There's a lot of ways to store a lot of data for an LLM to utilize, to simulate long-term memory. WorldData/WorldInfo, summarization, exporting-importing details into an external xml or javascript, etc. And ChatGPT isn't the best metric when it comes to workaround solutions and out of the box implementation.
Not that I think it's as drastic of a risk or problem as people seem to fearmonger over. Usually such fearmongering is intended to protect idiots from themselves, but idiots are doomed to do stupid things one way or another, regardless of what tools they're using.
A good analogy of how it can be used or misused would be some holodeck programs and hologram interactions in Star Trek. In the right hands it can be a handy simulation to stir up some inspiration and new ideas, just due to the "RNG" mix of potential responses. In other hands you might have... Janeway or Barclay (in the first episode he was in anyway).
Yes, in the near future LLMs will get much longer memory, and that's why they're only going to get more popular as companions. They're already very popular as companions despite fairly short memories.
Humans as a whole are idiots. Widely accessible AI is like widely accessible nuclear weapons. It only takes one idiot and we all die.
...I just indicated how it's already quite possible and fairly easy to implement simulated long-term memory with LLM's. Implying that it's already doable to a fair extent.
And we're an extremely long way off for AI to be something that can actually be used to cause any kind of serious destruction by just one person. And likening it to having access to nuclear weapons is completely ridiculous hyperbole and total bullshit.
I would have said the same thing in 2015 about AI being able to do college essays, make photo-realistic images or write a compute program, but here we are. Technological progress really does appear to be exponential, at least in the realm of computers. We also have emergent properties of AI that weren't expected by the developers. A lifetime of progress may now only take 5 years. There's already AI that will do scientific research and write a paper which is apparently fairly accurate. Won't be that long until it can do it properly and then it very much will be more dangerous than a nuclear weapon.
There's a lot of ways to store a lot of data for an LLM to utilize, to simulate long-term memory. WorldData/WorldInfo, summarization, exporting-importing details into an external xml or javascript, etc. And ChatGPT isn't the best metric when it comes to workaround solutions and out of the box implementation.
Not that I think it's as drastic of a risk or problem as people seem to fearmonger over. Usually such fearmongering is intended to protect idiots from themselves, but idiots are doomed to do stupid things one way or another, regardless of what tools they're using.
A good analogy of how it can be used or misused would be some holodeck programs and hologram interactions in Star Trek. In the right hands it can be a handy simulation to stir up some inspiration and new ideas, just due to the "RNG" mix of potential responses. In other hands you might have... Janeway or Barclay (in the first episode he was in anyway).
Yes, in the near future LLMs will get much longer memory, and that's why they're only going to get more popular as companions. They're already very popular as companions despite fairly short memories.
Humans as a whole are idiots. Widely accessible AI is like widely accessible nuclear weapons. It only takes one idiot and we all die.
...I just indicated how it's already quite possible and fairly easy to implement simulated long-term memory with LLM's. Implying that it's already doable to a fair extent.
And we're an extremely long way off for AI to be something that can actually be used to cause any kind of serious destruction by just one person. And likening it to having access to nuclear weapons is completely ridiculous hyperbole and total bullshit.
I would have said the same thing in 2015 about AI being able to do college essays, make photo-realistic images or write a compute program, but here we are. Technological progress really does appear to be exponential, at least in the realm of computers. We also have emergent properties of AI that weren't expected by the developers. A lifetime of progress may now only take 5 years. There's already AI that will do scientific research and write a paper which is apparently fairly accurate. Won't be that long until it can do it properly and then it very much will be more dangerous than a nuclear weapon.