Hacker News Reader

Comment by monkpit | Hacker News Reader

Is there any way we can prevent these models from being abused by large corporations for profit?

I was just thinking today, it’s interesting how I can ask ChatGPT something that is hard to get a straight answer for - such as “how much does X usually cost?”

I wouldn’t be surprised if this type of tech was quickly censored by a company that implements it at scale (say, Google or bing) for profit.

I feel like it’s inevitable, and it’s depressing.

Will the future be populated by LLMs that all want to control the information I have access to?

View on HN

I want an open version of this that I can train on any data I choose, can give any prompt I want, and runs on my hardware.

However, I feel like the best and most useful versions of this type of software would (by necessity) only be able to run at a huge scale, cost millions to train, etc.

Now instead of having to worry about the bias of the sources I read, I have to question the bias of the language model based upon who created it.

How do we even define 'best and most useful' for a language model?

If you have your own data and problems you are trying to solve I don't think using GPT4 in zero-few shot mode (even with CoT) is necessarily better/higher performing than using a smaller fine-tuned or retrieval augmented model, it certainly wasn't true for GPT3.5 although we don't have as much insight into GPT4.

If we're talking about something as general/open-ended as ChatGPT or a 'poor mans AGI' well yeah, the data curation and preprocessing alone takes a small team to figure out let alone training it. How useful this type of model is when it's still not good enough (i.e. hallucination) to be used in most production cases despite the exorbitant cost in development remains to be seen.

I would also like to avoid a future where we are slaves to corporate AI. If training the whole model individually isn’t feasible, would it be possible to decentralize the training?

That’s an interesting idea. Something like Folding at home, or coin mining.

Check out KoboldAI community, they're collaboratively finetuning and hosting GPT-J, NeoX, and now LLAMA models for free, as hobbies.

Sounds like the same problem we have with information sources in general. Whoever controls the source will bias what comes out to its advantage.

The good thing is that open source versions are popping up everywhere. This wont stop and everyone will be able to run their local AIs https://simonwillison.net/2023/Mar/11/llama/. Problem is the 'best' one will likely be centralized. We have open source everything; nowadays you can run on your own version of most software categories but users still choose to use proprietary versions bc the convenience outweighs their need for privacy/objectivity.

Are you implying that any company using AI for the purpose of increasing the profitability of the company is "abusing" AI? Or are you talking about something else? It kind of across to me like you thought the use of AI with profit as a motive is always abuse, not sure that I understood you correctly.

If it’s not disclosed then it’s abuse (in my eyes - not saying legally, just morally).

Do companies abuse linear regressions?

If you think it is always abuse if the use of AI is not explicitly disclosed, then you're going to have to recognize there are many thousands of situations where that view is absurd.

Imagine you and a group of friends decide to develop a game. You make a company and have a team of four people. Your budget is tiny and you can't afford the same resources a huge game company can. You want your new company to be profitable so you use large language models to generate realistic material textures for your 3D models, and to generate a diverse range of intetestimg character backstories and dialogs. This used to require a larger team of developers and content creators but now you can do more with fewer people and ship your game sooner. You make a great game and users love it and it's a financial success.

But oh wait, you didn't explicitly disclose that you use an AI powered tool to generate some art assets and character backstories. Therefore you are just a greedy company that is abusing AI for profit.

That's so ridiculous.

> Is there any way we can prevent these models from being abused by large corporations for profit?

Competition, from for-profits large and small, non-profits large and small, and the rest of the open community.

Don't trust what BigLLM™ is saying? Here's a dozen others you can ask...

competition is what got us the huge corps..?

maybe look for opportunities for cooperation instead.

Open source competes through cooperation.

could that not simply be stated "open source cooperates"

i find it a little dated that we cognitively frame everything through the myopic lens of competition.

and it's a perfect example where a systemic restructuring can fundamentally rebalance control.

Why is it myopic? OSS products and services absolutely do compete against other commercial services, but both improve in the process. It is the same as how biological evolution works. Competition is a good thing.

you are assuming competition as the dominant mode a-priori. i'm suggesting it's not the case. and the congnitive re-framing of these interactions can have a net-positive effect on social relations.

I'm really not sure what you're saying. All I'm saying is that OSS also has to compete against commercial software.

open source doesn't have to compete.

Of course not, sure. But it depends on what you're targeting. If you want to dump code into a git repo, slap MIT on the license and never touch it again, you can, I've done so many times. But if you want people to use it, or if you want to earn money from it (Docker, GitLab, Plausible etc) then you will have to compete.

> competition is what got us the huge corps..?

Yes, and consolidation also. Competition also gave us their disruptors.

> maybe look for opportunities for cooperation instead

I hear you, and I don't disagree with your line of thinking, but cooperation to compete is a thing, no?

My point being, an adversarial strategy prevents power from coalescing into the hands of one group, because it actively seeks to disrupt that process. The more diverse competition there is, the more pressure there will be against huge corps.

> Is there any way we can prevent these models from being abused by large corporations for profit?

No. I mean what, are you gonna get them to halt development? I'd also say heavy-handed regulation is unlikely because the US Government views itself as in a technological competition with China, it won't want to slow things down.

> Will the future be populated by LLMs that all want to control the information I have access to?

Nothing's stopping the open source community, or you yourself, from making your own. You have a computer connected to the internet after all. But it will require lots of technical expertise, data, and money.

I suppose regulation is part of what I had in mind, but it would be really tough to prove anything in court (I feel) so regulation might be impossible.

Maybe we require transparency in how the model is trained, what the prompt is, etc. It’s fine if a corporation wants to build something that increases their profits, and they can invest into training the model and use it as a competitive advantage.

But I feel there is some level of disclosure owed to the consumer. Not sure what it would be.

Big companies will have better technology for a long while because of all the money they can throw at it, but there's no reason to expect an ecosystem of useful open/instanced AI systems to mature along side it. We've seen that start with Stable Diffusion and now also with LLaMa, and we're only at the beginning of the road.

Yes, big companies that spend lots of money and are big targets for lawsuits and regulation will exercise a lot of control over what their state-of-the-art products offer. There's no getting around that.

But that's not all we have to look forward to by any means.

we need a tech consumer+employee co-op.

fedco is a seed company structured similarly. it's owned 60% by consumers and 40% by employees. this puts democratic control and profit distribution front and center.

> We are a cooperative, one of the few seed companies so organized in the United States. Because we do not have an individual owner or beneficiary, profit is not our primary goal. Consumers own 60% of the cooperative and worker members 40%. Consumer and worker members share proportionately in the cooperative’s profits through our annual patronage dividends.

https://www.fedcoseeds.com/about_fedco.htm

Reply on news.ycombinator.com