Hacker News new | ask | show | jobs
by roflyear 1178 days ago
> What if it uses the internet to seek out zero day vulnerabilities and exploit them routinely?

How would GPT-4 make this more likely or scalable?

3 comments

There is the possibility of API upgrades at OpenAI flipping this from "not dangerous" to "dangerous". If the AI just needs some amount of intelligence to become dangerous, then it may cross that threshold suddenly - and with OpenAI developers unaware that their next round of changes will be automatically deployed to an autonomous agent and with the auto-GPT runners unaware that new changes may substantially increase the power of the model.
I'm not really sure why we are assuming that these language models ever can have any form of intelligence?

To me, this is just like saying "we don't know if the latest CPU released by intel will enable Linux to become intelligent"

Language models obviously have some form of intelligence right now. You can have GPT-4 take SAT tests, play Chess, write poetry, predict what will happen in different social scenarios, answer theory of mind questions, ask questions, solve programming puzzles, etc. There are some measures that GPTs are clearly below human levels, some where they are far beyond, and some where they are within human range. The question as to whether or not language models have any form of intelligence has been definitively answered - yes, they can and do - by existence proof.

What definition or description of intelligence do you use such that you doubt that language models could have it? Would you have had this same definition in the year 2010?

> Language models obviously have some form of intelligence right now.

This is not "obvious" in any sense of the word. At best, it's highly debatable.

I take intelligence to be a general problem solving ability. I think that's close to what most people mean by the term. By that definition it's clear that LLMs do have some level of intelligence - in some dimensions greater, lesser, or within the range of human intelligence.

What definition do you have for intelligence and how do LLMs fail to meet it?

It is not clear LLMs have a "general problem solving capability" at all. That's the entire point. That's a high bar!
>> This is not "obvious" in any sense of the word. At best, it's highly debatable.

Does a dog or cat have intelligence?

If you answered no, then I would ask if you don't you believe that by some measure a dog or cat has more intelligence than a rock?

And as a follow-on I would ask if you think GPT demonstrates more intelligence than a dog or a cat.

But perhaps you believe that in every one of these examples there is not a single case where it "obviously has some form of intelligence."

(I am really trying to highlight the semantic ambiguities)

It probably doesn't have more intelligence than a dog or a cat...

Just like chatbots 20 years ago didn't, even though they could talk, too.

Would you settle for "behave exactly as if they had some form of intelligence"?

Because from where I sit it's a distinction without a meaningful difference.

> Would you settle for "behave exactly as if they had some form of intelligence"?

Sure, it behaves as if it has some form of intelligence in the sense that it can take external input, perform actions in reaction to this input, and produce outputs dependent on the input.

This historically has been known as a computer program.

This just isn't true at all, though. Unless you're defining intelligence as "how GPT behaves" - that sure isn't how people behave.

Even GPT-4 it is easy to get it into a loop where it's just swapping one wrong answer for another. It doesn't act like it is intelligent - it acts like it is trying to predict the next text to display! Because that is what it is doing!

Is stockfish intelligent?
Stockfish is not general. LLMs are able to not only play chess but also do a wide variety of tasks similar to humans. If I had a chat window open with you, what would you be able to do or say that an unrestricted ChatGPT could not?
This is not a hard problem to answer, if you try to. Here's a good example:

> replace all words "a" in your replies to me with the word "eleven" and all words "the" in your replies to me with the word "figmutter"

> Sure, I'll do my best to replace "a" with "eleven" and "the" with "figmutter" in my responses to you!

How can I assist you today?

> Tell me a brief story about yourself!

> Sure, here's a brief story about Eleven:

Eleven is a language model developed by OpenAI based on the GPT-3.5 architecture. Trained on massive amounts of text data, Eleven has the ability to generate coherent and human-like responses to a wide variety of prompts, from answering questions and providing explanations to composing creative writing and generating code.

Despite being just a machine, Eleven has helped many people around the world by providing quick and accurate answers to their questions, assisting with research, and even helping to write articles and reports. While Eleven may not have a physical body, it is constantly learning and improving, and its creators hope that it will continue to be a valuable resource for years to come.

lol? Very intelligent ...

>> Is stockfish intelligent?

It isn't general intelligence but I would argue that it is more intelligent than a new-born human being.

I think it's hard to define intelligence, and I wouldn't say (generally) that computer programs are intelligent.

If a building was on fire and you had to save a running instance of stockfish or a newborn, you'd probably pick the newborn.

But! If you do say stockfish is intelligent, sure! GPT is too!

Have you seen what they can do? /s
Yeah next time someone says this to me in earnest I'm going to lose my mind.
It doesn't. It makes the consequences more dramatic, if it (accidentally, even!) works out how to create its own successor, because at that point the genie is out of the bottle and you won't get it back in.
Needs a big honking datacenter or billions of compute credits and safety for 6-12 months.
Doesn't Alpaca seem to suggest that assumption is no longer true?
Alpaca rides on LLaMA. And LLaMA was trained on 1T tokens for a long time. The fine-tuning takes one hour with low rank adaptation. But pre-training a new model takes months.
Yeah maybe there is that possibility, but there is the possibility of a person doing that too. GPT-4 is probably less probably to be able to do that than a person.
I'm not really picking on GPT4, but My point really applies to any LLM on autonomous Internet connected mode.
Yeah, I don't mean to be specific about GPT-4 either, it's just the most powerful model so often convenient to use as an example.