough and "grocks" good enough the net learns to "understand" the problem.

The end goal and grading is "just predict next word" but somehow maybe magically model got large enough and good enough to actually learn English and concepts and how to understand as that leads to being better at "predicting next word".

Trivial example would be model learning addition and instead of a memorizing all possible additions the neruon weight variation eventually leading it to learn how to actually do addition.

Which is somewhat contrasted by (as far as I know) even Gpt4 having issues with addition if it's more than 3 digits, but somehow having theory of mind and ability to "understand".

There are neurolinguistic arguments that language is thought, and thought is language. It's exceptionally hard to self conceptualise any thoughts without using internalised language.

Perhaps it's not that the AI's are not reaching some sort of critical capability surprisingly swiftly but that we're over-estimating how much more complex sentience is then just something that talks to itself.

[DerWish]

Before you open you anus please learn how a couple of eager Samsung developers fucked up with ChatGPT.

We were also eager the automate couple of chores and were immeditaly stopped.

[Lowa [NSN]]

Before you open you anus please learn how a couple of eager Samsung developers fucked up with ChatGPT.

We were also eager the automate couple of chores and were immeditaly stopped.

We use a private instance just for this reason. We dont get the latest version but for what we currently use it for in engineering its fine and is a huge help.
Main source code and truly private IP is not allowed, at least not yet. General rule is that if its "open", like public APIs and automation tools we can use the public instance.
Long term we, and I must assume most others, view this as a must-have going forward to attract talent, AI-assisted development will be table stakes.

[epictetus]

[QUOTE=tulip;1971356]

...
Perhaps it's not that the AI's are not reaching some sort of critical capability surprisingly swiftly but that we're over-estimating how much more complex sentience is then just something that talks to itself.

So much this. It will be incredibly interesting to see where the current models reach their limit. Could be we are pretty much there already, could be we got a long way to go still with a bit of tricks and cheating with input from other directions etc.

[VARRAKK]

https://twitter.com/svpino/status/1650832349008125952

[helgur]

Maybe because it's not actually AI? It's a digital parrot with a big vocabulary.

So are many people but yeah.

djazic: I think the people who are successful with it don't necessarily get a good answer on the first try, but after a couple of nudges to refine the answer.

It doesn't get regex right because it is a language model. Regex are too complex and there are too many variations of it to assemble it from predicted text. Try to get it to play chess, it will never get that right either.

People who claim predictive text models like chatgpt will replace jobs like programming are abject morons who don't know what they are talking about. As are people who think they can use this as a tool to aid them in something when they are not qualified to fact check the output of the bot. It gets crucial things wrong all the time, because it has absolutely 0 understanding of how the world works.

[epictetus]

Maybe because it's not actually AI? It's a digital parrot with a big vocabulary.

So are many people but yeah.

djazic: I think the people who are successful with it don't necessarily get a good answer on the first try, but after a couple of nudges to refine the answer.

It doesn't get regex right because it is a language model. Regex are too complex and there are too many variations of it to assemble it from predicted text. Try to get it to play chess, it will never get that right either.

People who claim predictive text models like chatgpt will replace jobs like programming are abject morons who don't know what they are talking about. As are people who think they can use this as a tool to aid them in something when they are not qualified to fact check the output of the bot. It gets crucial things wrong all the time, because it has absolutely 0 understanding of how the world works.

Doesnt need to do the full job, only to speed it up a few % for it to replace a lot of jobs. If enough people can finish a part of the more simplistic coding tasks just a bit faster with new tools, that means a bunch of people will be doing other stuff than they would have (doesnt need to mean they are out of jobs, but they wont be spending their time on what they would have).

[helgur]

Doesnt need to do the full job, only to speed it up a few %

I mean, I can concede to it speeding up repetitive tasks up as a glorified autocomplete, but when I refer to the argument of replacing jobs, I am thinking of the meaning of the argument that programmers as a human profession being obsolete.

[Hel OWeen]

Doesnt need to do the full job, only to speed it up a few %

I mean, I can concede to it speeding up repetitive tasks up as a glorified autocomplete, but when I refer to the argument of replacing jobs, I am thinking of the meaning of the argument that programmers as a human profession being obsolete.

For the foreseeable future: I tend to agree. At risk are all those Mechanical Turk alike jobs, I guess.

