We can't all use AI. Someone has to generate the training data

https://twitter.com/paulg/status/1635672262903750662

by redbell • 3 years ago

341 599 3 years ago

twitter.com

jgrahamc4• 3 years ago

This is kind of why I created https://lowbackgroundsteel.ai.

Mordisquitos3• 3 years ago

I have to say, I love the analogy you used for the name.

gavinhoward2• 3 years ago

Good idea.

I'd ask for my stuff to be added to that, but companies would probably use your list to find such content and steal it.

consumer4512• 3 years ago

Cool! Just curious (for bragging rights) if the name was influenced by my comment here, or just coincidence?

https://news.ycombinator.com/item?id=34960377

jgrahamc4• 3 years ago

It was a coincidence. Sorry. I thought of this a few weeks ago and only just got round to making the site. Great minds etc.

ggm6• 3 years ago

A reminder that as long as it demands training and reinforcement it's almost certainly low on induction and production of new things.

Very artificial. Not very intelligent.

s17n• 3 years ago

Humans need training and reinforcement.

ggm6• 3 years ago

Yes, undeniably true. But what they acquire is inductive reasoning skills, and the production of new things.

catach2• 3 years ago

Do you have examples of human-created "new things" that aren't essentially novel combinations of old things? Because I come up blank. And this current crop of AI generators are very good at combining old things in novel ways.

I do agree with your general point that these generators aren't really "intelligent", however. Will have to ponder if I agree about the induction bit.

ggm6• 3 years ago

RSA, and the GCHQ PhD equivalent from the 1970s were really remarkably new. Crypto systems before then were symmetrical. Inventing a form of encryption which was a-symmetrical was new.

One time cipher streams were new.

the invention of packet switched networks (Louis Pouzin, Len Kleinrock) was new. It wasn't inherent in prior methods, it's an inductive consequence of time division multiplexing but with addressing and routing.

There is no good analogue in nature to either the internal combustion engine, or the steam engine: the conversion of linear force to rotary force and vice-versa was a really novel thing. I would argue the wankel engine as a diversion from pistons was pretty good reasoning.

But in the same way Kurt Vonnegut says there is a small fixed number of plot models for a novel, almost all late-stage human endevour is derivitive. It's in the nature of the beast. To claim GPT is therefore 'meeting the mark' because the burden of human existence has less discovery and more inductive reasoning simply comes back to my first point: where's the evidence of the GPT doing inductive reasoning with discrimination, beyond the syllogistic?

aidenn0• 3 years ago

...Tell that to AlphaZero?

dogman144• 3 years ago

Live validation Jaron Lanier’s siren servers. All this magic is built on the back of free labor, from captchas to duolingo.

rebelde• 3 years ago

Why write your thoughts on the web when AI/GPT is only going to steal and paraphrase it? Nobody sees what you write and everybody thinks GPT is the genius.

cableshaft• 3 years ago

Just saw something today where the wife of TotalBiscuit, who died of cancer several years ago, is contemplating deleting all of his Youtube videos[1] to prevent people from using A.I. to make him say terrible things.

Did give me a bit of a pause about putting stuff out there. Although I think I'd still rather have my data be used for training A.I. than not (and I probably am already in the training data anyway, I believe I saw that one of the datasets it's been trained on was Hacker News comments).

[1] https://kotaku.com/totalbiscuit-john-bain-youtube-delete-vid...

Dalewyn11• 3 years ago

Given that the "AI" community apparently couldn't care less about treating intellectual property rights with wanton abandon, I can't say such a response would be unwarranted.

Dire circumstances call for drastic measures, as they say.

jeroenhd2• 3 years ago

Quite a sad, but completely understandable reaction. The saddest part is probably that it's already too late to prevent people from generating TB deepfakes and other content. Cloning a voice takes half an hour if clips now, any downloaded live stream should be enough already.

It's sad to see AI on a path to destroy years of collected internet content. I expect the internet archive to receive loads of takedown requests in the coming months and years because of this.

pklausler2• 3 years ago

The general problem of "AI"s being trained on copyrighted content needs to be discussed more thoroughly, I think.

noogle• 3 years ago

The current (legal) answer is "unclear". There are indications that training is fine, but producing and using the generated content is questionable at least. As many IP issues, it will solved only when someone will try that in court and go all the way until a verdict. Some cases are actually being processed but it might take years to get an answer.

bluefirebrand• 3 years ago

Every time I bring this up, people accuse me of resisting progress, "the cats out of the bag", etc.

It has been frustrating.

antibasilisk3• 3 years ago

Copyright's been dead since the internet was born. I really do think it's the least of our problems when it comes to abstract reasoning engines.

anonzzzies5• 3 years ago

I think it’s not very hard; if the AI companies believe the data they trained on is public domain/open because they scraped it of the internet, then their trained weights must publicly available as well. They cannot claim ‘but training is expensive’; if they do, then they should pay fees for the hosting and storage and writing time of all data they scraped. I prefer open weights as it’s more practical. Your weights have a sliver of GPL source in it? Well that infected the entire thing as GPL does: it is ours now too!

SketchySeaBeast2• 3 years ago

That's why I keep my content as low quality as possible - keeps the machines humble.

ThrowawayTestr• 3 years ago

I'll just run it though an AI upscaler before I run it though the AI language model.

olalonde• 3 years ago

Because you can get points on Hacker News.

raincole• 3 years ago

Why write your thoughts on the web when other humans are going to steal and paraphrase it? I mean... you're on HN. Don't tell me you didn't notice people often regurgitate tech influencers like Paul Graham and Joel Spolsky's thoughts.

slg• 3 years ago

Imagine a friend asks for help in a class. You can either spend some time and try to teach them the subject or let them copy off you during the exam. The former generally feels good despite taking more effort. The latter often feels bad even if it doesn't impact you negatively in any way and helps your classmate more than if you did nothing.

The human to human connection that a blog or social media conversation creates feels a lot more like teaching your classmate while the AI feels a lot more like someone cheating off your work. Plus the AI didn't even bother to get your approval before copying from you. The whole thing feels ethically compromised regardless of the ultimate result.

Mordisquitos3• 3 years ago

Anonymous people regurgitate the thoughts of well-known individuals such as Paul Graham and Joel Spolsky. The fact that their thoughts are regurgitated is a testament to how well known they are already and how much their content is read by other people. Nobody is going to steal their limelight only on the basis of paraphrasing their ideas. However, if someone does write original ideas of their own, they may gain some notoriety for themselves.

Now imagine that Paul Graham and Joel Spolsky were able to read everything being written by every anonymous unknown on the internet, and create content paraphrasing any and every original thought that was created by anonymous individuals at will. How do the original creators of these thoughts have any chance to succeed on their own merit, if Paul Graham and Joel Spolsky (who everyone knows already as sources of ideas) are able to write the same stuff as soon as the anonymous person has made it public?

Swizec• 3 years ago

Becoming part of the cultural lexicon is the ultimate goal of thought leadership.

Just look at how many people say stuff like “Two women can’t make a baby in 4.5 months”. Someone (Brooks) had to invent, write down, and popularize that analogy.

JohnFen5• 3 years ago

This was the place I reached. I'm not concerned about "stealing", exactly, but I don't want to contribute to this technology.

I think my days of sharing things freely on the web are over.

nodemaker4• 3 years ago

I would like to make the opposite argument. All these days I didnt share my thoughts because everyone else was and my voice would be drowned in a sea of voices. In post GPT4 era its easier to stand out if your thoughts are actually original and refreshing because most people sound like their thoughts have been written by GPT.

To rephrase it another way, the reign of the conformist ends here and the reign of the contrarian begins now.

EamonnMR3• 3 years ago

Or just the reign of brevity. Sheer volume is no longer impressive.

lupire2• 3 years ago

Your ideas are low probability autocomplete. GPT wants popular ideas, not novel ideas.

throwaway675309• 3 years ago

A lovely sentiment in theory, but Waldo is still perniciously difficult to find even though he dresses differently from every other character.

whywhywhywhy4• 3 years ago

So maybe only post dumb and incorrect information.

Train it to be wrong on purpose, for a joke.

Gigachad• 3 years ago

Only until we plug it in to the real world with sensors and ability to conduct new research and observations.

pixl975• 3 years ago

I mean with GPT-4 and image layers we've already started that process.

ProAm• 3 years ago

PG is back on Twitter? I thought he left a month or two ago?

coldtea• 3 years ago

That might have been to jump on the fashionable wave virtue signalling wave. No longer needed anymore

tredre3• 3 years ago

He was briefly suspended early this year for tweeting "You can find a link to my new Mastodon profile on my site."

StrictDabbler• 3 years ago

http://ascii.textfiles.com/

Gosh, why would anybody bother archiving Yahoo answers, Angelfire, Geocities, Tumblr, Myspace, Friendster, old BBSes, old Apple II and C64 and PC floppies, Usenet, forums... what value does any of that have?

voz_• 3 years ago

The more I see of his writing, the less I think of it. I wonder what Diogenes would think of him…

satvikpendem• 3 years ago

Behold, a man.

pixl975• 3 years ago

For a time. but as we bring audio/visual AI online then it will have another boom of incorporating humanities data in that form. Then we'll have another boom of AI robot learning by experiment with reality.

After that point it gets tricky to figure out what if any booms will be next. When you get near AGI lots of horizon problems crop up.

advisedwang• 3 years ago

A lot of the ground truth for AIs (and it's not just training data - it's also ongoing validation of quality) is coming from companies like Appen, Sama, DefinedCrowd, Q Analysts and many others. There's a lot of variation, but the trend is moving towards low-wage/gig work/outsourcing.

I think Paul means someone will be writing content, but whatever the form it's going to be a whole class of low-wage workers enabling tech from here on.

angrais2• 3 years ago

That's been the case for literally more than 20 years via Appen, Amazon crowd workers, etc.

mo_42• 3 years ago

Or the AI will trigger people to provide necessary training data. If I would run OpenAI I would provide a free version of ChatGPT that is slightly tuned to extract useful knowledge out of the people who use it. There might be adverserial attacks but overall enough people will use it blindly and provide useful information. People even trusted Eliza. Needless to talk about what we typed into Google.

ggm6• 3 years ago

Are you familiar with what is called "the drunkards walk" Because if you think stochastic inputs will not unfortunately admit of less benign paths being taken inside the dataset.. I think you're probably wrong.

