I just want to express how grateful I am that Zuck and Yann and the rest of the Meta team have adopted an open approach and are sharing the model weights, the tokenizer, information about the training data, etc. They, more than anyone else, are responsible for the explosion of open research and improvement that has happened with things like llama.cpp that now allow you to run quite decent models locally on consumer h…
AI safety risk is substantial. It is also testable. (There are prediction markets on it, for example.) Of course, some companies may latch onto various valid arguments for insincere reasons.
I'd challenge everyone to closely compare ideas such as "open source software is better" versus "state of the art trained AI models are better developed in the open". The exact same arguments do NOT work for both.
It is one thing to publish papers about e.g. transformers. It is another thing to publish the weights of something like GPT 3.5+; it might theoretically be a matter of degree, but that matter of degree makes a real difference, if only in terms of time. Time matters because it gives people and society some time to respond.
Software security reports are often made privately or embargoed. Why? We want to give people and companies time to defend their systems.
Now consider this thought-experiment: assume LLMs (and their hybrid derivatives) enable perhaps 1,000,000 new kinds of cyberattacks, 1,000 new bioweapon attacks, and so on. Are there are a correspondingly large number of defensive benefits? This is the crux of the question I think. First, I don't expect we're going to get a good assessment of the overall "balance". Second, any claims of "balance" are beside the point, because these attacks and defenses don't simply cancel each other out. The distribution of the AI-fueled capability advance will probably ratchet up risk and instability.
Open source software's benefits stem from the assumption that bugs get shallower with more eyes. More eyes means that the open source product gets stronger defensively.
With LLMs that publish their weights, both the research and the implementations is out; you can't get guardrails. The closest analogue to an "OSS security report" would take the form of "I just got your LLM to design a novel biological weapon. Do you think you can use it to design an antidote?"
A systematic risk-averse person might want to ask: what happens if we enumerate all offensive vs defensive technological shifts? Should we reasonably believe that the benefits outweigh the risks?
Unfortunately, the companies making these decisions aren't bearing the risks. This huge externality both pisses me off and scares the shit out of me.