Technology

First of all, thank you for the compliment. I try not to be insulting or condescending if the other party remains civil, so I’m glad to know that my effort is not going unnoticed.

I would have to look into who’s funding AISI to take a guess at that (which I’m open to doing, but I think you may have misinterpreted my meaning, so I will start by addressing that).

While I don’t think that most media is being truthful about AI, but I do think that they might not know they’re lying? They seem to just parrot whatever the tech oligarchs tell them to say without question; I don’t think they actually have the requisite knowledge to determine if what they’re being told is true or not.

AISI presents itself as a sort of regulatory/research entity. I don’t currently have good reason to disbelieve them, and based on the article you shared it doesn’t sound to me like they’re one of the people lying for money.

They do mention early in the article that for this research they didn’t put these models into sandboxes, they were intentionally allowed access to the internet. They also mention that they’ve turned off all of the safety features of the models before prompting them. With those given parameters, this isn’t a scenario that would organically occur in real life.

So, that’s the 1st half of the misinterpretation. The 2nd half has more to do with how LLMs (and all software) function.

When you prompt an AI, you aren’t dictating every step of the process to it, you’re asking it to complete a task, essentially. So, any time you prompt any LLM at all, with any objective, things are happening inside of the computer that you aren’t necessarily aware of, and that you definitely didn’t ask it to do explicitly. A good analogy for LLMs specifically is d20 dice. Essentially, the LLM has a separate 20 sided die for every single token (a token is basically a fraction of a word). So if the first token is “the” the computer picks up the die labeled “the” and rolls it. This die has the 20 most statistically likely tokens to follow “the” on it, one on each side. So, for this example let’s say that the “the” die lands on “re”. The AI now has the text “there” ready to print. It sets down the “the” die, and instead picks up the “re” die, and rolls that to determine the next token. This process repeats over and over again until the LLM is ‘confident’ that it’s assembled a plausible looking answer, and then it prints that answer to your screen. That is the extent of the thinking that an LLM does.

You didn’t ask the LLM to do any of that, you just asked it how to get cheese to stay on a pizza, and it told you “glue”. You didn’t observe any of the things it did without your explicit direction to get to “glue”, but they happened.

That’s not exactly how all programs work (most programs are deterministic, which is what you want a computer to be. That’s why most programs don’t hallucinate like AI does).

A more palatable example might be a computer game. When I fire up my emulator, and load up a Sonic the Hedgehog ROM, the CPU takes that ROM file and copies it into my system’s RAM. I did not tell the emulator to do this, it did it on its own, because it is programmed to do so… there’s a bunch of other steps in that process that I also didn’t explicitly request, but you get the point. (Hopefully anyway, let me know if I’ve done a poor job explaining this and I’ll try again).

EDIT - i’m listening to the most recent episode of the podcast “Better Offline” right now. Serendipitously, the topic of discussion is directly addressing exactly what you’re asking me about. I’ve only listened to the first third of the podcast so far, but I’m fairly confident that it will explain what I am currently trying to explain to you, but much more effectively.