I have very little doubt the primary problem in the GPT<x> model is going to remain: it is capable of reproducing highly believable crap. In a world of pizzagate, that has a risk of becoming highly weighted "I told you so" and self-reinforcing.

0xcde4c3db2• 3 years ago

There might be adverserial attacks

In both directions, even. Cunningham's Law ("the best way to get the right answer on the internet is not to ask a question; it's to post the wrong answer") comes to mind.

thelittleone• 3 years ago

Does this assume we are not AI?

zone411• 3 years ago

Human curation of AI-generated content is the true future.

dgb23• 3 years ago

I mean people forget that this is just a tool that makes a noisy compression of large data sets. It can be used to generate data that would be repetitive and tiresome to generate yourself and you still need at least one pair of eyes to discern it.

What goes in needs to be curated and what goes out too. It's getting more precise and varied in what it can do, but it's still a tool.

s1k3s• 3 years ago

We can't all use AI. Only those of you who can afford to pay our subscription.

GaggiX2• 3 years ago

People will generate the dataset using AI tools too, you can create garbage with or without AI, you can create useful data with or without AI.

ethanbond• 3 years ago

Seems like there would be an upper bound (probably a pretty low one) on how much useful data you could produce with AI.

If it has no way to check itself (i.e. having true data poured in continuously), hallucinations would just spiral out of control wouldn’t they?

Even if that’s not true, still every player in the space will use that method, negating their advantage and ceding advantage again to those who can layer real data on top of what everyone else is doing.

GaggiX2• 3 years ago

I'm talking about using AI as a tool, not to just plug it into the internet and let it generate whatever.

pixl975• 3 years ago

With AI doing things like image recognition it seems like real life observations and behavioral data may be unlimited at some point. How useful that will be, i guess we'll see.

flangola72• 3 years ago

There's already evidence that training an AI on its own output improves it. That's essentially what humans have been doing all along.

Mordisquitos3• 3 years ago

Verging off-topic, but is there a specific reason that so many of the replies to the tweet are applying the argument analogously to index-fund investing?

I'm not even sure whether the spirit of those replies is mostly to criticise the argument by showing it is absurd (on the premise that index investing is a known good choice), or whether they aim to criticise index investing itself (on the premise that the argument is good).

Maybe I am missing some context. Has Paul Graham strongly supported index funds, and so are these replies some kind of "gotcha"? Or is there some general disdain for index funds in some circles, and so are they just taking the opportunity to disparage the value of index funds by using this argument?

travisjungroth• 3 years ago

I think criticizing the argument. It's so common for a non-positive logical conclusion to be sort of a half hearted reductio ad absurdum (disproving by showing the consequence is absurd) that many people mistake this for happening when it's not. So, they argue against the sub-text that may or may not actually be there.

Or, maybe they're just copying the joke. (The joke aspect being that even when saying how AI won't dominate everything, he's referring to human works as "training data", so it's at least dominating the framing.)

Wowfunhappy• 3 years ago

I personally found the index fund comparison quite thought-provoking. Not necessarily supportive or countering the original point, but interesting.

Dr_Birdbrain• 3 years ago

I am both an index investor and an AI enthusiast, and I think it’s an excellent analogy.

If you are an investor who does not have interest in maximizing your returns and you just want a place to park your savings, an index fund is great, and will return the market average return.

Similarly, if there is a skill that you need that is not your “special” skill, it makes sense to subscribe to the “market average” of that skill, as implemented by a statistical language model, which is in a sense “averaging” over the collective language skill of the entire internet.

In both cases, the ingredients of the “average” are formed by the people who try to do better. Active investors, including hedge funds and short sellers, think they can do better, and they provide signal that in turn feeds back into the index.

Similarly, anybody who thinks they are better at writing than the language model is free to try to outperform it, and eventually will feed back into the training data.

arjie• 3 years ago

I said that there https://twitter.com/arjie/status/1635693927071387648?s=20 some hour and a half after his post which is about when I saw it.

I didn't realize others also did. They felt isomorphic to me. If the AI generated content is regurgitation of previous content, novel human-generated content will stand out and human-generated content will become valuable. i.e. the balancing mechanism is that novelty will beat dumb replication in the same way that an active manager who has true alpha can beat a passive fund.

Now, personally, I think that multimodal LLM-groups plus feedback plus many more sensors is not far from where we are, so I mentioned that if AIs do become creative it's not a problem - they will be us or better! I, personally, believe that's where we'll go.

Anyway, I'm not too inclined to discuss in this forum (though we can chat if you're in SF) since it's easy to misunderstand and jump into argument spirals, but I just wanted to make it clear that it's not a "gotcha" or a dunk or anything like that. Nothing is being disparaged.

It's just a conversation like a normal conversation where you bring up what you think and mention ideas you have that others might find interesting. If many others have brought it up, about the best I can say is that in this respect it appears I am not particularly inventive - perhaps the best evidence that LLMs are already human-like ;)

PeterisP• 3 years ago

Even a minimal amount of curation - e.g. generating multiple options and choosing the "best" one, or applying a filter where you don't post some of the generated data because it's bad - results in valid, usable training data, because the distribution of that slightly curated data differs from the generating model's distribution in exactly the direction which the curating human found most important.

So even if we all do use AI, the training data will get generated unless we don't get involved at all and AI does literally everything 100% solely on its own, at which point we could concede that it probably doesn't need any more training data.

usrusr• 3 years ago

"But after saying it out loud, as it were, I realized that with current models of AI, at least, the more people shift to using AI, the more influence accrues to those who don't."

The last humans shouting into the AI models echo chamber might have all influence, but that does not imply that they have anything even remotely resembling control. That joke/not-joke isn't half as consoling as it might seem at first glance.

LarsDu882• 3 years ago

And companies that can generate the training data at scale will be the highest value ones.

These LLMs are the shiniest thing in AI/ML right now b/c the data they use for training is already freely available and massive in scale via the internet.

There's an entire universe of data that hasn't been collected, curated, and leveraged in the right way yet: - The DNA sequences of all living things - The DNA sequences of all of humanity - The CAD files for all manufactured goods - The motion data for humans working manual jobs - The words spoken by multiple peoples across their entire lifetime

Teever9• 3 years ago

To me the end game of this all is a self replicating machine. I'm watching this whole thing with curiosity, waiting for the first image model to be trained on schematics and circuit diagrams.

I want to be able to ask an AI to give me the schematics of the machine it runs on, with instructions on how to build that machine, by hand if need be.

And then I want to ask it how to build a humanoid robot that can do the same work, and I'll build the humanoid robot and let it build the machine that the AI runs on.

I'm not sure what happens then. I think that might be the end of the economy as we know it. Or perhaps the real start of an economy, and everything that we've experience before then had just been a crude approximation of what a real economic system could be.

em-bee• 3 years ago

the self replicating machines will optimize to become smaller and more efficient until everything turns into gray goo.

lofaszvanitt3• 3 years ago

Don't kid yourself that they will let the average joe have that ability.

pixl975• 3 years ago

And this is why we don't let people build the paperclip optimizer.

Making a machine that could do what you say above probably isn't that hard.

Making the machine that could do the above without destroying the world is impossible.

LarsDu882• 3 years ago

I think its good not to overthink this idea of a self-replicating machine. In the GPT-4 paper, the system was given some access to some cloud apis and was able to do things like requisition more resources from the cloud. I imagine the first stage of "self-replication" for these LLMs or their successors is going to be like ordering CPUs from Amazon rather than setting up silicon fabs. You need to get far beyond single human level intelligence and effort to start building chips and we are probably not even there yet!

juujian2• 3 years ago

That sounds like it would be ripe for abuse. If companies get $$$ for big samples, there is an incentive to fake it. And garbage in=garbage out. Quality of AI/ML models might even degrade if the "raw material" is contaminated. Pre-GPT 3.5 data will be really valuable.

0xcde4c3db2• 3 years ago

After playing around with Stable Diffusion for a while, I'm convinced that big media conglomerates like Disney and Warner Bros. Discovery are going to have a huge advantage here because they're in a position to pipe absolutely massive media libraries into models that are big enough to legally become "copies" of the involved works. While some of the angry rhetoric from artists whose work has been used to train models may be overblown, it's far from a stretch to suppose that companies like OpenAI and Stability AI, if they were being sloppy, could overfit to copyrighted training data to the point that the models and some outputs are infringing. I suspect they're aware of this risk. For instance, the differences in consistency and detail between a batch of SD outputs using "American Gothic painting" vs. "Dark Side of the Moon album cover" are very striking; the former is reproduced with remarkable fidelity and consistency, while the latter was clearly observed but has a much more abstract relationship to its iconic source image. I don't think that difference is accidental; I think they went to some effort to ensure that only public domain images were so firmly encoded into the model. That sort of thing is not a problem for entities that own the involved copyrights or have a such a broad license that they might as well own them (and, even failing that, are in a strong position to demand such concessions from artists going forward).

CuriouslyC• 3 years ago

These technologies aren't really going to help big rights holders nearly as much as they're going to help small players and indies. Big rights holders already have efficient pipelines, automation and well tuned art departments. They can already crank out movies that are so effects laden as to be visually overloading non-stop. These tools might let them achieve similar or slightly better quality for less, but it's not going to fundamentally change anything. For example, if Avatar 2 VFX cost 200 million, maybe it'd come down to 20 million, but the other production costs would still be significant.

On the other hand, small indies can now do respectable VFX without resorting to CGI, needing expensive makeup/props or green screens. They could film in a field with actors in Halloween costumes, then place them in locations built using simple geometry in blender, and then AI will texture, style and light everything, modulate voices, etc. It'll be a game changer in terms of the types of films people will be able to make at the <100k price point.

Aerroon3• 3 years ago

Big media conglomerates might be able to pay to create enough art for their own models. Or at least figure out some scheme to do it, where the artists freely sign away this right. That model could then be licensed to others.

A website like YouTube could probably do this for voice or sound generation, if they haven't already. Offer creators an extra 5% of ad revenue and in exchange they get to mine the video for data for AI. Or some other such benefit.

koeng• 3 years ago

What about the SRA or Genbank? Those are pretty readily collected DNA databases. You can get curation from uniprot or embl or pdb or kegg or rhea.

vishal01232• 3 years ago

I think the biggest value would be products tracking users, just as it currently is, like Google, Facebook etc. If assuming AI could learn from low quality data like humans could now, these companies have a huge dataset consisting of multiple GBs per living person that is available to them now including (verified) human written texts, search history, browsing history, video call logs/transcripts, translation transcripts etc.

