Greg Notch had just said a frontier-class open-weight model could be run at home for roughly the price of a car. He explains why that matters for security: models run locally are unmonitored, and their safety controls can be stripped out.
I think that price is plummeting over time. Right now, if you want a frontier Opus class to use the
…as long as you have a free dryer plug. I mean, if you want to run a smaller model, so some of the quantized models, you can build a completely reasonable rig for about $20,000. And I think those prices are just going to keep plummeting, especially as GPUs are cycled out of big data centers. There's
effectively no moat around that technology and the price. Is going to drop. That's the thing where it should give everyone pause because there's no rules about what you run on your hardware. No one's watching. And so all of the safety guards can be easily removed.
What's a hypothetical cyber attack where an unrestricted AI model could change what the attacker could accomplish?
A good example is the reverse. There's a new repo of reverse engineering software that's making the rounds now called Rhea, and it's quite excellent. But if you run it and you attempt to say, take apart something and remove its cryptography, or you take a look at the firmware for a medical device…
version of it is you're able to accomplish reverse engineering. You're able to build exploits sort of away from prying eyes and with no restrictions. And everything that you do will be novel when you weaponize it. So I found a zero day in a firewall, or I found a zero day in a medical device, or…