@jeffjarvis An interesting thing about that article is that "When did The New York Times first report on AI?" is a question that's almost the perfect case of something a language model will be unable to answer correctly
Explaining why that is to people is so hard though!
One of the big skills involved in using this tech productively is developing robust instincts in terms of what questions it will likely answer well and what questions it's going to blatantly mess up
@jeffjarvis A key message I've been trying to get across to people is that, while this stuff looks trivially easy to use, it really isn't - it takes a great deal of experience and knowledge to use it in a way that avoids the many, many traps and pitfalls
Once you climb that steep and invisible learning curve it's one of the most powerful new tools I've ever encountered - but figuring out how to help guide people there is proving really difficult
@simon@jeffjarvis this is the message of the classic xkcd 1425: “In CS, it can be hard to explain the difference between the easy and the virtually impossible”
As a developer who works daily with very highly skilled and technical people (financial asset traders), the opacity for them of what I do continually surprises me. The gentle “I’m not sure how long this will take but…” is just as often followed by a request that’s a one-liner change or a request that needs three teams to coordinate work on.
@simon@jeffjarvis I don't think it takes extraordinary skill - or at least it would not need to, if the mental models of people who drive public discourse were better informed.
Of course you can't understand or explain a thing well, if the metaphor you are using to describe it is ill equipped and leads you down unhelpful paths.
@simon@jeffjarvis incidentally, "answering" is already an unhelpful frame.
To answer a question, as an act of human-like communication, an interlocutor needs to access a knowledge model that binds truth values to propositions. An LLM cannot answer a question in that sense.
Some prompts can generate useful outputs, though. That's how I would try to approach defining sensible constraints.
@jakob@jeffjarvis that's an illustration of the problem: despite lacking a knowledge model, they CAN answer many questions really effectively - they do that all the time
I'm not convinced that trying to explain to people that what they see LLMs do every day isn't technically correct or possible is a useful path
@simon@jeffjarvis it is epistemically false to purport there is a schema of question and answer that LLMs can partake in. LLMs generate outputs from large N data pools, stochastic training algorithms and prompts. That you frame the input prompt as a question is a choice.
People see things all the time where an initial model of what they see is incongruent with the underlying causes. That's where journalists and Benoit Blanc come in. Hopefully only one of them is fictional.
@jakob@jeffjarvis I understand how they work - the thing I find so interesting about them is how many things they manage to be useful for despite the very clear flaws and limitations in the way they are built
@simon@jakob@jeffjarvis I find the same thing is the case with human beings. Despite the very clear flaws and limitations in the way WE are built, we still often manage to be useful and occasionally even inspired.
@jeffjarvis Classic both-sides trick from the grey lady. Use the bad but industrial term but then include the sentence that “critics” oppose using the term because it’s bad and misleading.
@jeffjarvis I've come to conclusion that using words like hallucinate is part of a deliberate effort to make people think of these technologies as similar to human Intelligence. Similar to the vast amount of time spent making robots dance so we anthropomorphise them.
@robertklaschka It is in the best interest of the publisher that journalists paint LLMs as fully capable of taking their positions as journalists. And some journalists seem to hallucinate about LLMs, ironically. @jeffjarvis
@robertklaschka@jeffjarvis it also makes what the ai is doing seem more deep and interesting...it's having a mental adventure! Instead of more negative anthro language like: it's messing up, it's bullshitting..m
@jeffjarvis I also know nothing, understand nothing. I just assemble words into patterns that I’ve observed from other humans. It’s not that I think that AI is super smart, it’s that I think we’re overestimating how smart humans actually are.
Yeah, the word hallucinate is terrible. Luke Lafreniere (that's probably wrong, I'm terrible with french words) from LinusTechTips called it "LLMs are Confidently Wrong", aka, bullshitting.
@jeffjarvis "Figuring out why chatbots make things up and how to solve the problem has become one of the most pressing issues facing researchers as the tech industry races toward the development of new A.I. systems."
Yeah, I don't know if "why chatbots make things up" is that big of a question. Of course language models will do that. It's really a bigger question why they give an acceptable answer more often than one might expect.
@jeffjarvis Yep - I built @StochasticEntropy and asked OpenAI to confirm it's not returning anyone elses responses from my empty prompts - they confirmed it's just stochastic chains
@jeffjarvis fair enough, but the term was coined by the researchers in LLMs themselves and predates the (unexpectedly) public release of ChatGPT. Like "temperature", the term has a specific and unique meaning in this context.
@jeffjarvis I understand your sentiment, but I also agree with the label of “hallucination” as commonly applied to LLMs.
We’ve been anthropomorphising technology long before LLMs.
Computer “viruses,” “worms,” and “Trojan horses” come to mind. None of these are “real” in the physical sense, but they convey the concept accurately to someone familiar with the physical objects these words describe.
@ravigupta@jeffjarvis Most of those "it got an A on this test!" things are multiple choice questions
Multiple choice questions are the kind of thing a language model can be incredibly effective at - much, much easier than "when did the NY Times first mention AI?"
Add comment