Comment by idle_zealot | Hacker News Reader

Comment by idle_zealot | Hacker News Reader

Parent Full thread

The advantage that open AI (not the company) has is that if using copyrighted content as training data without licensing it is found to be illegal, they can just keep doing it. There's plenty of FOSS software basically designed to violate copyright law (comic readers, home media center servers/clients, torrent clients) that big tech cannot compete with lest they face legal consequences. Basically what I'm saying is that the open source community will continue to use books3 and scraped images to train while facebook and the like get stuck in legal quagmire.

Of course, this ignores the fact that popular "open" models of today were actually trained with facebook or other companies' computational resources, so unless some cheaper way to train models were developed we would actually be stuck with proprietary models trained with lots of compute but unable to use unlicensed training data, and open models that can use whatever data they like but must operate in the shadows without access to much compute for training.

As long as the cost for training a model is in the 7+ figures, that means that any such open model is bound to be tracked to someone with deep enough pockets to sue.

Consider that you just spend a few millions on training a model on copyrighted data. Release it would reveal that, problem.

I _guess_ you can try doing training in the public, like Seti @ Home or something like that, which distributes the risk? But no idea if this is even possible in this context.

Is the training cost equally that high if you do adversarial training?

But if we have 100,000 people with RTX 4090s mining some new AI based cryptocurrency, that happens to train the model in the process, it's going to be a highly effective system. And we can do it anonymously over Tor or I2P.

You really can't, because bandwidth matters as much as compute power - you can only utilize as much power as you can transfer data to/from.

The training methods for current LLMs are parallellizable only by very frequently transferring all the data back and forth, and needing a node to gather and merge all the updates very frequently, and redistribute it to every other node so that they can make any progress. And a GPU that's not connected you with a high-speed link is pretty much useless as you can't make useful progress until you get their part pack, and "their part" is very large (i.e. the update size you need to get back is comparable to all of the model size) and you need to do that very frequently. Training on nVidia many-GPU pods works because of high-speed interconnect (e.g. 600 gigabytes per second for 8 GPUS in A100 pod), and if your internet bidirectional speed is much less than 600gbps, then if you have 100000 free remote RTX 4090s, you simply can do the compute locally faster than you can exchange information with the other GPUs.

"This paper presents a distributed model-parallel training framework that enables training large neural networks on small CPU clusters with low Internet bandwidth." Low bandwidth being <1Gbps. They've also tested it with GPUs as well.

https://arxiv.org/pdf/2201.12667

Maybe there's the possibility of a completely new AI architecture that can still be efficiently trained when there are very low bandwidth connections between nodes? Specifically targeting this use case would make sense, given all the millions of underutilized GPUs out there in peoples' desktop computers.

Also https://arxiv.org/pdf/2106.10207 ?

The cost of modifying open-ish source models with copyrighted data is much lower.

For example, the cost of modifying Mixtral with uncensored data is currently about $1200 [1]

[1] https://youtu.be/GyllRd2E6fg?si=SJmPLsPlCRRT0uPV&t=236

> I _guess_ you can try doing training in the public, like Seti @ Home or something like that, which distributes the risk? But no idea if this is even possible in this context.

Let's say that's a field of active research. Right now you need very low latency, to the point that people connect training clusters via infiniband instead of ethernet despite the computers being meters apart. But approaches that tolerate internet-level latency are being developed

I’m willing to bet that there are projects at every chip company looking into how to drive down these costs. The first step is to put the architecture of the model and the backprop into silicon. The weights are the only variables so if you can reduce those to cheap and fast modules that can be mass produced you could come up with something cheaper than a GPU. If the profitability of running and training this specific model is high enough than its worth the investment of time and money to set up the production line.

The costs are falling 10-100x every few months. If you told people in 2022 that they could run 70b paramater models at 2 bit quantization on your home 4090, they'd have laughed in your face. They're not laughing anymore. SFT/DPO are ultra efficient compared to RLHF, and innovation in this space is only just now getting started.

You're right that there are areas like torrenting where big tech can't really go for fear of legal consequences, but there's also the cat-and-mouse game played by IP holders against, say, torrent trackers, which leads me to think that

> open models that can use whatever data they like but must operate in the shadows without access to much compute for training

sounds pretty analogous to private trackers, where high-quality stuff is available but not in full public view. If rightsholders and big tech, abetted by states, crack down on open models I think you're right that the open source community will continue to train on liberated content, but it's not going to be as open and free-flowing as things are now. Going back to the compute problem, I can imagine analogues of private trackers where contributors, not wanting to expose themselves to whatever the law-firm letter/ISP strike analogue will be in this space, use invite-only closed networks to pool compute resources for training on encumbered content.

> There's plenty of FOSS software basically designed to violate copyright law (comic readers, home media center servers/clients, torrent clients)

The key difference is that those projects don't violate copyright themselves, but facilitate users doing so. If training without a license is infringement then projects that are doing so will struggle to host code/models publicly or access other parts of internet infrastructure.

>struggle to host code/models publicly or access other parts of internet infrastructure.

To be fair, IPFS/Torrents/Usenet exists.

Yes, there are workarounds that are already used in other areas of copyright infringement. It won't stop the activity completely but will curtail the scale of that activity.

That may be partly true but are small open source groups going to have the financing to do AI training that can compete with the big techs? I think the most likely scenario is for some other country with a less strict view of copyright (ahem) to continue developing AI without caring for the input data.

Reply on news.ycombinator.com