AI technology thread

[rotting bones]

Discuss the technology of AI. I would appreciate it if the discussion of social effects were kept at a minimum because we all know how that would go. Personally, I think the scale at which data centers are being built is a bad thing to say the least. Back to technology:

Latest news on corporate LLMs: https://youtu.be/B3LXEW4Pc-w

Latest news on non-corporate LLMs: https://youtu.be/w9RDunJACkc

[alice]

As someone mentioned somewhere on another website: it's OK as a large repository of information which can be accessed and processed far more quickly and precisely than humans can, e.g. noticing flaws in brain scans. It's not OK as a pretence to, or replacement for, human creativity or empathy (e.g. call centre workers), but unfortunately that €qual$ mor€ $$$ in $om€ p€op£€'$ €¥€$..

[Richard W]

I've a question on AI art, which may be too specialised for this thread. If two people without deliberate collaboration are commissioning detailed scenes for story-telling with the main characters being similar stereotypes in similar cultural environments, should the results look as though they were drawn by the same person? By 'detailed scenes', I am thinking of what might be described as having a lot of set decoration.

I am seeing a lot of similarities in the pictures for stories by nominally different authors, but the names are tied to Youtube channels, so I wouldn't dignify the names as noms de plume.

I can conceive of several mechanisms for what I am observing:

• I am not observing well enough; there are actually differences in style that I'm just not noticing
• Rampant image theft, or mutual derivation. I've seen three similar images for the cover picture for three versions of the same story, where the drawing of the characters looks independent but their positioning is the same. (Theme of at least one is a fat cook on the run, who finds her employer's heart via his stomach.)
• There is a single source of instructions for the AI processes. It is quite conceivable that a common origin is being hidden for financial reasons, such as evading shifting partial prohibitions on the use of AI.
• They're all based on the same processed database
• The same few human-generated images are being used via different databases. That would be plausible for some background details, such as crowded streets.

[zompist]

I am seeing a lot of similarities in the pictures for stories by nominally different authors, but the names are tied to Youtube channels, so I wouldn't dignify the names as noms de plume.

A little of everything, I think. Your experiences remind me of what I got trying out an AI image generator a couple years back. It was not one of the major firms, so it was probably a smaller training set. Its limitations were pretty easy to find.

AIs are trained on human images, so with some work you can reconstruct something close to those images. Someone demonstrated awhile back how easy it was to get trademark violations. The more obscure your prompt, the more you're likely to get into a sparse area of the training data.

The image generator I tried allowed to you input a random seed. This was a feature, since it would change up the results, for good or bad. You generally want a range of images in case your prompt isn't working out. However, you could reverse that if you really wanted a uniformity of style.

I'd also note that if you're still looking at orcs, the source is ultimately human drawings, which are limited in their positioning. E.g. on book covers, or cartoons, or porn, you usually get entire figures centered on the page, or close-ups in predictable ways. There probably aren't a lot of drawings of orcs where e.g. the face is way in a corner and half off the page.

[Richard W]

I'd also note that if you're still looking at orcs, the source is ultimately human drawings, which are limited in their positioning. E.g. on book covers, or cartoons, or porn, you usually get entire figures centered on the page, or close-ups in predictable ways. There probably aren't a lot of drawings of orcs where e.g. the face is way in a corner and half off the page.

I've a feeling one can start with a picture of people and replace some or all of them with orcs. There are also several cases where a cover picture (or thumbnail or whatever you call it) has been designed by replacing alleged Apache women by orc women, but not with as stunning a lack of other changes. My point was that the copying of designs is quite rampant. It's not surprising, as the return on investment may be quite low.

As to getting orcs off-centre, panning and trimming of images has been used quite a bit. There are quite a few videos where there is a single basic picture which is then panned and trimmed, often seemingly just to get round one of Youtube's demonetisation rules. However, sometimes the panning matches the narration.

[malloc]

OpenAI claims to have achieved AGI with their latest model, ChatGPT 6. So much for suggestions that AI would hit a wall in advances before seriously challenging humans.