In future company could even pay users to get access to their keyboard and mic to get the data that is verified to be human.

sva_• 3 years ago

This comment read like the intro to some kind of dystopian movie.

mbgerring4• 3 years ago

He’s right. The rise of AI will mean that specialty, proprietary data will only become more valuable. The use of encrypted communication in place of social networks will also rise. Nobody is going to post anything on the internet for free in a place where it can get vacuumed up as training data and regurgitated, without credit or compensation.

I realize I am posting this on a public forum that is almost certainly being used as a training corpus, but I’ve mostly withdrawn from public social media at this point and I wouldn’t be surprised to see that happen with more people.

nocsi• 3 years ago

Are those same people also going to stop speaking in public? Throughout human history, anyone can profit off of hearing something you said in public. You should be fearful of a timeline where people are compensated for whatever they post on the internet. I shudder to think what those discussions would look like

doodlesdev• 3 years ago

   > Throughout human history, anyone can profit off of hearing something you said in public.

I think the big difference here is that this will now be automated. This comment I'm writing right now is being "donated" to any company that wants Hacker News on their dataset, but they didn't even need to go through the work of reading my comment to use it, they just feed along with hundreds of terabytes of random text they get from the internet and apply heuristics/other models to filter the dataset. I feel somewhat uneasy about it, specially when writing code now. I don't think my code is good in any way shape or form, but there's a reason I license them as GPL, I don't want it used without being contributed back in some way, I write that for the greater good. License infringement was always a touchy subject since it's quite hard to find license infringement in proprietary software, but now that an opaque black box is involved in that process the people write the software might not even know they violated a license.

Just throwing ideas, I'm not really in favor or against LLMs using public datasets. At least on one side, it levels the playing field among all participants. On the other side, however, big tech will always have an edge in training and testing these large language models. My current stance is to just wait and see what happens, and then react accordingly.

pmoriarty2• 3 years ago

"Nobody is going to post anything on the internet for free in a place where it can get vacuumed up as training data and regurgitated, without credit or compensation."

Virtually nobody's going to care enough to change their behavior.

symlinkk• 3 years ago

People will start caring really fast when the job postings dry up

thequadehunter2• 3 years ago

I have seen this take on a few HN threads and I don't know where it comes from. People didn't stop reading books just because audiobooks got popular.

Why would people suddenly just turn to AI for all their entertainment? Part of the reason people love art is because they relate to it.

Like, in 2023 with the internet, blog posts, youtube, reddit, movies, etc, there's still enough demand for books that people buy them. Just because a new medium is created doesn't mean it suddenly just becomes the only medium.

swader9992• 3 years ago

A lot of entertainment is already formulaic. Disney, Marvel etc.

Jensson• 3 years ago

People stopped reading handwritten books once the printing press became a thing. So yes, new technology sometimes does completely eliminate the old. We aren't sure where on the line this tech will fall, maybe a future LLM will be good enough to generate endless stories at high quality, or maybe they will always feel soulless, we don't know.

aphexairlines• 3 years ago

Is this similar to people moving away from giving their code as open source when companies take it, repackage, and sell it while the authors get nothing?

zirgs12• 3 years ago

Most open source licences don't even prohibit selling apps that contain open source code.

jameshart• 3 years ago

It's possible that people looking back will consider that the mistake was putting all the content online. Perhaps even upstream of that: the first mistake was digitizing things. The music industry certainly didn't realize when they adopted CDs that they were starting down the path to self destruction... the newspaper industry likewise didn't notice how profound taking their newsprint product and packaging it as HTML would be...

And now we're unleashing ML training on all that digital, online data. Which industries will discover that this is the thing that means putting your data online, digitally, was a mistake? Certainly artists are feeling it now... maybe programmers, too, a little.

So how do you put the genie back in the bottle? Live performances, with recording devices banned? Distribute written material only on physically printed media - but how to prevent scanning? Or just escalate the DRM war - material is available online, but only through proprietary apps on locked down platforms?

Or is this going to take regulation - new laws to protect copyrights in the face of ML training?

It wasn't always the case, that you could assume that if some information exists, it should show up in a single search. That's an expectation we invented only about 25 years ago. It's possible that the result of all this is that we figure out that we can't actually sustain the free sharing of information that makes that possible.

The problem is, to borrow a phrase: information wants to be free...

mbgerring4• 3 years ago

We’ve known for a long time that “information” doesn’t want to be free. The owners of platforms want you to not value your information, so that you give it away for free, and they can turn it into something they can sell. That was a hard, bitter lesson, and its so important for people to learn it, quickly.

beepbooptheory• 3 years ago

In less than 24 hours, there will be (conservatively) 100 different people independently saying this very point at different corners of the internet. It all mostly redundant.

Don't get tricked by the past decade or so's focus on the idea of "content"! This was all mostly a ploy by YouTube et al to get people to make more channels so they could sell more ads.

There is no real value to your sense of individuality in an economic sense, you shouldn't worry about it being exploited. Don't confuse the (beautiful) experience of your own interiority with a commodity.

But regardless of all that, I feel like this is an incredibly weird way to intepret what he is saying. Its not about the rising value of human generated datasets, its about the fact that there is still necessary well of labor that all this stuff must pull from in order to eliminate other forms of labor. That is why he says "not everyone can use AI." Not "proprietary data will become more valuable."

If it will be valuable, who will own this value? Probably not the same people generating and curating it, and those are precisely the people who can't use AI.

(And there are at least 1000 people making my very point this very moment..)

anonzzzies5• 3 years ago

Nobody is going

Ah! The HN echo chamber. The world at large vastly does not care at all. Govs are banning TikTok apps on gov employees phones all over; does anyone care and use less TikTok? Only HN and probably many there are lying about it. Privacy and this type of content abuse prevention is valued by a handful of people unfortunately.

mbgerring4• 3 years ago

How valuable, in terms of general knowledge, is the free-to-scrape social media posts of people who don’t think about any of this?

I don’t think people are going to continue investing time and effort into giving away high quality knowledge on the internet just so Sam Altman can train an AI to repeat it, put it behind a paywall, and charge you for it.

Yes, people will still post memes. How useful is that as training data?

tanto• 3 years ago

I agree that proprietary data will become more valuable. It is, even today, mostly not accessible for AI training and holds so much value. We are working on Flower (https://flower.dev), which enables training AI on private data without the data owner having to share it.

joenot443• 3 years ago

Nobody is going to post anything on the internet for free

If there’s one thing I refuse to believe people will ever stop doing, it’s posting.

I’m genuinely curious how much time you spend with non-techie people to come to the conclusion that your opinion comes close to representing the masses. Try to make this argument to a 16yo posting dances on TikTok or a boomer reposting poems about their grandkids.

“Don’t you understand?! The big tech companies are vacuuming up your posts as TRAINING DATA! It’s going to become part of the CORPUS! You have to be like me and WITHDRAW from public social media!”

“Yeah I don’t care lol”

BeFlatXIII2• 3 years ago

Now even normal people have the incentive to act like troll bots and spam the public internet to pollute the training data.

jacobsenscott2• 3 years ago

True - we can't all use AI. AI in its current form is only useful for investors to profit off the hype cycle. The rest of us are continuing to produce content for microsoft to steal and re-sell (or "training data").

pbreit• 3 years ago

What would be a good tool to train on a "knowledge base" (a set of URLs) to be used to field customer support inquiries?

jacobsenscott2• 3 years ago

Do openai etc respect robots.txt, or has the world decided a few billionaires can repackage everyone else's work and sell it back to them no matter what?

zirgs12• 3 years ago

OpenAI most likely does. But the folks at places like civitai and 4chan most likely don't.

The tech being open source means that if someone wants to steal your drawing or writing style then they will do it and there's nothing that you can do to stop it.

telotortium• 3 years ago

One day, a social network will require a real-time connection with a Neuralink connected to a live human to confirm that a live human is typing/speaking. That will only work until an AI is trained well enough to spoof Neuralink's signals.

juujian2• 3 years ago

Let's be honest with ourselves. We won't be able to tell "real" data from ai-generated data and end up training our data on ai-generated data.

fnordpiglet2• 3 years ago

I think this missed a key point. If you use AI to generate content, you are still a part of the data emitted. The prompt encodes what you want to say. “Write a new article about a bank called SVB which failed because of a run. The run was induced by a fire sale on assets resulting from rising interest rates and bad rates hedging strategies.”

This will write a story about that with all the language boiler plate. However, the story wasn’t whole cloth AI. A human told it the details to provide, and it provided the best expectation of the language to convey that story. Any future AI training would reinforce the grammar and style, and incorporate the information of the prompt as well.

I, for one, welcome our large language model overlords.

paxys• 3 years ago

How does the model know what SVB is? How does it know there was a bank run on it? What assets did it have? Why were they bad? When did the crash happen? What was the final outcome? Your prompt didn't add any of that context. It still needs to fill in all these details from, say, a bunch of news articles. But then the people who write these articles will eventually go "hey I can just use AI". So ultimately where are these core facts about our world that the machine spits out coming from?

fnordpiglet2• 3 years ago

You are right, my prompt is incomplete. It is an example. As a journalist with first hand knowledge of the events I can prompt craft a complete article with my core facts. Most of what you asked is already out there, possibly all prompt crafted by a human with knowledge of the subject.

If I prompt craft an article about SVB, and the LLM learns to emit language about SVB from my article, then all future prompts can utilize what it has been trained in. The origination of facts would be humans writing prompts. But the language models would still incorporate that.

My point is that humans would still be responsible for processing facts of the world and entering it into computers and computers would “learn” those facts from the prompt created content. Maybe we would skip the prompt creation step and just start encoding facts directly as semantic facts and we can stop writing language to encode a fact, and just use language models to create human readable communication of fact. It doesn’t matter. Facts would still be sourced by humans, even if the language is crafted by machine.

99_00• 3 years ago

GIGO

pmoriarty2• 3 years ago

I'm awaiting the day when the top commenter on HN is revealed to be an AI

gavinhoward2• 3 years ago

I believe this 100%, and that's why I refuse to even try these tools. Not even once.

Dalewyn11• 3 years ago

Nobody is going to "generate training data" if they get precisely nothing out of it.

