Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

What we still don’t know about how A.I. is trained

newyorker.com

Post reply on HN
Print view View original

191–200 of 211 posts

Re: What we still don’t know about how A.I. is trained

#191

Post by imranq »

bob1029 wrote: post #26

I feel like there is an emerging consensus that [Chat]GPT 3.5/4 is not just 1 big model. A large part of the magic in the final product appears to be many intermediate layers of classification that select the appropriate LLM/method to query. The cheaper models (e.g. Ada/Babbage) could be used for this purpose. Think about why offensive ChatGPT prompts are rejected so quickly compared to legitimate asks for code. Imag…

I suspect they must be selecting different models based on the question asked. I've noticed this myself asking simple questions vs difficult questions, chatGPT is much slower on the difficult questions even though each token has the same computation cost if given to the same model

Re: What we still don’t know about how A.I. is trained

#192

Post by bigfishrunning »

drowsspa wrote: post #153
bigfishrunning wrote: post #81

Earlier quoted context omitted.

yes, but for you all of that text is associated with ideas. The word "dog" has an associated object. For a machine like GPT-4, the word "dog" has no meaning or object, but it does have an associated likelihood for adjacent words. The words themselves aren't the intelligence, the ideas behind them are.

I feel like in a few months this human exceptionalism will be proven wrong by construction.

Perhaps, but i think there's just as much hubris in saying "We've created something smarter then us" as there is in saying "We're smarter then the thing we've created".

meanwhile, we've created the worlds most complicated set of dominos, and we're delighting in knocking them over.

Re: What we still don’t know about how A.I. is trained

#193

Post by jimmySixDOF »

2OEH8eoCRo0 wrote: post #147

> GPT-4’s predecessor, GPT-3, was trained on forty-five terabytes of text data Is this correct? I've seen varying reports of the training set size.

Dr Alan D. Thompson has the best tracking of this training set data I am aware of also a 1hr YouTube on the same subject.

https://lifearchitect.ai/whats-in-my-ai/

Re: What we still don’t know about how A.I. is trained

#194

Post by standardly »

kossTKR wrote: post #69
matheusmoreira wrote: post #16

Earlier quoted context omitted.

Just say what's on your mind and don't mind the votes. One thing you'll discover is that you're not alone in your views, whatever they are. Few days ago I came across this bone chilling AI generated Metal Gear Solid 2 meme with Hideo Kojima characters talking about how the purpose of this technology is to make it impossible to tell what's real or fake, leading directly to regulation of information networks with ident…

Yes, this is the real dark future right here. It will be a continuation though. The uppermost echelons of society have waged information warfare since the dawn of modern PR in the beginning of the 20'th century. Lots of theory on this have apparently been memoryholed, but it's easy to just start with the genealogy around Edward Bernais and the plutocracy and robber baron families still existing in the interwar period…

From manufacturing consent to manufacturing reality

Re: What we still don’t know about how A.I. is trained

#195

Post by rvba »

beaned wrote: post #116
rvba wrote: post #112

Earlier quoted context omitted.

I wonder how much better the labeling of everything (not just the bad stuff) would be if it wasnt outsourced to the lowest bidder

Is there a specific capacity in regards to labeling that is enabled by more money? I can see it for like.. heart surgery. But don't most of us know what things are to be called when we see them? ChatGPT seems to be pretty good at knowing what bad stuff is.

I assume that it would lead to higher quality labels - > higher quality output.

Analogy to the "Garbage in, garbage out" rule.

Re: What we still don’t know about how A.I. is trained

#196

Post by 1vuio0pswjnm7 »

1vuio0pswjnm7 wrote: post #78

"To avoid this problem, according to Time, OpenAI engaged an outsourcing company that hired contractors in Kenya to label vile, offensive, and potentially illegal material that would then be included in the training data so that the company could create a tool to detect toxic information before it could reach the user. Time reported that some of the material "described situations in graphic detail like child sexual a…

This passage comes straight from the OP.

On the positive side, 21 up 0 down. Readers appear to see the issue.

On the negative side, several "tone deaf" commenters that appear to have a blind spot for human suffering. This is not like a clinical trial of a therapeutic. It is not even like animal testing of a consumer product. OpenAI knows the material is harmful.

So-called "tech" companies like Facebook and Google are engaging in this sort of practice every day, paying people some embarassingly low wage, usually through contractors, to subject themselves to psychologically harmful content so the company can sell more online advertising services.

