Hacker News Reader

Comment by pphysch | Hacker News Reader

> I picked AP Bio because the test is more than a simple regurgitation of scientific facts—it asks you to think critically about biology.

Of course, the LLM is just regurgitating the facts from the textbooks as well as thousands or millions of test samples.

It's certainly possible to formulate a biology question that a decent student would be able to answer easily, but LLM would be helpless at (by carefully choosing heterodox wording & structure).

The Age of Bullshit continues unabated.

View on HN

Yeah, I think that we need to probe how much these models are merely memorising their inputs rather than learning a more genereal, compressed (sortof) representation.

Honestly though, ten years ago NLP was using Bag of Words and Markov Chains, it's pretty phenomenal how far it's come.

To be fair, I mostly agree with your main point, but it's still remarkably impressive progress in a relatively short period of time.

It is impressive and useful technology but Gates is misleading by implying that the LLM is doing "critical thinking", when it is doing the opposite -- shallow regurgitation of "information" (i.e. tokens).

Experiments have demonstrated that GPT-4 does have a basic theory of mind, so it does not simply regurgitate.

No, experiments have demonstrated that GPT-4 can respond as if it had theory of mind. Basically all measures of psychological constructs need to be rethought given the development of LLMs.

> Of course, the LLM is just regurgitating the facts from the textbooks as well as thousands or millions of test samples.

Indeed. For some of us in the world of business, this is actually a very compelling "just".

Most use cases are indeed still bullshit, but I think we are getting to a point where maybe there exists at least one that isn't, and many of us don't want to be left holding the bag when the threshold hits 50%.

> Of course, the LLM is just regurgitating the facts from the textbooks as well as thousands or millions of test samples.

Sure, it is regurgitating facts. What's impressive is that it knows which facts to regurgitate.

It also makes me question how different this is from what we (humans) do. We learn a bunch of information from a variety of sources and experiences and synthesize that into answers. Sometimes we are correct and sometimes we are not. I suppose the difference is that we can extrapolate from our synthesized information vs simply formulating answers, but what the best LLMs do doesn't seem radically different than the average person.

The difference is that we are reading and comprehending information, while the model is making a prediction of what word is most likely to come next based on a huge dataset of words. As such, there have been many documented cases of the model seeming correct at first glance, maybe even second glance, but its actually totally wrong in a way that a human isn't totally wrong. Like chatgpt will give you a summary and a title and an author and a year and be entirely confident that this is the truth, whereas the title was written by a different author in a related field perhaps or was made up entirely. A human would not behave like this at all. If I asked you for some info you hardly knew, you wouldn't word vomit a paragraph like this, you'd reply with a confidence score, "I don't know," or "I know about this but not that," perhaps. Seems that if there is any money or time on the line, you'd rather use the human.

Reply on news.ycombinator.com