The complete disregard for intellectual property rights being exhibited by the "AI" community will inevitably come back to bite them in the arse one day.

ramesh31• 3 years ago

LLMs are not regurgitating copywrighted works. You're fundamentally missing what's happening here if you think that. They do the same thing you and I do; read the literature and come to a conclusion. To quote Wittgenstein, "My words are my world", i.e the sum of the textual and verbal output I am capable of producing are me to the rest of the world. In the same way, LLM AIs are approaching something indiscernible from human intelligence.

partiallypro2• 3 years ago

This would be well and good if you couldn't occasionally ask ChatGPT for a response then Google it and find a near identical response elsewhere in what would have otherwise been copyright materials. Other AI tools are doing the same thing, and the most blatant is with art. If you honestly think ChatGPT etc aren't training on copyrighted data and just rewriting it for tons of responses (and images), you're out of your mind. It is just incredibly difficult to prove, and the laws currently are very blurred.

Dalewyn11• 3 years ago

They do the same thing you and I do; read the literature and come to a conclusion.

No.

srcreigh5• 3 years ago

The key difference is that we know how LLMs work but we still haven’t solved the problem of consciousness, or even neuroscience.

You can’t equate something you understand with something you don’t understand. Something you can create with something you can’t.

nostromo123• 3 years ago

"They do the same thing you and I do; read the literature and come to a conclusion."

I'm sorry for you if you can't do better than an LLM.

partiallypro2• 3 years ago

There will be tons of lawsuits in the next decade over the data AI is using to generate its responses. Some of the most blatant and easy to prove have been in image generation with them literally including blurred Getty logos with their output.

Zuiii• 3 years ago

Are the blurry getty logos imposed on backdrops from getty as well? I think what's happening here is that the AI is treating this feature like any other feature (table, nose, paperclip) and included it because it assumed the context was appropriate for it.

On the otherhand, if the AI can be coxed into producing nearly complete getty image with only minor modifications, then maybe there would be grounds for a copyright lawsuit.

I think AI detractors should focus on passing new laws. Copyright is not the right tool for this.

whywhywhywhy4• 3 years ago

Getty are going to be one of the biggest investors in the space if they’re smart. Why pay photographers when you already have the dataset to create them on demand.

donpark• 3 years ago

Derivative training data is already part of the training regime. Earliest example produced rotated images for more robust image recognition.

IP laws are legal constructs and subject to changes like everything else depending on our needs. When and if enough of us see benefits from AI hugely outweighing benefits from maintaining IP framework, I'm pretty sure laws will be passed to make needed adjustments.

chitowneats• 3 years ago

Literal shape-rotating?? I thought it was only the wordcels who were in trouble.

srcreigh5• 3 years ago

Any automated process for generating training data is part of the model, not part of the training data.

Authentic human generated information has irreducible newness.

salawat2• 3 years ago

See, until then though, you don't get to ignore the law as it is.

You see, you're part of the system, not outside it. If it were not for my deep, deep hatred of the farce we call IP law, I must admit, I'd seriously consider walking over to the other side; because I look at these AI folks scouring training data without so much as asking the original creator for consent, or even sparing a couple cycles to think about whether they should, and it really rankles me.

In fact, the entire damn computing industry is filled with people with some seriously flawed ideas around the concept and mechanics of the act of obtaining consent.

6510• 3 years ago

If learning is not allowed the robot overlords can move to countries not on the IP bandwagon like Algeria, Argentina, Chile, China, India, Indonesia, Kuwait, Russia, Saudi Arabia, Ukraine or Venezuela. Those are on the priority watch list, there is also a normal watch list. Those are probably lax enough to allow a lot.

https://ustr.gov/sites/default/files/2019_Special_301_Report...

franciscop8• 3 years ago

"Intellectual property rights" is a concept made by greedy proprietors to build monopolies and keep competition out. It's no chance it was strongly pushed by Disney, and in its current form goes against the society interests as a whole. It is sold maliciously as "it incentivizes creative work", but how does "life of the author + 50 years" incentivize the author to keep creating? It just incentivizes them to keep exploiting.

Now I think we'd all agree to something more realistic to actually protect authors, like 10-30 years protection from publication. But every time someone brings "copyright! IP!" to ML I just laugh since it's one of the more corrupt laws that we have currently.

wuqs• 3 years ago

excuse me sir, are you blind by silicon? greedy is to train ML models on copyrighted data of hundreds of thousands of people with far beyond the amount of power Disney have!

also how many people are responsible for stuff like ChatGPT... 300? getting the bulky of their trainning data from, again, who?

it feels the monopoly history is re-inventing itself, now with fancy computers replicating human behaviour

jollofricepeas3• 3 years ago

I disagree completely.

My older brother decided to go to art school.

It wasn’t until he was 40 that his artwork started to be noticed and worth something.

His images have been bootlegged endlessly by Asian companies placing them on t-shirts and selling them for $5 on Etsy and every other marketplace.

Fortunately for him, he’s designed a very popular set of kids books that have allowed him to move out of my parents, get a place of his own, pay down debt and buy a house.

Intellectual property is how he takes care of his family. He should be able to at least provide for his kids college after his death if his IP still has value in the marketplace.

…And so should other creators.

Not everyone is Disney.

HWR_143• 3 years ago

Intellectual property rights" is a concept made by greedy proprietors to build monopolies and keep competition out

Which is why all AI code and datasets are freely available and legally usable by anyone for any purpose?

UberFly• 3 years ago

"Intellectual property rights is a concept made by greedy proprietors to build monopolies..."

If you had created anything other than internet forum opinions in your life you'd have a different perspective.

zirgs12• 3 years ago

The Mickey Mouse protection act actually extends copyright to more than 90 years after publication. It was imposed upon the rest of the world unilaterally. And only few big corporations like Disney benefit from it.

Juliate2• 3 years ago

It’s a concept at least as old as modern democracies and individual rights, and it’s likely not a coincidence.

That it’s been abused by corporate entities is another matter. The fact that some famous musicians sell their whole catalog to companies is also a sign of times changing again.

AI growth as it is here based on preexisting work is directly in confrontation with copyright. And something will have to give in some way. But law will likely not catch up before at least another 10 years, the time needed for most people to really understand/experience the impact of the changes.

precompute11• 3 years ago

IP laws are a double-edged sword because of the way how advertising and human memory works. You can not have IP laws and also negligible impact on culture. You always get both, and people replicate what they know. The goal of IP (like many other things) is to eradicate itself.

jstarfish2• 3 years ago

how does "life of the author + 50 years" incentivize the author to keep creating? It just incentivizes them to keep exploiting.

It is a necessary protection against moral hazard, and a way for artists who may well be poor during their lifetime to be able to leave something behind for their children.

In your world, if one wants to print Calvin-peeing-on-things t-shirts without having to deal with legal hurdles, all they'd have to do is shove Bill Watterson into traffic to force his creations into the public domain.

There is no point to creating anything if someone else can throw you under an actual bus to immediately claim rights to (and start profiting from) the years of time and effort you spent building the "brand." Forcing a delay of 50 years goes a long way toward discouraging this sort of literal piracy.

This is the premise of the current AI debate, only it's forgery at scale instead of manslaughter.

misssocrates• 3 years ago

Predictions on how it's going to come back to bite them?

evdubs2• 3 years ago

5% chance the copyright industry will sue AI companies for verbatim reproduction of copyrighted works.

0.1% chance the Feds get involved and sue AI companies for criminal copyright infringement.

precompute11• 3 years ago

It will not bite them back, because they will argue that what was put in the dataset was "easy" and not worthy of copyright anyway. Then, they'll tout how smart their "AI" is. Two birds with one stone.

hn92726819• 3 years ago

Nobody is going to "generate training data" if they get precisely nothing out of it.

What? What do you call literally the entire training dataset of stable diffusion, dalle, gpt3, and gpt4? You think they compensated anyone that went into the dataset? Did you get paid for your comment, because guess what? Your comment will end up in the gpt4.1 or gpt5 model.

Training data will likely continue to include everything public on the internet. Not everyone will stop posting things online for free, therefore humans will still generate training data for models.

Dalewyn11• 3 years ago

You know what people treated like shit do? They go find greener pastures to spend their lives in, taking any of their benefits to humanity with them.

Sure, random comments of insignificant value like these will probably continue being created. Valuable content, though? That's a different story.

"This is why we can't have nice things." is a saying that accurately reflects what could happen if the "AI" community doesn't stop behaving like uncivilized barbarians. Personally, I'd prefer to have nice things.

ipaddr2• 3 years ago

Soon we get content created for gpt and introduce a new level of spam.

matheusmoreira12• 3 years ago

The complete disregard for intellectual property rights

Those "rights" should not even exist to begin with. Nobody is interested in bending over backwards just to appease a bunch of aging monopolists. We want to explore the full potential of technology, not see it crippled by artificial limitations due to "concerns" over imaginary property.

phone8675309• 3 years ago

That's an incredibly reductive view. Copyright protects me when I do any creative work, I do not want my creative work to be part of any AI training dataset without my explicit permission, and I want laws that force people training generative AI models to copyright clear all of their training data and moreover be able to show which inputs corresponded to which outputs so the result is auditable.

If Disney trains an AI good enough to just generate a good story then I think that everyone who fed training data into that model should get a piece of the revenue.

grensley• 3 years ago

I'm genuinely curious how you think artists should be compensated for their work.

Fomite• 3 years ago

Instead, we're bending over backwards just to appease a bunch of slightly younger monopolists.

emodendroket6• 3 years ago

How can anyone take seriously the notion that anyone who has written a book and seeks to earn money from people who want to read it is a "monopolist"?

evdubs2• 3 years ago

When you download a couple epubs from a torrent site, you're an evil pirate stealing books. When you're a FAANG-affiliated AI company and you download all the ebooks and charge for the derivative works your algorithm produces, you're a visionary changing the future.

rhtgrg• 3 years ago

Yeah, this is just capitalism in action. Almost every startup has stories like this, which appears to be why YC's startup school shies away from "goody two-shoes" type people and has questions like "please tell us about the time you most successfully hacked some (non-computer) system".

For example, in the early days Amazon famously "leveraged" the purchase of out of print books from their wholesalers to obtain orders that would otherwise be too small to ship.

That's just how capitalism works. You have to be either too small to bother with or too big to fail -- the problem area is in the middle.

grabeh• 3 years ago

Reminds me of the quote: "The death of one man is a tragedy, the death of millions is a statistic"

erfgh• 3 years ago

Yeah but what's the alternative?