Re: What we still don’t know about how A.I. is trained

#197

Post by precompute »

bob1029 wrote: post #176
precompute wrote: post #149

Earlier quoted context omitted.

I agree. But I think do think they use a tiered approach: they have an uncensored model at the very bottom, to which the queries go to first. And then there's the "clean" model, and (for lack of a better explanation) the two outputs are then "XOR"d to get the clean public version.

I agree the core models are uncensored (as this would introduce a lot of noise), but I don't think any queries are allowed to touch them before a cheaper model first performs screening. If someone is spamming toilet humor, there is no reason to exercise the full 175B+ parameter model each time. The only reason ChatGPT is affordable is because of caching, filtering bad prompts, etc. The underlying LLM would be far too…

Agree 100%. Have you noticed any lag with more "difficult" queries?

Re: What we still don’t know about how A.I. is trained

#198

Post by bob1029 »

precompute wrote: post #197
bob1029 wrote: post #176

Earlier quoted context omitted.

I agree the core models are uncensored (as this would introduce a lot of noise), but I don't think any queries are allowed to touch them before a cheaper model first performs screening. If someone is spamming toilet humor, there is no reason to exercise the full 175B+ parameter model each time. The only reason ChatGPT is affordable is because of caching, filtering bad prompts, etc. The underlying LLM would be far too…

Agree 100%. Have you noticed any lag with more "difficult" queries?

Yes. I think there are multiple levels of LLM for each use case now.

The stuff that absolutely must run on 100B+ models can be pre-classified by something like Ada/Babbage. Attempts at arithmetic or dimensional analysis can be binned and shipped to a monster model. Anything that is more routine information retrieval needs maybe goes to a 7B parameter model.

Lots of models with lots of temperature levels/hyperparameters/etc is the only way to achieve the kinds of performance we are seeing. The secret sauce is looking increasingly like a huge classification layer.

Nothing you see in ChatGPT is as it appears. Lengthy conversations are managed with recursive summarization. Every conversational turn is potentially handled by a different LLM.

Re: What we still don’t know about how A.I. is trained

#199

Post by scottyah »

kossTKR wrote: post #125
scottyah wrote: post #107

Earlier quoted context omitted.

Who are the "microscopic elite that controls the media"? How many people are in the "uppermost echelons of society" to where they try to influence other people? From my middle-of-the-road perspective, everyone is trying to change how everyone else thinks, from the small insignificant details to a cult-like brainwash. Even here and now both you and I are trying to change each others and everyone who reads this's mind.…

From a european perspective even asking the question "why should be scared of these so called elites" is so bizarre it's almost frightening, i'm sorry. It's a testament to the absurd amount of philanthropic whitewashing, PR and media control these billionaries hold. "Elites" have conspired to exploit the masses throughout 5000 years of civilisation, it's simply a fact of history. It's almost physically impossible to…

Assuming Kings/Feudal Lords can be considered elites, we were taught that these systems formed as a means of protection and overall ability to not worry about a lot of aspects of life- raiders, setting up trade routes, negotiating prices, farming, etc.

I see a direct mirroring in today's corporations. If you join a big company, it's because you don't want to risk branching out on your own. You exchange lots of potential money for a steady paycheck, and you don't have to worry about things like finding customers, figuring out taxes, etc.

It's clear that humans need some sort of hierarchy, and I just don't see why we should be frightened of the people whose skill is organization/mediation between people. I surely don't want to play the power game with them, and I don't think it's because of their brainwashing?

Re: What we still don’t know about how A.I. is trained

#200

Post by scottyah »

noelherrick wrote: post #184
liamconnell wrote: post #173

Earlier quoted context omitted.

If you study history you’ll notice that groups and their leaders are rising and falling, conquering and pillaging and then losing it all. The world is too dynamic for your reductive theory to fit in. Elites compete with each other. They don’t sing kumbaya and cooperatively share the keys to power.

While they will absolutely backstab each other given an opportunity (see the VCs who caused a bank run that primarily affected other VCs), they do seem pretty chummy. Davos is quite literally the summit of the elites.

But is the summit "let's all get together and stomp on the poor while holding hands", or is it just another high stakes poker game to show off who's winning or losing, making alliances and seeing who talks to who?
Post reply on HN