As AI advancements threaten the crowd-worker industry, a recent study reveals ChatGPT consistently outperforms human crowd-workers in classification tasks.

And not only in theory, but in pratice already.

A new study reveals that OpenAI's GPT-4 outperforms elite human annotators in labeling tasks, saving a team of researchers over $500,000 and 20,000 hours of labor while raising questions about the future of crowdworking.

Stanford/MIT Study: GPT Boosts Customer Support Agent Productivity by up to 35%.
Boosting Productivity for New and Low-Skill Agents

The researchers observed a 13.8% increase in the number of chats agents successfully resolved per hour. Specifically, the gains the researchers witnessed were:

- A decline in the time it takes an agent to handle an individual chat.
- An increase in the number of chats that an agent is able to handle per hour.
- A small increase in the share of chats that are successfully resolved.

The latter could benefit end users/customers, if customer support/helpdesk quality improves. But the cynic in me believes that these improvements will be turned into profits instead, i.e. fewer agents needed to handle the same amount of requests.

[Spaztick]

Doesnt need to do the full job, only to speed it up a few %

I mean, I can concede to it speeding up repetitive tasks up as a glorified autocomplete, but when I refer to the argument of replacing jobs, I am thinking of the meaning of the argument that programmers as a human profession being obsolete.

For the foreseeable future: I tend to agree. At risk are all those Mechanical Turk alike jobs, I guess.

As AI advancements threaten the crowd-worker industry, a recent study reveals ChatGPT consistently outperforms human crowd-workers in classification tasks.

And not only in theory, but in pratice already.

A new study reveals that OpenAI's GPT-4 outperforms elite human annotators in labeling tasks, saving a team of researchers over $500,000 and 20,000 hours of labor while raising questions about the future of crowdworking.

Stanford/MIT Study: GPT Boosts Customer Support Agent Productivity by up to 35%.
Boosting Productivity for New and Low-Skill Agents

The researchers observed a 13.8% increase in the number of chats agents successfully resolved per hour. Specifically, the gains the researchers witnessed were:

- A decline in the time it takes an agent to handle an individual chat.
- An increase in the number of chats that an agent is able to handle per hour.
- A small increase in the share of chats that are successfully resolved.

The latter could benefit end users/customers, if customer support/helpdesk quality improves. But the cynic in me believes that these improvements will be turned into profits instead, i.e. fewer agents needed to handle the same amount of requests.

There are a few companies already doing this. I know Zendesk for sure uses some kind of language-model AI. If it thinks it can help you by fetching documentation or contacting a certain department it will, otherwise it escalates to a real human. I think of the self-checkouts at a grocery store more than a thinking robot replacing all the admin jobs. That's coming, but not here yet precisely because the output isn't equivalent to a person.

It will be the future for every industry, but the first rollouts are the tedious busywork roles that are the easiest to automate. Programming is still too specialized, but I can see an AI being implemented for analytics and reporting next. Give an AI access to your database and it could generate what you want without prompting it based on past requests.

[Hel OWeen]

This is a ... concerning? interesting? ... statement by basically the founding father of the technology used in current AI systems:

Geoffrey Hinton was an artificial intelligence pioneer. In 2012, Dr. Hinton and two of his graduate students [Ilya Sutskever and Alex Krishevsky, the former now is OpenAI's Chief Scientist, Hel] at the University of Toronto created technology that became the intellectual foundation for the A.I. systems that the tech industry’s biggest companies believe is a key to their future.

[...]

Around the same time, Google, OpenAI and other companies began building neural networks that learned from huge amounts of digital text. Dr. Hinton thought it was a powerful way for machines to understand and generate language, but it was inferior to the way humans handled language.

Then, last year, as Google and OpenAI built systems using much larger amounts of data, his view changed. He still believed the systems were inferior to the human brain in some ways but he thought they were eclipsing human intelligence in others. “Maybe what is going on in these systems,” he said, “is actually a lot better than what is going on in the brain.”

Source: ‘The Godfather of A.I.’ Leaves Google and Warns of Danger Ahead

[VARRAKK]

A US Air Force AI drone simulation ended with the drone going rogue and killing its simulated human operator because the operator wouldn't let the drone kill everything it wanted to

https://www.pcgamer.com/a-us-air-for...-it-wanted-to/