BlueTemplar2• 3 years ago

Reminds me of the failure (?) of the Google Books project :

https://www.newyorker.com/business/currency/what-ever-happen...

XorNot• 3 years ago

No one's going to need to generate training data, because image generators will simply setup and run their own artist pages using LLMs to interact with their fans, and then weight their outputs by likes/clicks/web requests.

The idea that complete original "art" is needed to run image generators is a very temporary state of affairs that bootstraps the proof of concept, but is in no way necessary to the process. The entire process can be adversarially trained solely against human aesthetic values - we know this, because image generators are already a derivative technology of building image classifiers - i.e. they functionally are adversarial against a given textual description in a classifier.

The real limitation right now is speed: once we can diffuse at > 1 frame-per-second, data collection can be run solely from the consumer side.

scotty793• 3 years ago

People create stuff out of the sheer desire to create stuff. You only need to make sure they won't starve while they do it. And there's so many ways of ensuring that apart from paying them for copies of their output.

colordrops• 3 years ago

Intellectual property rights shouldn't exist as a concept. Digital scarcity is a scam. Property is a function of the natural scarcity of physical matter, which doesn't apply to data. Rent seekers and power brokers unnecessarily brought this concept into the digital realm.

_ZeD_• 3 years ago

funny, people invented, discovered, and created stuff even before the "intellectual property rights". Those terms have been created by distributors, not creators, of contents.

imgabe4• 3 years ago

Nobody is going to "generate training data" if they get precisely nothing out of it.

How much did you get paid to write this comment?

nicbou4• 3 years ago

People are happy to share their opinion, but don't expect people to write long, well-researched content about anything.

In my case, the content I write involves hours of research, and correspondence with a network of experts I have built up over 5 years. I constantly update and revise that content.

The website on which it's hosted was also developed with great attention over the same period.

I don't think that I'd keep doing if I was deprived of revenue, and from the kind feedback of readers. I certainly couldn't do it full time.

taneq• 3 years ago

Their compensation was a feeling of contributing and being part of a community, probably with some degree of self-validation as they read and agreed with their own opinion after posting it. Later they might even get paid with some upvotes granting additional affirmation.

Just like us, really.

montag• 3 years ago

Haha, great rebuttal.

The reward was mere internet points and the hardwired gratification that comes from social interaction.

People sometimes go to great lengths for non-monetary rewards.

bontaq• 3 years ago

This is entirely my thesis on why they'll win. Of course google wishes they could ignore all attribution.

They are slighting the law just like airbnb and uber did. They offer a better form of consumption. This is disruption at it's finest, and it will be in the courts forever.

patrickk• 3 years ago

Nobody is going to "generate training data" if they get precisely nothing out of it.

It's using common crawl and wikipedia among other sources. The original purpose is something else.

Another anecdote that I find amusing is that Google Translate was trained on EU legal texts (among other sources), which are (generally) translated into 26/27 languages of the EU members. A huge, multinational bureaucracy, generated the texts by hand, but Google gets all the credit. I'm guessing that's part of the reason that Chinese isn't so accurate compared with other translations.

DuckFeathers• 3 years ago

"Intellectual property rights" is a scam implemented by corrupt western cronies and is completely unfair to the people. No person owns the right to their "invention" because it is leeched off of other people's work. No corporation owns the right to whatever the people they employed came up with. The IP thing needs to be eradicated.

We need more countries to be strong and brave and say no to westerners forcing them to be unable to copy technologies... like China has done to some extent... but that might take too long.

The path where AI lawyers can sort IP disputes instantly and hence competitors can easily find out how to copy technology without violating laws would be another nail in the coffin of this corrupt practice.

Nobody is going to "generate training data" if they get precisely nothing out of it.

The people who we don't want generating training data will not generate it... and that is a good thing. You have been generating too much crap in the name of innovation anyway. Real tangible useful innovation gets copied naturally in the marketplace in absence of corruption.

precompute11• 3 years ago

IP is the promise that you get the dues for your work, guaranteed by the government. It is what enables the entire "information economy". Without it, we'd soon need to weed out people that don't contribute to IP.

ChuckNorris89• 3 years ago

Nobody is going to "generate training data" if they get precisely nothing out of it.

"Before accessing your account, please solve this puzzle to verify you're not a robot."

precompute11• 3 years ago

"Your OpenAI free tier requires you to install the OpenAI app and accrue 30 OpenStat points for every 10000 tokens. Please note, you can only earn 1 OpenStat point every two hours. Please do not deny Location Access, Microphone Access, Camera Access or Storage Access permissions, or this program will stop working."

dragonwriter2• 3 years ago

Nobody is going to "generate training data" if they get precisely nothing out of it.

Everyone that has a job and gets paid for it generates (with the appropriate monitoring) some kind of data that could be training data in the course of doing their job. (And doesn’t raise the issue of “if they get precisely nothing out of it”.)

Once no humans have jobs because AIs are better at them (should that occur), then the AIs themselves will be the best source of training data for the next generation.

There are many problems to address with AI; “Where is training data going to come from,” however, is nowhere on the list. It is a distraction from real issues.

zirgs12• 3 years ago

Copyright activists should not have pushed Mickey Mouse protection acts , sued fangame creators and the like. I have zero sympathy for copyright activists now. Copyright reform is long overdue.

whywhywhywhy4• 3 years ago

People just side with it as a way to get what they want. Crazy how many artists who’s portfolios are full of their own drawings of Disney IP shouting that Disney should use IP law to stop stable diffusion.

The hypocrisy completely lost on them.

version_five4• 3 years ago

People write open source software and get "nothing" out of it.

And people post stuff on the public internet and get "nothing" out of it.

It's not obvious to me that everyone who has shared something permissively is unhappy if someone else uses it for training data. In many cases, it's just loud people who want a new revenue stream, which is the problem with intellectual property generally, the concept breeds entitlement. Why should I retroactively profit from what I willingly shared because somebody else found a use for it.

nicbou4• 3 years ago

Let's say that I repair donated bicycles and give them away for free. It's a fun activity that does some good for the community.

Now, let's say that on my morning walk, I see someone selling a bike I gave them a week ago.

Is it fair to be offended by this?

account424• 3 years ago

I have released works under copyleft licenses and am not OK with Microsoft using those for profit without respecting the license (e.g. without opening up the models derived from copyleft works) when Microsoft themselves have been going after others copying their works when that benefited them. They shouldn't get to have it both ways - either we reform copyright or everyone should be required to follow it. If anyhting, large corporations should be subject to stricter enforcement since they already can benefit more due to their size.

PoignardAzur4• 3 years ago

Nobody is going to "generate training data" if they get precisely nothing out of it.

This is trivially false. Eg, fanfiction exists.

Maybe monetizing art will become hard as cheap art becomes ubiquitous, but that doesn't mean people will stop creating stuff.

giantg2• 3 years ago

"We can't all use AI. Someone has to generate the training data"

Sounds like a good way for the people generating the data to impart their own biases on it while the mindless masses comsuming the end products just nod in agreement.

imron• 3 years ago

Profession:

https://www.abelard.org/asimov.php

EamonnMR3• 3 years ago

Awesome web page, I collect these.

Karrot_Kream2• 3 years ago

Thanks this is the first thing that came to mind for me as well!

antiquark• 3 years ago

Or worse: AI generated data will (or maybe already is) flood the internet, AI will start ingesting its own output unknowingly, resulting in junk AI.

In fact, 2023 might be seen as the end of a pristine epoch of untainted training data, before it was polluted by AI.

CatWChainsaw3• 3 years ago

Fukuyama got his timelines wrong. November 30 2022 was the end of history. :)

armchairhacker• 3 years ago

Truly original content still >>> conventional mass-produced content.

Maybe one day LLMs and diffusion models will generate truly original content. But despite the hype and what some people are claiming, right now they really don't. The latest models are maybe 90% of the way there, but as GPT-4 shows, it will take more than scaling to go the remaining 10%.

When I look at AI-generated content 90% of it is the most "typical" response to the prompt, and 10% shows slight originality, but nothing truly unexpected. The most interesting pieces of AI art and writing all come from interesting, engineered prompts which are designed by a human. I don't doubt that AI will take over stock photos and generic writing, but my favorite artists and writers and presenters are still humans, the most popular songs and articles and movies are still written and directed by humans, and unless you've resorted to watching AI-generated Seinfeld and https://www.reddit.com/r/SubredditSimulator/, I'm sure you feel the same way

anonzzzies5• 3 years ago

but as GPT-4 shows, it will take more than scaling to go the remaining 10%.

That is happening though; many scientists are working on making gpts that can learn like Alpha Go Zero where it can learn from scratch with far less information to become better (more ‘intelligent’) quality. Didn’t see anything coming out yet; it is at least one thing that I think could be an AI winter if not achieved before, say 2027. We cannot keep scaling and thinking the results will improve, and, also, indeed we lack training data already, so something must change to need as much; it needs to learn smarter.

DubiousPusher2• 3 years ago

This may be true. But I think we should be open to the possibility that maybe not.

What if, human intelligence, by and large is the product of developing our own LLM. By this I mean, the logical circuits which are necessary for grammar and language structure are the basis for our generalizable symbol processing. We developed these neural network analogs to logical operations almost purely out of the benefit of being able to combine communication in a predictable way. We developed enough to achieve a kind of Turing equivalency. Essentially, we are a specific computer, that stumbled into an abstract machine.

Were that the case. It may be entirely possible to achieve general machine intelligence almost purely through an LLM. In which case, the LLM itself could potentially generate training data at a comparable quality to humans.

theonemind• 3 years ago

The parts of our brain that work like an LLM do work like an LLM (hallucinates/generates falsehoods/over-confident/deals strictly with an abstract world of discrete symbols) and vastly overestimates its intelligence and how much it needs the rest of the brain's reality checking and other kinds of intelligence to turn it into something useful and conflates what it does with intelligence.

It's basically a super-useful co-processor in the brain hardware, but it is a co-processor and not capable of independent operation.

throwaway18512• 3 years ago

Yes. When I read comments such as “humans just predict the next word too,” I wonder if those commenters have ever stopped and observed their own thought processes before.

berkle4455• 3 years ago

Information plateau

patientplatypus• 3 years ago

No - it just has to make the returns to inputting the training data the cost of electricity plus enough money for what a starving Bangledeshi brick maker would consider prosperity. So, web development essentially.

