Statement on US government directive to suspend access to Fable 5 and Mythos 5

First I want to see them play video games at a high skill level, preferably without any access to game state beyond the same visual output that humans have access to, like a raster frame X number of times per second. One LLM model played Factorio, albeit at a very, very poor level, which can be seen if you slow the video to 0.25 playback speed and pause frequently.

https://old.reddit.com/r/factorio/comments/1u1blr6/claude_fa...

There have been streams of other games, where LLMs and AIs have likewise performed very poorly.

I recognize that LLMs might be better at language processing than these sorts of tasks. But being able to play video games is part of general capability. And this kind of hardcore video game playing, with no access to game state, is also a general task where feigning skill can be harder. If LLMs excel at pretending to be competent without actually being competent, like this AI training approach is arguably about

https://en.wikipedia.org/wiki/Generative_adversarial_network

Then some AIs might be trained and designed for deceiving humans instead of actually being competent and capable. And thus, one response is that they should be met with more difficult tests.

Basically, make tests that AIs or LLMs will not have an easy time cheating. Hopefully, that will engender research in greater LLM/AI competence, not in greater ability to cheat or deceive, neither for LLM/AI researchers and companies, nor for LLMs/AIs themselves.

That's separate from the question of whether Fable 5 and Mythos 5 are unsafe. I don't really know. Here's a few things that seem real, though: These models probably have some level of capability to assist with bioterrorism, Anthropic has self-admitted that their own safety measures are imperfect [1], so it should come as no surprise that jailbreaks seem far more possible than Anthropic is leading you to believe in this blog post [2].

[1] https://www.anthropic.com/news/fable-mythos-access: "We suspect that perfect jailbreak resistance is not currently possible for any model provider."

[2] https://x.com/elder_plinius/status/2064776322979676227

If Amazon sold a book that taught someone how to commit bioterrorism, would there be action against them to stop selling it? Its an imperfect analogy, but the parallels are there. LLMs don't get a free pass because they're also so good at writing typescript for beige CRUD apps and bedtime stories.

One thing I hope we align on: Synthetic safeguards (steering, rejections, etc) on top of models to block illegal/sensitive topics isn't good enough. Anthropic has self-admitted that it isn't good enough. We need the technology to lobotomize these capabilities the public deems too unsafe to allow out of the models at the most fundamental level. And, we need to align on what the scope of these forbidden fruit topics are. This is, actually, the only way open source continues to thrive. I want open source models to thrive, but they won't be allowed to thrive, nor should we want them to thrive, if they're teaching people how to engineer novel viruses and other horrible stuff.

Found the time traveller.