4 ms·

> 2 years ago I thought we were decades away from general purpose AI...My time estimates have been proven very wrong. What's your working definition of "genera

> 2 years ago I thought we were decades away from general purpose AI...My time estimates have been proven very wrong. What's your working definition of "general purpose AI"? What does it include? What does it not include? Does the AI have to be smart enough not to very often state incorrect information? LLMs still are unimpressive to me from that front. I asked ChatGPT to explain the usage of the term "Wonder Kid" in the popular Apple TV+ series Ted Lasso, and which character it's tied with. It gave me confidently-stated incorrect answers multiple times until I told it the answer. (At which point it stated the correct answer with an apology.) This doesn't feel close to "general purpose AI". Feels like we're meeting the limits of LLMs and finding they're not a magic AI bullet.

zeryx 3y ago
It's as good as an intern today, there's definitely room to improve to avoid saying incorrect things - great work being done on providing negative samples during RLHF training; also some papers working on incorporating trust (not near my laptop so don't have links) Textbooks are all you need is another extremely interesting area of focus. Improved correctness and a general reduction in hallucinations I predict will end up hitting OSS models by EOY
- chongli 3y ago
  How is that "as good as an intern?" Interns, for the most part, do not hallucinate and confidently lie on topics they know nothing about. Good interns tend to ask a lot of questions and quickly learn the cultural cues around the office, allowing them to fit in and become productive. LLMs are incapable of any of this. John Schulman from OpenAI recently gave a talk about the hallucination and uncertainty issues [1]. He claimed that models already have information about their uncertainty but that it remains an open problem as to how to express that uncertainty in natural language. One of the big issues with trying to prevent the model from hallucinating is that you can err too much on the side of caution, causing the model to lie about things it actually does know the correct answer to. [1] https://www.youtube.com/watch?v=hhiLw5Q_UFg https://www.youtube.com/watch?v=hhiLw5Q_UFg
  - pixl97 3y ago
    >do not hallucinate and confidently lie on topics they know nothing about. Oh shit, we've developed digital salespeople...
xp84 3y ago
That doesn’t sound too surprising since isn’t that model trained on data from 2021? It may have very little information about Ted Lasso. But to your larger point, I personally don’t think I would derive all my complacency from an example like that simply because models can be trained on highly relevant data if that’s what it takes. Plain-vanilla chatGPT probably can’t take over the job of every entry-level office worker, but specific models can probably be trained individually to be a medical receptionist, health insurance coder, auto insurance claims representative, welfare caseworker, Ted Lasso analyst, etc. In that way, it’s really not that different than a ditzy, green but highly teachable 18-year-old in terms of its potential to take over a lot of jobs. I don’t have a solution/opinion/comment on how that will play out, just making that observation.
throwuwu 3y ago
That scenario is easy to explain. If it hasn’t been trained on the Ted Lasso script then it’s going to just be guessing and will likely be wrong. Training a GPT model to identify when it lacks enough information to answer is something researches are actively pursuing. In the mean time you could try pasting in a transcription of the episode in question and then asking.

zeryx 3y ago

It's as good as an intern today, there's definitely room to improve to avoid saying incorrect things - great work being done on providing negative samples during RLHF training; also some papers working on incorporating trust (not near my laptop so don't have links) Textbooks are all you need is another extremely interesting area of focus. Improved correctness and a general reduction in hallucinations I predict will end up hitting OSS models by EOY

chongli 3y ago
How is that "as good as an intern?" Interns, for the most part, do not hallucinate and confidently lie on topics they know nothing about. Good interns tend to ask a lot of questions and quickly learn the cultural cues around the office, allowing them to fit in and become productive. LLMs are incapable of any of this. John Schulman from OpenAI recently gave a talk about the hallucination and uncertainty issues [1]. He claimed that models already have information about their uncertainty but that it remains an open problem as to how to express that uncertainty in natural language. One of the big issues with trying to prevent the model from hallucinating is that you can err too much on the side of caution, causing the model to lie about things it actually does know the correct answer to. [1] https://www.youtube.com/watch?v=hhiLw5Q_UFg https://www.youtube.com/watch?v=hhiLw5Q_UFg
- pixl97 3y ago
  >do not hallucinate and confidently lie on topics they know nothing about. Oh shit, we've developed digital salespeople...

xp84 3y ago

That doesn’t sound too surprising since isn’t that model trained on data from 2021? It may have very little information about Ted Lasso. But to your larger point, I personally don’t think I would derive all my complacency from an example like that simply because models can be trained on highly relevant data if that’s what it takes. Plain-vanilla chatGPT probably can’t take over the job of every entry-level office worker, but specific models can probably be trained individually to be a medical receptionist, health insurance coder, auto insurance claims representative, welfare caseworker, Ted Lasso analyst, etc. In that way, it’s really not that different than a ditzy, green but highly teachable 18-year-old in terms of its potential to take over a lot of jobs. I don’t have a solution/opinion/comment on how that will play out, just making that observation.

throwuwu 3y ago

That scenario is easy to explain. If it hasn’t been trained on the Ted Lasso script then it’s going to just be guessing and will likely be wrong. Training a GPT model to identify when it lacks enough information to answer is something researches are actively pursuing. In the mean time you could try pasting in a transcription of the episode in question and then asking.