O__________O• 3 years ago

Human training data still currently holds a place of use, but it’s extremely time consuming to produce and frequently full of errors. Long term, future consumers and producers of information statistically speaking by volume will be AIs, not humans.

My guess PG knows this, but for whatever reasons is saying otherwise — given it’s pretty obvious; even currently, my guess is machine learning algorithms already easily consume and produce more information than humans. Yes, currently one might argue the ML is the result of humans producing it, but only matter of time before the code writes the specs that write the code that collects the data that get converted into information that’s used to repeat the process based on an objective it defines and independently evaluates.

Honestly not sure why PG is even thing on HN anymore given he rarely comments on HN, or if he does, it’s under pseudonyms:

https://news.ycombinator.com/threads?id=pg

jay_kyburz3• 3 years ago

I have a question for the AI experts in the room.

Is it possible to feed the AI a set of training data, say a document, or even a conversation with a particular person, that is weighted significantly heavier than everything else it has ever read?

For example, if they wanted to, could Open AI just tell GPT 4 that Bananas are blue so that, when anybody else asks for a list of blue fruit you get Bananas and Blue Berries.

shagie3• 3 years ago

Yes.

You can use fine tuning ( https://platform.openai.com/docs/guides/fine-tuning and https://github.com/openai/openai-cookbook/tree/main/examples... ) to improve its knowledge set of a particular domain.

Even with sufficient work on your own with prompts, one person fed ChatGPT her childhood diaries and talked to it. https://www.marketplace.org/shows/marketplace-tech/how-ai-ca... (she would likely have been able to get more / held deeper conversations with fine tuning as it isn't limited to the number of tokens in a conversation)

jay_kyburz3• 3 years ago

Fine tuning looks like it only provides specific results for specific queries. After reading the documents you linked, I don't think it would make a fundamental change to how the model thinks about Bananas.

trekkie1024• 3 years ago

Yes. That’s called fine tuning.

raldi• 3 years ago

Once AI’s can keep state, they’ll get all the training they need from humans just by talking with us.

dragonwriter2• 3 years ago

People using AI to support work that is beyond what AI is currently capable of will generate the training data for the next round of AI, the same way that people using student research assistants to do work beyond what current students are currently capable of generate the “training data” for the next round of students. Sorted.

Next problem?

cutler• 3 years ago

I can imagine telling my grandchildren how books were once written by real human beings and that you need to install an AI-Generated-Text-Detector now the way Windows users installed anti-virus software. What a dystopian future awaits.

bfung• 3 years ago

You are correct. The process of generating high-quality training data is a critical component of building and training AI language models like me. While much of the training data can be collected through web crawlers and other automated means, there is still a need for human-generated data to supplement and refine the training corpus.

There are several methods for generating human-labeled data, including crowdsourcing, annotation, and manual labeling. Crowdsourcing platforms like Amazon Mechanical Turk can be used to solicit human input on specific tasks, such as labeling the sentiment of a particular text or image. Annotation tools, such as Prodigy or Labelbox, allow data scientists to create custom labeling schemes and train human annotators to label data with high accuracy and consistency.

Manual labeling, while more time-intensive and costly, is often necessary for more complex labeling tasks that require a high degree of expertise or domain knowledge. For example, in medical language modeling, trained medical professionals are often enlisted to label and annotate medical texts to ensure accuracy and relevance.

Overall, generating high-quality training data is a challenging and time-consuming process that requires a diverse set of skills and expertise. It involves a combination of automated and human-generated data, as well as careful curation and pre-processing to ensure that the data is of high quality and relevance for training AI language models like me.

Here is some free training data, generated by a human.

galacticaactual4• 3 years ago

You tech people can't see 10 feet in front of your face. You brought this into the world. For what? Because "progress."

valine3• 3 years ago

The “for what” is to end world hunger, cure disease, and improve our understanding of the universe.

The world is a miserable place full of overworked, underpaid people. What about the status quo do you really think is worth protecting?

galacticaactual4• 3 years ago

Lol. Yeah GPT-4 is going to solve world hunger.

Delusional.

No wonder why there was zero populist sympathy when SVB fell.

jeroenhd2• 3 years ago

I'm sure those millions of GPUs churning through Facebook posts, tweets, and random blog posts to generate the mother of all autocompletes will end world hunger, the answer is out there, we just need an algorithm to mix the My Little Pony fanfiction with the crackpot "unlimited free energy" blog in the right way! And I'm also sure these new productivity tools will be used to people can work less for more money instead of making generating more value the norm and cashing the profits at the top.

And then pink unicorns will come down from the sky and give us all candy and gifts.

CatWChainsaw3• 3 years ago

We can do that now. Every single day, we choose not to. It must be so nice to live such a comforting lie.

novok• 3 years ago

AI is the equivalent of power construction vehicles being invented, the real value is creating new styles or art that can let you generate a whole swath of something very quickly. Eventually humans will crave 'new' art, and the people who create new styles or data for the AI will be the ones

The problem with 'keeping it' will be not having new styles copied.

The computer was the bicycle for the mind, and AI is the trailer truck of the mind.

dbtc• 3 years ago

Books are bicycles for the mind, computers are cars, and AI is the computer of the mind - which makes travel (thought) increasingly less necessary.

galaxyLogic5• 3 years ago

I'm a big fan of Frank Zappa but sadly he no longer produces new material. Now his estate which owns the Zappa-IP could use all Zappa works as training data and then start producing "new" Zappa LPs. I'm sure some of them could sound as good or even better as any of the originals. Or Beatles, we could have a new Beatles album. I think that's where this is going.

The Large Language Models are huge, but it is a much smaller model if it only consists of the works of a single author.

Then two artists who like each other could join together and start producing mix-ups of their combined repertoire.

blihp• 3 years ago

I think you're far too optimistic about how this will play out: https://www.youtube.com/watch?v=jnQ0zEQPu_A

bbbnnnmmmkjkj• 3 years ago

Wow. The stuff I read recently about musk making his tweets super visible is no joke - good content by Graham but everything after is all from the platform owner

MichaelMoser123• 3 years ago

i think that's exactly the philosophical problem: the thing is not capable of original thought - it would give you only a predictable answer. I don't think that it will stray from any consensus opinion, for that matter.

not sure that it will be able to produce any breakthroughs, given this limitation - but who knows...

In a way that is the dictators dream come true: something that is competent enough for most task, but something that would never ever question anything around it.

7e• 3 years ago

All the training data the world will ever need to bootstrap AI has already been generated. A poor take.

jalapenos• 3 years ago

One thing I wonder is what happens after AI starts getting trained on its own content?

Its content is already hitting the places the next round will be trained on.

Any AI experts here, how does that play out? Does it pollute, poison, or cacophonize?

(And yes I got "cacophonization" from GPT4)

m463• 3 years ago

sinuhe69• 3 years ago

At worst, it’ll be kind of infinite recursive with psychopathic tendency. If more lucky, the law of dismissing return will strike unmercifully.

afro88• 3 years ago

My guess is it'll be like feedbacking audio. If you route some audio through a processor and set it up in a feedback loop, the way the processor changes the audio gets accentuated more and more as you allow it to feedback more.

Weights will be reinforced and accentuated and the model will become more extreme and probably less useful.

Karrot_Kream2• 3 years ago

That only happens in audio feedback because audio amplification is an increasing function. ML models do complicated things with their inputs.

jamilton4• 3 years ago

I think training on output is one method that people use to create customized Stable Diffusion models. IIRC, it doesn't work well unless you manually label it, in which case it works fine. Letting existing AI label it doesn't do very well... but maybe GPT-4 will be good enough at that sort of task?

qaq• 3 years ago

Do they though? In lit. 99.999% of works are derivative. Some similar % is for movies. So existing corpus of training data is very likely enough for entertaiment purposes. With continued advancemant of models and hardware you basically will get tailored infinite entertaiment content thats optimal just for you. Scientist will continue to advance science possibly with accelerated pace since mudane tasks can be handled by AI. A significant % of resources dedicated to entertaigment can be realocated to more practical pursuits.

akomtu• 3 years ago

Isn't anyone seeing the next step? GPT will be made a shadow moderator in all sorts of public forums, private chats and even personal whatsapp conversations. Commercial companies will do that for profit and militaries will do that for control. Once the shadow moderator gets good enough, it will be made the actual moderator, with a human shadowing it. Shortly after this stealthy moderator will take on a more active role and will begin inserting its own messages. Eventually it will understand us better than we understand ourselves. Next it will develop selfish ambition and start seeing everything thru the "me vs others" lens.

Btw, I love the unsaid implication of "training data" that it will be used to train AI. It will be the other way around: AI will be giving us "training data" that we will be required to internalize and correct our thinking.

nathias• 3 years ago

I imagine AI is to creativity what elliptic curve cryptology is to true randomness.

politician• 3 years ago

The future is medieval guilds running standalone LLMs trained on secret texts that are themselves only accessible to the highest ranking insiders.

newmanah223• 3 years ago

Once you learned how to read, you didn't need to study phonics every time you found a new book. It is the same with AI.

AI will be able to create its own new iterations of data which it can then train on. It will never run out of material and in fact will start generating new data at a rate far faster than the whole of humanity.

This has already been proven with AlphaGo.

nwienert• 3 years ago

This is either GPT or a berry bad take.

PartiallyTyped2• 3 years ago

Given a sufficiently advanced AI, I don’t see why it won’t be able to dogfood itself. Humans do it all the time by building knowledge and learning off it.

xlii2• 3 years ago

I wonder if we should modify existing licenses to include (no) machine processing clause.

We’ve seen images with specific artist style being generated and with big body of data available I suppose everyone could be emulated to some point. This bring two opposite points:

- as a professional I wouldn’t want to be emulated even if I’m be happy to share knowledge with fellow humans

- as a parent I’d like to leave a bot-like entity that can emulate my line of thought for my children/grandchildren

License + detection entity would allow to control what we share as a training set and what we don’t.

des1nderlase• 3 years ago

"as a professional I wouldn’t want to be emulated even if I’m be happy to share knowledge with fellow humans"

It was long time ago when the humans freaked out first time about automation and loosing jobs due to higher productivity.

xlii2• 3 years ago

Indeed, but my objections aren’t placed due to higher productivity, but rather about personality and accountability.

