Hacker News Reader

Comment by morkalork | Hacker News Reader

Eventually these products are going to be uselessly nerfed to anyone studying above a highschool level. Can't ask about advanced biology, you might be engineering a virus. Can't ask about chemistry because you might be manufacturing drugs. Can't ask about physics, navigation etc. because you could be building missiles.

View on HN

Which is why I think very soon there will be development of dark AI, think of it as darknet AI. You can ask anything and it will tell you. It's only a matter of time. How to make a bomb? Sure, here's a shopping list for your local hardware store…

there is already a "darknet AI" site on Tor. I don't recall its name, but I found it on a list published by Sam Bent, a Youtuber who talks frequently about his experience as a former darknet vendor of moonshine (seriously).

Hm run on Tor, will answer anything, an takes payment in crypto? Cool

It is entirely reasonable to run a model like DS Flash 4.1 locally. Not easily. But feasible. If you drop down 100b - 200b models they are far more feasible.

DS Flash 4.1 is close to Opus 5. Beats Opus 4.8 in all of my little evals. You can abliterate local models. The genie is out of the bottle. There is no going back. This is why you see Anthropic and others talking about slowing down the frontier. Because they realize that you can’t even stop local users at this point. We basically have the equivalent of frontier models within reach on local hardware that cost less than $10k

It isn’t even a matter of time. You can just do it right now if you know how and have a bit of hardware.

e: this raises serious questions for me btw. The most obvious lever is controlling hardware. Controlling models and their capability is over. Controlling access to models is hard. Controlling access to GPU and fast RAM is more feasible. In 3-5 years every nation will understand how to distill, build, and abliterate safeties in models. This technology is fundamentally not difficult to work with.

It already exists, you're describing the open weights AI. The only reason they aren't "dark" at the moment is that it has no reason to be dark.

In the fullness of time, though, we'll also see previous-gen frontier models leak out of their labs and then get their censorship removed. It won't be as good as never having had it in the first place but it'll work just fine. It is likely a momentary and transient occurrance that the AIs can only be run with super specialized hardware. Eventually the weights, which are just big piles of numbers after all nothing terribly special about them other than that, will not be something that can be kept locked away completely and forever.

With all technology for the past 25 years people have said "You can just run your own" and that has worked approximately zero times.

I'm not sure what you're trying to say. We've been around the "wheel of reincarnation" multiple times [1]. It has worked many times. And there's even more things that you can do, even if you don't, e.g., self-hosting Jellyfin isn't really that hard.

[1]: https://www.winestockwebdesign.com/Essays/Eternal_Mainframe....

i would be shocked if it doesn't already exist. Surely there's research being done with such models that have no guardrails or brand safety trained into them. It's similar to gain of function research on viri, the work needs to be done to understand what is possible. If it exists at all then it's been stolen by nation states because of the value similar to how BigDeal(tm) weapon research always ends up stolen (nuclear bombs, stealth, etc).

Fable 5.1 has a much better classifier for biology than when it came out. It has no problem talking about the chemistry of cell signaling pathways. When it first came out even basic biology questions were tripping the guard rails but not now.

> Eventually these products are going to be uselessly nerfed to anyone studying above a highschool level.

And are already banned to many studying at that level.

Reply on news.ycombinator.com