Have you seen the work where they use another instance to fact check the first? The MS Research podcast made it seem like a really viable way to find hallucinations without really needing to code more. I’m curious if other people find that works or if MS researchers are just too invested in gpt.
I’ll check out that podcast but I’m deeply skeptical that one LLM can correct another since neither of them truly understands anything: it’s all statistics. Very detailed stats but still stats.
And stats will be wrong.
Before chatgpt released most Google AI engineers were looking into alternatives to LLMs as the limitations of an LLM were increasingly clear.
They’re convincing facsimiles of intelligence and a good tool for maybe 80% of basic uses.
But I agree with the consensus: they’re a dead end in our search for intelligence and their output is vastly overestimated
this was recorded early 2023 during the peak hype of generative AI
the guest immediately started making outlandish statements like “cancer will be solved in 10 years”, a statement entirely outside of his field of expertise: bad start but I kept listening all the way through
statements like “we have no idea how it answered a “give me a reason” to an AP bio question” demonstrate how out of touch both he and the head of open AI are with the work, if that story was even true. There are clear and easy explanations for it: the model has extensive training in formal education question and answer formats being the first.
the guest is the head of AI at Microsoft and has been in the field for 20 years: which is less of a flex than you might think. It means he has a literal vested interest in this being the next big thing. He spends 1/4 the episode selling Microsoft as the big integration for AI into everyone’s lives.
the solution to hallucination suggested hasn’t born fruit as far as I’m aware: hallucinations cannot be consistently detected by other instances.
he immediately makes claims about superhuman AI appearing in the next 5-10 years when there is 0 indication that’s close
he immediately anthropomorphizes the ai talking about it “reasoning”. It’s literally weighted functions. It doesn’t reason: it pushes input through a predetermined path and outputs a response. There’s no consideration, no extra steps: it just transforms input into output by training. Stochastic parrot.
He seems like a salesman who has fallen for his own pitch.
Thanks for listening and echoing some of my own doubts. I was kind of getting the feeling that MS Researchers were too invested in gpt and not being realistic about the limitations. But I hadn’t really seen others trying the two instance method and discarding it as not useful.
The tldr is nobody has solved it and it might not be solvable.
Which when you think of the structure being LLMs… that makes sense. They’re statistical models. They don’t have a grounding in any sort of truth. If the input hits the right channels it will output something undefined.
The Microsoft guy tries to spin this as “creativity!” but creativity requires intent. This is more like a random number generator outputting your tarot and you really buying into it.
They’re treated like something more than they are because we anthromorphise everything, and in our brains we assume anything that can string a sentence together is intelligent. “Oh, it can form a sentence! That must mean it’s pretty much already general intelligence since we gauge the intelligence of humans by the sentences they say!”
Have you seen the work where they use another instance to fact check the first? The MS Research podcast made it seem like a really viable way to find hallucinations without really needing to code more. I’m curious if other people find that works or if MS researchers are just too invested in gpt.
I’ll check out that podcast but I’m deeply skeptical that one LLM can correct another since neither of them truly understands anything: it’s all statistics. Very detailed stats but still stats.
And stats will be wrong.
Before chatgpt released most Google AI engineers were looking into alternatives to LLMs as the limitations of an LLM were increasingly clear.
They’re convincing facsimiles of intelligence and a good tool for maybe 80% of basic uses.
But I agree with the consensus: they’re a dead end in our search for intelligence and their output is vastly overestimated
I don’t know if you’ve already found it, but I’m pretty sure this is the episode.
Appreciate this link. Grabbed it on Spotify
Follow-up: I found the episode very unconvincing.
A few points:
He seems like a salesman who has fallen for his own pitch.
Thanks for listening and echoing some of my own doubts. I was kind of getting the feeling that MS Researchers were too invested in gpt and not being realistic about the limitations. But I hadn’t really seen others trying the two instance method and discarding it as not useful.
Here’s a recent story about hallucinations: https://www.cnn.com/2023/08/29/tech/ai-chatbot-hallucinations/index.html
The tldr is nobody has solved it and it might not be solvable.
Which when you think of the structure being LLMs… that makes sense. They’re statistical models. They don’t have a grounding in any sort of truth. If the input hits the right channels it will output something undefined.
The Microsoft guy tries to spin this as “creativity!” but creativity requires intent. This is more like a random number generator outputting your tarot and you really buying into it.
They’re treated like something more than they are because we anthromorphise everything, and in our brains we assume anything that can string a sentence together is intelligent. “Oh, it can form a sentence! That must mean it’s pretty much already general intelligence since we gauge the intelligence of humans by the sentences they say!”