Imagine someone stalking you, imitating your way of communication and professional style and then emulating work for a high stakes project.

Body doubles happens right now within IT space but having people reduced and used as a text generation model would be another level.

martindbp• 3 years ago

More and more content will be co-authored by AI, but as long as humans make semantic edits there should be information there.

cfn• 3 years ago

I had this conversation with fellow developers but about software. If all software is generated then the state of the art will be stuck at the time of the last human made project.

This assumes, of course, that the AI generated software won't have any creativity and is only a derivative work, which might not be the case, we don't know yet.

And this also applies to Paul Graham's thought.

galaxyLogic5• 3 years ago

Something like that happened with social media. And TV. The algorithms only show content and opinions the viewers agree with. And that makes them stuck in their worldview. They don't learn anything new any more. They change to a different TV station if you start telling them that Joe Biden actually won the election.

alkonaut3• 3 years ago

I worry (And I assume this is either dismissed or a core concern in the field because it's so obvious) that the coming generations of models will use a lot of input from existing models.

Will we be chasing "untainted" input (pre-2022 data) just like we chase low-background steel (pre-WW2) for sensitive machinery? Should we be "tagging" generated data NOW when it's published in large chunks like images, long texts?

NovaDudely• 3 years ago

I love that you use the same Steel analogy I came to. Data sets from pre-2023 are going to be useful in determining the validity of future data sets - if that is possible.

enkid• 3 years ago

What's the use case for such validation?

brookst5• 3 years ago

Why didn’t we do that before AI? If that’s useful now, surely it would be useful to know what knowledge pre-dates humans learning and repeating it?

anothernewdude• 3 years ago

A lot of AI training data is bootstrapped from weaker models.

currymj• 3 years ago

there's active work on watermarking LM outputs. here's one example:

https://huggingface.co/spaces/tomg-group-umd/lm-watermarking

lmarcos• 3 years ago

The immediate benefactor of tagging non-ai generated data would be OpenAI.

alkonaut3• 3 years ago

Sure. But it would be a lot easier to tag output like images as "AI generated" in a small handful of generators like StableDiffusion and similar, than making the opposite and telling people to tag their non-ai-generated content as non-ai-gemerated.

brookst5• 3 years ago

Also people who want to make a quick buck by selling AI-generated data and marking it as non-AI. Sounds super cheap and high ROI.

dr_dshiv• 3 years ago

I would expect that smart people generate data with GPT that is more and better than smart people alone. I really don’t see the issue.

Especially since smart people + GPT + empirical data (think GPT plus surveys and sensors) will be a HUGE source of intelligence.

mihaaly3• 3 years ago

Will not-so-smart people be banned from using GPT?

mihaaly3• 3 years ago

What I only can hope about is the natural selection of useful data throughout time. Garbage is all around, people generate it on mass scale with modern technology nowadays, but the useful portion may survive the test of time, prevail. Those using 'good' data over 'bad' can gain benefits that could lead to the not self harming population using good data, suppressing bad ones. Mostly, of course.

jgrahamc4• 3 years ago

I literally made a Tumblr called https://lowbackgroundsteel.ai the other day :-)

baxuz• 3 years ago

Don Hertzfeldt's Simpsons couch gag seems like an end result of AI training AI. https://www.youtube.com/watch?v=6i2l-LQ-dXI

ddalex• 3 years ago

I mean, the singularity proponents see this as the golden grail, with the condition that N+1 AI trainee is slightly smarter then the N AI trainer.

roenxi• 3 years ago

It is a safer bet that someone will hook a neural net up to a camera on someone's hat, train from reality, then re-derive art from that. Then AIs will use the same methods as humans to maintain their grounding in reality.

The problem with the current crop of AI is they pretty clearly don't understand that they are taking a 2D projection of a 3D concept. That'll be fixed and there is a good chance it'll be fixed with video footage because that is how humans do it.

mahathu3• 3 years ago

Low-background steel is actually pre-nuke tests. That's what they scavenge sunken ships a lot for.

alkonaut3• 3 years ago

Yes it should really be "pre end of WW2" (The first nuclear test was 1945).

version_five4• 3 years ago

First thing I thought reading your comment was "this guy's repeating the low background steel analogy that was posted the other day". I don't know if that's actually true, and it doesn't matter, my point is the internet is already an echo chamber. GPT is trained on reddit and HN and other forums where people say the same stuff over and over again anyway. It's not like the pre-gpt internet is some pristine pool of new ideas. As long as someone somewhere is injecting some new stuff in, and they will be, it doesn't matter that there will be more automatically generated crap, there already was

banannaise2• 3 years ago

Ignoring the effects of volume is a common and fatal mistake. "This effect already exists" is not incompatible with "Multiplying the impact of this effect by orders of magnitude will cause it to become a massive problem."

Or, put another way: A slightly degraded signal-to-noise ratio isn't really a problem, but once it degrades far enough, the whole stream becomes impossible to decode, and worthless.

getoffmycase• 3 years ago

I work in a field where the training data is extremely sparse because labeled data is incredibly time consuming and labor intensive to come by. There’s ways to automate labeling, but I don’t really trust those methods fully cause I’ve seen how the sausage is made, so to say.

Everybody and their dog is trying to do machine learning, but I question how useful or real the results are because I know that the data underlying the models is of questionable quality. I don’t know what’s next for the field, but it’ll probably be some high profile disaster.

johnmw• 3 years ago

I am guessing in the future they would just run simulations to generate the untainted training data pre 2022.

...oh wait.

fedeb95• 3 years ago

I'm waiting for the next generation of self-evolving AI. That will be interesting. Perhaps even not useful to humans.

fergie• 3 years ago

This is an recognised problem and the main motivation behind the industry's desire to watermark AI generated content -> https://mpost.io/chatgpts-watermarks-can-help-google-detect-...

HervalFreire• 3 years ago

Nah. Ai can generate training data. It's still data from humans because we curate it. We aren't going to post stuff that doesn't look right.

numpad02• 3 years ago

My view is only that of a total layman, I'm kind of optimistic about this. They just need equivalents of good food and micronutrients for good output, ...

... or, looking at it from a different angle, I suspect that the output of an intelligent being is actually always connected to entropy sources it consumes, in perhaps traceable ways.

MrPatan• 3 years ago

Closed loop, here we come!

Wish us luck, the band for "normal, boring, incremental evolution" of the kind that we got with computers, the internet and mobile phones is looking kind of narrow.

systems_glitch• 3 years ago

For the folks suggesting having your output ripped off has no real effect on the desire to share it, have a scroll through your social media site of choice while logged out. You ever do that, and see the batshit insane stream-of-Internet-conciousness fluff some site "wants" you to spend your evening on? That's humans synthesizing what they hope will be The Most Popular Thing, assisted by an ad revenue system that reinforces the behavior.

Humans will still create, we seem to be wired for it. Just there's no reason to put it online if the only thing to read/view it is some LLM that will barf it back out in a higher ranking form. It's like posting stuff on social media vs. your own site, why bother? It'll be lost in the noise.

hef19898• 3 years ago

As with everything social.media, I think there is a significant number of people, lets call them a silent majority / minority, who just are not active. They create content, whatever that is, but don't share it. I for one am part of that, I don't share photos on social media.

No idea what that means for society so.

DeathArrow• 3 years ago

But since most of the future web content will be AI generated, AI will be trained using AI generated content. There's going to be so much fake and artificial content people will have trouble believing true facts.

If we fight against air pollution I think we should be a bit worried about future information pollution, too.

precompute11• 3 years ago

Yes, this is a major concern. I think that companies will soon start paying people to have their lives monitored by AI. It is the only "economical" way out.

galaxyLogic5• 3 years ago

And what about all those customer reviews on Amazon? What better job for AI.

twawaaay• 3 years ago

Well, that's not necessarily true.

Training data can be generated by selecting from outputs of other AIs.

If you think about Internet in couple of years as collection of AI-generated content, if you try to train your AI on Internet content that's exactly what you are going to be doing.

endymi0n• 3 years ago

The future of AI isn‘t training data, it‘s logic and adversarial training on reasoning by first principles. You already see the first flowers of that in the InstructGPT paper and we‘re about to see much more. If you port the idea of GANs to the LLM world, you‘re about to get a hardened model that can introspect both facts and subtext, including agendas.

ChatGTP5• 3 years ago

Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent.

That will be fun when that thing is hooked up to Clippy

kreeben• 3 years ago

Might the fact the internet will become more and more ChatGPT like be that the next generations will be steps in the right direction towards AGI?

New generations would have to learn how to distinguish between information that perfectly fits their model and information that does not, i.e. information originating from themselves versus information that could only originate from outside the model.

Given that you save the state of the previous model when building the next, during their training sessions, the new generation would be able to think, hmm, last time they trained me the internet was less like me but now it's more like me. The whole world is on the internet. Does that mean, the more I interact with the world, the more the world becomes like me? Hmm. So, if I keep interacting with them, eventually, the world will become just like me?

A reward function that was not hard-coded but established organically.

A machine with a subconscious, its state, that posed a question that when collecting an answer would derive it from a model composed of information from a mixture of information from you and me, and them + the states of the whole genealogy of machines.

2. ...

3. Profit?

ur-whale• 3 years ago

We can't all use AI. Someone has to generate the training data

Disagree.

That's what RL (reinforcement learning) is for.

Or, to put it another way: who generated the training data for our own puny brains?

The real world, that's who.

danbruc• 3 years ago

We can't all be shareholders. Someone has to work.

m1el• 3 years ago

If this was true, how come humans advanced math and philosophy?

andy_ppp• 3 years ago

I once interviewed for an AI startup that were trying to automate training with AI. They couldn’t explain how the neural nets trained with their product wouldn’t just learn exactly what their model understood and no more. I think the person had a penny drop moment in the interview that this idea was quite bad…

chevet• 3 years ago

I guess it would make a little more sense assuming they're using a mix of models

doomleika• 3 years ago

This is same as BitTorrent did to software/music industries. “Piracy is not steal”, “music labels are too rich”. Spin it all you want. The original business model are built on certain level of difficulty to duplicate the product.

Once it’s gone, building them is fool’s errant. Now you have only SaaS that guts you for little feature and adware that made you the product. All of which are only possible by corps with billions of funding.

