Enjoy!
@MevetS, thanks for surfacing this excellent article!
Some in this forum have shown such an enthusiasm for artificial intelligence, such a willingness to believe in its imminent rise, and have made such claims for its usefulness that I’ve wondered where their usual skepticism has gone. They should read Bjarnason’s article (link just above) and find out.
I think the skepticism goes away when one finds its “a-ha” moment. While I’ve been using LLMs for some time, last week I used GPT to generate a tabular listing from a bunch of text that I had simply copied and pasted from my bank website, and then asked it to make it in hledger format. No CSV or Excel involved, just copying and pasting with the mouse and a prompt. I have been using computers for almost 40 years now and I had never interacted with them in this way.
I agree that LLMs are not intelligent. But to me they are useful, and that’s all that matters. Intelligence is in the eye of the beholder.
Which is actually what the linked item warns against.
Intelligence is not “in the eye of the beholder”, if it were, then the definition of intelligence would be vacuous.
I use LLMs and like you find them useful. But the hype, and the level of snake oil sales pitches, does everyone a great disservice. And could lead to quite a number of problems in the future.
I think this is a relevant quote:
“The first principle is that you must not fool yourself and you are the easiest person to fool.”
― Richard P. Feynman
That statement is emblematic of the confusion surrounding artificial intelligence.
And it has to be until science can fully explain the processes that make intelligence emerge in our brains. I honestly think that LLMs successfully capture a small subset of those process (namely, language recognition and the association with abstract concepts) and that’s what makes them useful. But by and large that does not make them intelligent.
I strongly disagree. By this measure we cannot define anything until we fully understand it.
As an example, we do not fully understand the physics of gravity. But gravity is not something that “is in the eye of the beholder”.
I use LLMs for this as well; makes text. to tables so easy. I like your last observation. Being able to ask questions of a pdf or a search removes an extra layer of frustration and work. It is like having a conversation with the data rather than rooting through a pile of information.
I’ve noted on this very forum that LLMs are likely similar to how humans process language (and as a result had a whole bunch of pop psychology BS thrown back at me showing how I was badly mistaken).
And I use these tools and they can appear magical.
The issue is in extrapolating from what they can do (and how they do it), to what is being claimed. LLMs are not intelligent and by themselves they will likely never be. They are statistical syntax without semantic understanding. They are the first module of a language understanding stack. That doesn’t mean that the current tools are not helpful. They certainly can be.
But realizing, “that’s a neat tick” is a good way to keep grounded and not buy into the hype.
Exactly. Does one understand the physics of gravity armed with Newton’s universal gravitation? Or does one need to know Einstein’s field equations? We cannot define what gravity is in the most fundamental way because at the end of the day we don’t know where those equations come from (and I’m not even sure there is a definite answer to that but that’s another point). In this sense, our understanding of the physics of gravity is, as you said, vacuous. But that does not mean that what we empirically know cannot be generalized to make useful predictions. And that’s what LLMs are: an empirical approximation to intelligence.
Didn’t want to skip this, on this we agree!
Huh? Either gravity is in the eye of the beholder or it is not.
We have a very good understanding of the physics of gravity. But we don’t have full understanding of it.
Likewise, we have good understanding of intelligence. Neuroscience, Cognitive Science, Psychology, and Artificial Intelligence are all fields of study that with a rich literature on the subject.
I think we agree a lot more than we disagree.
No, I won’t go any further than to call an LLM “a useful tool.” A tool based on an application of statistics, clever algorithms, and a massive amount of compute cycles and storage. And don’t forget, human feedback. An LLM without human feedback in its training is laughable.
A different consideration. Since these are just random number generators, you must know in advance how you will verify that it gave back the correct results.
Even worse, when a tool is right 9 times out of 10, we gain a false sense of security.
Useful - sometimes yes.
Reliable - not really.
Beware
I fully agree.
In my experimenting the best use is for things you are somewhat proficient in. So while I have some programming experience, I do not know Python. But using VS Code with the Continue plugin with Ollama I can get enough Python code to figure out what I want the code to do. And sometimes it does seem magical. And others, not so much. But at least I can tell the difference. (And one can of course run the code to determine if it works or not, lowering the barrier to entry for some coding tasks.)
Likewise, I am using a tool called CoTypist (currently in beta), as a context aware autocomplete. Again, the results are hit and miss, albeit promising. As with Ollama, everything is local to my Mac. And at all times I am in control of what is written.
What I would not do is use an LLM in a non-supervised production system. So fine for my hobby Python coding. Not so good for a production system that needs to be robust, secure, provide data and error validation, cover edge use cases, etc. A helpful tool for experienced developers, yes. A way to replace experienced developers, no.
Or in a situation where one does not have some knowledge of the subject matter. One comment I’ve seen is that the LLM, while not an expert in all subjects, is at least mediocre in all subjects. So an LLM would be better than me at writing Chinese. While no doubt true (a very low bar!), I would have no way to know when it was wrong. And if it was 95% accurate, I would miss the 5% of the times it was not. Fine perhaps if I’m trying to learn Chinese. Not so good when translating in diplomatic or business setting.
Again, as long as one understands the limitations of LLMs they can be useful. Falling into the trap of ascribing human level competence will eventually result in disappointment (or worse).
AI has already proved to be useful in many areas. Researchers have won Nobel prizes in physics and chemistry using AI. And we’ve recently seen what appears to be a major improvement in weather prediction. But all that is running in big data centers.
So far I’ve not been impressed by anything running on a smartphone or other mobile device. With the possible exception of some of the prototype “smart” glasses that are starting to show up. And those won’t be allowed to tell me the name of the person who just waved at me.
Ha! I wonder if I’m one of those people … or not?
Either way, it’s possible to be enthusiastic and sceptical at the same time, in much the same way as cars haves accelerators and brakes.
LLM’s are also useful when you know a programming language but are stuck on how to do something. They usually come up with useful things to try to help solve your problem. Another useful LLM example is the terminal application “Warp” which gives you advice when a terminal command doesn’t work.
Good read. It’s more about the limerence some experience talking to chatbots than usefulness or how emergent intelligence should be defined.
If y’all can take a joke: self-selecting AI skeptic readers thought the article described their experiences uncannily!