[linguistcat]

OpenAI claims to have achieved AGI with their latest model, ChatGPT 6. So much for suggestions that AI would hit a wall in advances before seriously challenging humans.

I promise you they haven't, it just sounds good to the shareholders.

[Raphael]

Please think of this paragraph right at the start of the opening post:

Discuss the technology of AI. I would appreciate it if the discussion of social effects were kept at a minimum because we all know how that would go. Personally, I think the scale at which data centers are being built is a bad thing to say the least. Back to technology:

[alice]

OpenAI claims to have achieved AGI with their latest model, ChatGPT 6. So much for suggestions that AI would hit a wall in advances before seriously challenging humans.

I'd like to point out that I can similarly "claim" to have achieved AGI with a five-line¹ Python script, but that would be contrary to the spirit of the thread.

¹standard lines of 80 characters, that is

[malloc]

I'd like to point out that I can similarly "claim" to have achieved AGI with a five-line¹ Python script, but that would be contrary to the spirit of the thread.

¹standard lines of 80 characters, that is

Would enough people find it convincing to produce numerous articles and YouTube videos? I understand the value of skepticism toward mainstream media and corporate press releases, but I also worry that we are letting cynicism blind us to legitimate threats. Sometimes the boy cries "wolf" without good reason, but sometimes there really is a wolf.

[zompist]

Would enough people find it convincing to produce numerous articles and YouTube videos? I understand the value of skepticism toward mainstream media and corporate press releases, but I also worry that we are letting cynicism blind us to legitimate threats. Sometimes the boy cries "wolf" without good reason, but sometimes there really is a wolf.

So Sam Altman issues some propaganda. What has actually changed? Is AI more energy-efficient? Did it suddenly become profitable? (Hint: no.) Did it at a stroke solve all the problems of hallucinations or poorly written code or poorly managed agents? Hint: Also no.

What there is to worry about is exactly what there was to worry about last week: techbros, not AGI.

[malloc]

So Sam Altman issues some propaganda. What has actually changed? Is AI more energy-efficient? Did it suddenly become profitable? (Hint: no.) Did it at a stroke solve all the problems of hallucinations or poorly written code or poorly managed agents? Hint: Also no.

None of those have any bearing on whether AI has matched or exceeded humans in intelligence. Humans hallucinate and make mistakes all the time so the fact that AI makes similar errors does not particularly undermine its purported intelligence. Likewise not all cognitive abilities lend themselves to profit. One can excel at theoretical physics, or language construction for that matter, without ever striking it rich.

[zompist]

So Sam Altman issues some propaganda. What has actually changed? Is AI more energy-efficient? Did it suddenly become profitable? (Hint: no.) Did it at a stroke solve all the problems of hallucinations or poorly written code or poorly managed agents? Hint: Also no.

None of those have any bearing on whether AI has matched or exceeded humans in intelligence. Humans hallucinate and make mistakes all the time so the fact that AI makes similar errors does not particularly undermine its purported intelligence. Likewise not all cognitive abilities lend themselves to profit. One can excel at theoretical physics, or language construction for that matter, without ever striking it rich.

In other words you have no evidence that AI has matched human intelligence and you can't think of a single thing that has changed.

[malloc]

In other words you have no evidence that AI has matched human intelligence and you can't think of a single thing that has changed.

If you're really interested, I could point to numerous examples of AI mastering cognitive skills once considered exclusive to humans. Not quite the same thing as AI matching human intelligence but alarmingly close and getting closer everyday. Quite honestly, I struggle to name any cognitive skills that remain exclusive to humans though presumably there still are some.

[Travis B.]

In other words you have no evidence that AI has matched human intelligence and you can't think of a single thing that has changed.

If you're really interested, I could point to numerous examples of AI mastering cognitive skills once considered exclusive to humans. Not quite the same thing as AI matching human intelligence but alarmingly close and getting closer everyday. Quite honestly, I struggle to name any cognitive skills that remain exclusive to humans though presumably there still are some.

You're ignoring the fact that these AI's have to be specialized in what they do, i.e. they are not Artificial General Intelligence. Your chess-playing AI cannot simultaneously, without modification, fold proteins.