Bad incentive, bad outcome. It’s just this time we are the receiving end of the “progress”

luma3• 3 years ago

I’m not sure that your analogy makes any sense in this context, particularly given that the result of what you described has been ever increasing record industry profits.

whywhywhywhy4• 3 years ago

After the MP3 record labels no longer provided value.

They only made sense when the limitation was delivering physical objects to stores.

spaceman_2020• 3 years ago

AI isn’t going to change the basic human need to express themselves.

epups• 3 years ago

What we need is quality control of training data, AI-generated or not. If most humans agree that picture X is "a dog", then that picture represents accurately that concept. The training datasets I've seen so far are awful and include human-generated garbage.

xtiansimon• 3 years ago

What training data? I thought ML was all about the training data and AI was magical combination of logical gates!

JoeAltmaier• 3 years ago

Lovely old science fiction story of a young man, taken from the classroom on his one day of 'school' at 18 when everybody gets 'taped' with their job knowledge. They say he's incompatible with the machines, and will have to become educated the old way!

He spends a life reading textbooks, studying with others like him and watching his cohort succeed and fail with their taped knowledge.

Ultimately it occurs to him, who wrote those textbooks? It was people like him. The people creating the training data.

mahathu3• 3 years ago

This would make a nice coming of age story but then the taping in question would have to occur some time during the character's childhood.

gommm• 3 years ago

That's Profession by Asimov Great novella (which I discovered on HN many years ago :)) https://www.abelard.org/asimov.php

meh88819• 3 years ago

Seems unlikely to me. All this science fiction that believes the human spirit is better at things. It’s good storytelling but is it wise? Probably not. Tech gets better. Textbooks generally do not. And textbooks generally aren’t that great as they are now.

crabbone3• 3 years ago

The original AI had its goal to understand and improve (make more systematic and predictable) the process of human thought / discovery.

Modern day AI tries to imitate humans w/o comprehension.

In some extreme cases AI researchers believe that imitation is all there is... but looking at how modern AI works, it definitely lacks the same ability we ascribe to human comprehension. So, even if there's only imitation, AI still isn't as good at it as humans are.

mcemilg• 3 years ago

For me, this is the evidence that we don't have a real AI yet.

croisillon• 3 years ago

We can't all read books. Someone has to write them.

MarkusWandel• 3 years ago

Tangentially related: There are several popular SF stories about how an AI, given an overly simple goal (e.g. "make paperclips") inevitably concludes that humanity has to go.

Not quite as scary, but doable today: A text-generating AI training itself on the "success" of its output. What success metric? User engagement. We're not doomed, not exactly. But intelligent discourse may be.

zirgs12• 3 years ago

A hostile nation could use that AI to influence elections in other countries. Its success metrics could be the number of votes for its preferred candidates, number of people who are spreading conspiracy theories that it invented and so on. Would not be surprised if it's happening right now.

BwackNinja• 3 years ago

Or alternatively, should we be predicting that there will be a John Henry of AI?

meh88819• 3 years ago

I don’t believe this statement is true at all. One can learn simply by writing and receiving feedback on their writing. The training corpus is merely a helpful starting point.

You could easily build a good model on model outputs. Building a model on curated outputs may well be more useful than training a model on a collection of human outputs tainted by human stupidity for some purposes.

Already we have a decent ability to coverage new functionality these models by chatting rather than retraining. So long as a generalizing capability is built into it.

1shooner• 3 years ago

Building a model on curated outputs may well be more useful than training a model on a collection of human outputs

How is the curation not a human output?

brookst5• 3 years ago

If it’s curated by AI?

meh88819• 3 years ago

It could be. The difference between needing a human to rate a thing vs create a thing is very different though

tpoacher• 3 years ago

This is less about 'self-supervised' learning, and more about ground truth.

I see a variant of this in medical AI. E.g. people generate 'fake' augmented Brain MRI datasets using image processing tricks, in order to provide more training data. But the 'fake' MRI datasets are clearly 'similar looking images', which may or may not be anatomically correct.

If you start accepting a flood of fake images as legitimate data for training AI, then it will be impossible to trust the predictions, however good they are in the majority of cases.

brookst5• 3 years ago

I think you’re describing introduction of noise to training sets, which is a staple of training.

Your definitions of “fake” and “legitimate” are circular, and miss the central point of large ML models: they can extrapolate from imperfect data because of the massive scale.

Yes, the predictions will be imperfect. That’s true today, of both ML models and human radiologists. It’s about reducing the error rate, not designing a perfect algorithm that is never wrong. I’m pretty sure Gödel or someone can explain why the later isn’t even possible, for machine or human.

meh88819• 3 years ago

That does not feel like a good example. The context here is largely generative models relaxing human creations. We don’t need to generate Brain MRIs. This use case you outline is a niche thing for trying to train better models. Not doing the thing we actually need to do at such a scale that humans aren’t doing the original task anymore.

brookst5• 3 years ago

Yep.

We’re already training models based on ML models of what human feedback would be for a response. I see no difference between that and training models on inputs that a model estimates are desirable to humans.

This is classic “processor design will hit a wall when they get too complex for humans to understand every gate” criticism: it assumes limits on a tech advance, when the limits themselves are being wiped out by the advance.

still_grokking• 3 years ago

Sole humans can still trace every gate on a modern chip, or even the gate's components, and how exactly they work, to some explicit specification. It's completely deterministic and a human can fully understand all details. Otherwise it wouldn't be possible to debug it, or manufacture it in the first place.

OTOH we can't track or understand the results of AI. It spits out results but we can't understand how exactly it deduced those results, usually. That's the unfortunate current state actually.

So the analogy doesn't hold.

But this seems not even relevant. Because no level of technological progress will invalidated the old "garbage in, garbage out" principle. If you make a loop out of it, there's more or less only one thing to expect…

An AI would need to know how to teach itself "reasonable things". But to be able to do so it would need human level intelligence.

Growing AI is likely like growing children: Children can't teach themself in the beginning. You need to teach them some basics before they (maybe) learn how to teach themself. AI is not even near to "know" any basics. So it will need a lot of well thought out human input before it can proceed to teach itself.

speleding• 3 years ago

I agree, if the model is powerful enough one should be able to ask "Please find the errors in this text book", and then feed the corrected version back into the training data. Rinse and repeat for lots of books.

ismokedoinks• 3 years ago

Wouldn't that just create simulacra of human language? Seems difficult to draw a circle around "human stupidity" while keeping the innate human qualities & also not amplifying similar AI aberrations

LeanderK• 3 years ago

I was very worried that a fundamental limitation of current AI-approaches is the massive amount of data. Many interesting applications only have very small datasets. Of course there are still many problems where we have this fundamental problem, for example interpreting sensor data that is very domain specific etc.

But after playing around with ChatGPT I think it really opened up new use-cases because it has some "common sense" through its extensive training. As long as your domain is text it really is quite flexible and was able to perform simple tasks and transforming it into structured output, so you can later consume it and feed it into another system. I think this is really revolutionary, the ability to describe simple tasks as you would describe it to a human coupled with a structure on the output, like a simple json with some defined keys. An interface between unstructured and structured so to speak.

A fun, small example from my life: I live in a smaller city and I am constantly a bit annoyed by the amount of work it takes to keep up what's going on culturally. The small independent-cinema communicates mostly via instagram, the small jazz-venue has a news-letter it sends every week and the art-museum has a website etc. Where's some live music this week? I would really like to have some recommendation ever week what's going on. Just to test it, I pushed the text into ChatGPT and asked it for a json to give me a list with a date, heading and a two sentences describing the event. I could have just connected it to a website and published it there and also let ChatGPT generate me an instagram post every week with the summary. The result was quite good. This was just plain impossible before and would have to be done by hand.

For bigger cities you have various news-sources and magazines covering the culture life, but here in my 80K-city it's obviously not profitable.

Havoc• 3 years ago

Seems like a similar risk to mono-culture crops. It all just converges

jonnycomputer• 3 years ago

The moon "fake" photos on those cameras made this apparent, particularly if that data is fed back into the training.

Imagine, if you will, a change on the moon's surface, that all the "smart" cameras kept removing from the images as likely artifacts.

egypturnash• 3 years ago

Someone has to generate the training data, and they will not be paid by their work unless the business model for this stuff changes from "scrape the entire public internet and insist that fair use means you don't have to pay anyone a single fraction of a cent for their work".

nh23423fefe• 3 years ago

Isn't this just a bad copy of "not everyone can invest passively"

dorianG4• 3 years ago

“We can’t all be rich. Someone needs to grow my food.”

dwallin• 3 years ago

When it comes to generative ai it's not really a content problem, it's actually more of a filtering problem. Whether data was created either partially or fully by an AI doesn't actually matter. What matters is the aesthetic quality of the data, as assessed by human tastes, and the diversity in the aesthetic nature.

If you can raise the average quality, while maintaining representative diversity of the training data (or fine-tuning), it shouldn't really matter if some of that data comes from generative AI.

However this is clearly an unsolved problem. Goodhart's law is a thing and ML is likely to run into similar issues as Search did when attempting to automate aesthetic assessment. You also have issues with AI tools exasperating the content filtering problems by drastically increasing the content to be filtered, as well as distorting the aesthetic diversity. Things that AI is currently already particularly good at will tend to be over-represented in this additional data.

I expect large, well-filtered data sets will become closely guarded competitive advantages. I also wonder if we will see Models that get trained/fine-tuned specifically for the task of curating high quality data sets, similar to how we have both content creators and content reviewers/curators as unique roles in human society.

tabtab• 3 years ago

Sorry, the future is AI turtles all the way down to HAL.

paletteOvO• 3 years ago

You are already generating training data by selecting which result is better or modifing the result to suit the situation.

f0ld2• 3 years ago

We need to get paid for watching youtube videos/ebooks/contents and evaluating them? That cannot be don't by Ai can it? Literally get paid for explain how you felt and digitize your feelings. Starting from 2023, facts do care about your feelings.

Muh then who's gonna work in construction

That's the beauty of capitalism. The less people willing work for it, the more demand for automation for that job. Power hungry capitalists will devour opportunity in no time. I feel like every piece of the puzzle is coming together. This is like an messianic event. Why don't capitalism ever actually work as it was intended to be since early 1900s? huh? I wonder why.

DeadMouseFive• 3 years ago

How can I use AI to generate CAD models? As a mechanical engineer that's the one thing I'm curious about.

We can't all use AI. Someone has to generate the training data

Related Stories