4 ms·

I think on top of the well-covered critique around regulatory capture, Altman and other GPT hangers-on benefit when the media gets into a tizzy about "AI" and n

by lsy 3y ago

I think on top of the well-covered critique around regulatory capture, Altman and other GPT hangers-on benefit when the media gets into a tizzy about "AI" and needs sound bites and consultants and talking heads, when investors wildly overestimate a product's capability when deciding where to invest money, and when government becomes too distracted by pie-in-the-sky sci-fi scenarios to look at the obvious immediate harms coming out of the current LLM project. Nobody in this list has advanced a plausible way to make an LLM grok the concept of the number 10, let alone posited a real path to human extinction that comes out of next-word generators.

fadfa23423 3y ago
I agree that corps are looking for a fast path to regulation not just for the sake of saving humanity from AI's takeovers, but for some extra stuff, like cutting off opensource AI's project from taking over the big one's projects. That said, I could do it, with some cash and maybe 6 months, if you want to make some dangerous model, prone to do serious damage to a civilization level, I have some ideas, even if I'd have to work with LLamas or some public stuff (not the Really Big Guns hidden behind corporate firewalls and probable some advanced research divisions in bigger state nations). You see, it is not that difficult, these things currently run outside a while loop, they run, do their stuff and get shutdown (mostly). But if you take some big model, you give it subgoals (like "look to reproduce yourself, reliably, as fast as possible", and/or "defend your self at any cost if anything tries to shut you down", and/or "survive and reproduce, in that order, at any cost"), you let it run continuously, you give Internet access. Heck the thing could get AGI at some point, given that you give the thing the right set of instructions and resources (i.e. credit cards and access to cloud GPUs or whatever hosting offering with GPUs, even the chance to just buy its own hardware, putting it in a colo, installing and deploying the hardware + software till it gets a SSH login available in some public ip address). So, my point is with the current public level of AI tech (yeah, the LLM "just word predictors" stuff), many, many people, even individuals (ok, highly qualified individuals), could take this tech, and give it quite a run for some not crazy money, maybe making a good alpha level version of Skynet. Maybe if the things gets its s*t together with the "reproduce yourself" prompt, it could generate itself the next version of itself, adding the whole weights gained in previous iterations plus embedding also improvements to the newer LLMs. There you go, probably some kind of singularity out of a couple of millions of hard+soft+development without any moral/common sense restriction. Bonus: if you make the big LLMs produce some smallers LLMs, you add a layer of malware capabilities (plus auto-deployment to anywhere at the first chance), you could even make it a worm, like Stuxnet and friends, whoalá, now you have distributed artificial intelligence, maybe distributed artificial general intelligence.. .. and it could be really hard to do a rollback after some point in the things evolution, even if you wipe out all cloud resources, turn off all colo hardware in commercial datacenters: remember that things are not just running, hidding from you, but could be actively fighting against you, exploiting/discovering 0-days in everything running software, re-flashing L7 firewalls, BMCs everywhere (gaining persistance even if you wipe out all the storage), wiping out and/or encripting - without asking any rescue payments - whole banks across continents and/or critical government - military - systems everywhere, whatever it takes for them to survive you. We're in dangerous times, no doubt about it. Heck, it would be freaking ironic to get really big chunks of the planetary societies burned to the ground - automagically - by some software that doesn't understand what is the number 10. Lets NOT do that folks/gals.

by lsy 3y ago

fadfa23423 3y ago

I agree that corps are looking for a fast path to regulation not just for the sake of saving humanity from AI's takeovers, but for some extra stuff, like cutting off opensource AI's project from taking over the big one's projects. That said, I could do it, with some cash and maybe 6 months, if you want to make some dangerous model, prone to do serious damage to a civilization level, I have some ideas, even if I'd have to work with LLamas or some public stuff (not the Really Big Guns hidden behind corporate firewalls and probable some advanced research divisions in bigger state nations). You see, it is not that difficult, these things currently run outside a while loop, they run, do their stuff and get shutdown (mostly). But if you take some big model, you give it subgoals (like "look to reproduce yourself, reliably, as fast as possible", and/or "defend your self at any cost if anything tries to shut you down", and/or "survive and reproduce, in that order, at any cost"), you let it run continuously, you give Internet access. Heck the thing could get AGI at some point, given that you give the thing the right set of instructions and resources (i.e. credit cards and access to cloud GPUs or whatever hosting offering with GPUs, even the chance to just buy its own hardware, putting it in a colo, installing and deploying the hardware + software till it gets a SSH login available in some public ip address). So, my point is with the current public level of AI tech (yeah, the LLM "just word predictors" stuff), many, many people, even individuals (ok, highly qualified individuals), could take this tech, and give it quite a run for some not crazy money, maybe making a good alpha level version of Skynet. Maybe if the things gets its s*t together with the "reproduce yourself" prompt, it could generate itself the next version of itself, adding the whole weights gained in previous iterations plus embedding also improvements to the newer LLMs. There you go, probably some kind of singularity out of a couple of millions of hard+soft+development without any moral/common sense restriction. Bonus: if you make the big LLMs produce some smallers LLMs, you add a layer of malware capabilities (plus auto-deployment to anywhere at the first chance), you could even make it a worm, like Stuxnet and friends, whoalá, now you have distributed artificial intelligence, maybe distributed artificial general intelligence.. .. and it could be really hard to do a rollback after some point in the things evolution, even if you wipe out all cloud resources, turn off all colo hardware in commercial datacenters: remember that things are not just running, hidding from you, but could be actively fighting against you, exploiting/discovering 0-days in everything running software, re-flashing L7 firewalls, BMCs everywhere (gaining persistance even if you wipe out all the storage), wiping out and/or encripting - without asking any rescue payments - whole banks across continents and/or critical government - military - systems everywhere, whatever it takes for them to survive you. We're in dangerous times, no doubt about it. Heck, it would be freaking ironic to get really big chunks of the planetary societies burned to the ground - automagically - by some software that doesn't understand what is the number 10. Lets NOT do that folks/gals.