[linguistcat]

In other words you have no evidence that AI has matched human intelligence and you can't think of a single thing that has changed.

If you're really interested, I could point to numerous examples of AI mastering cognitive skills once considered exclusive to humans. Not quite the same thing as AI matching human intelligence but alarmingly close and getting closer everyday. Quite honestly, I struggle to name any cognitive skills that remain exclusive to humans though presumably there still are some.

You're ignoring the fact that these AI's have to be specialized in what they do, i.e. they are not Artificial General Intelligence. Your chess-playing AI cannot simultaneously, without modification, fold proteins.

Or create images, navigate a maze, identify baked goods or cancers, create conlangs, do basic math correctly, etc. And even if we have individual AI for each of these (or calculators in the last case), none of them do the others, or do other things well. LLMs are bad at math, AI for identification cannot write.

[alice]

Note for those of a malloc persuasion: this post says some things about AI which are not wholly negative. If this is likely to upset you, please skip ahead to the next post.

I've spent much of the past few days looking at AI-generated videos on a popular video-hosting website. Some obversations:

- In general, the visuals are good enough to survive a cursory watch; in particular, some AI-generated steampunk is seriously impressive. There is probably a place for it here where things in the imagination simply can't be built in real life.
- It's rather less convincing when you start looking at the details. Notable examples are lettering, people eating things, and gears rotating. And continuity is another good source of inconsistencies.
- I suspect there are only a few programs which are used to generate the videos; certain faces keep popping up over and over again, for example.
- It's unnerving when characters' accents repeatedly switch between near-RP and GenAm.
- Some of the sillier videos - Star Wars meets Star Trek, gender-swapped LotR, and popular films with 1980s'/1970's/1950's/1940's casting (e.g. Harry Potter as film noir) - are actually good fun, and showcase what AI is best at - making things out of existing things without needing to invent anything new..
- When all is said and done, there's nothing AI humans can be or do which non-AI humans can't be or do more convincingly, however realistic they may look..

In summary: for broad sweeping scenes which don't need much action or detail, it's often at least passable and sometimes much more so. It would be interesting to compare the amount of effort (and energy!) which goes into such things with the amount which would be needed to do something comparable without AI. But for anything which needs to be convincing to even the mildest of nitpickers, a lot of cleaning up is likely to be necessary. If all you want is the equivalent of a thin slice of beef on white bread, it'l probably do, but if you're after anything more, you'll be left unsatisfied.

Finally: honourable mention goes to the beginning of this, which has clearly had a lot of work post-production: not only are the names all legible, but they're also consistent with the following scene, which is also unhurried and very well done. I very much doubt AI will ever be able to do this without further substantial human intervention.

[Richard W]

- It's rather less convincing when you start looking at the details. Notable examples are lettering, people eating things, and gears rotating. And continuity is another good source of inconsistencies.
- I suspect there are only a few programs which are used to generate the videos; certain faces keep popping up over and over again, for example.
- It's unnerving when characters' accents repeatedly switch between near-RP and GenAm.

I've recently seen advertising for a product, Pippit, that seems to claim to be able to do something resembling animating a story (perhaps just a series of scenes). One thing I've heard is that AI is lacking in pushback - essentially, GIGO lives. Humans have some resistance to doing GIGO.

One thing that LLMs are very bad at is a arithmetic. But perhaps this is just a matter of it being found worthwhile to bolt on the capability. This may relate to its lack of consistency checking.

[alice]

As an addendum to my previous ramblings, my beloved opined that (paraphrased) an AI-generated film is basically a glorified cartoon.

[malloc]

Or create images, navigate a maze, identify baked goods or cancers, create conlangs, do basic math correctly, etc. And even if we have individual AI for each of these (or calculators in the last case), none of them do the others, or do other things well. LLMs are bad at math, AI for identification cannot write.

Sure but AI in the aggregate can do all of those things, which is enough to challenge humans in the workplace and administration. Meanwhile the technology just keeps advancing with no signs of slowing down.