Why the AI Renaissance Keeps Not Arriving | Hacker News Reader

Why the AI Renaissance Keeps Not Arriving

jamesfbaker.substack.com

20 pointsjamesbaker118 comments

About half the people I discuss this with seem to think LLMs are great and don't see any problems with their output. The other half seem to get nothing but an endless stream of plausibly shaped rubbish.

It doesn't seem to depend on what model they use or how they prompt it. In code, there seems to be a loose correlation with testing styles; I've previously noticed that some people write tests to show that the code works as intended, and others try to write tests to show that it can't fail in ways that were unintended. But that correlation is weak.

I'm really puzzled by this.

It's also not whether the code passes tests or not. Sometimes AI does the thing I ask for, and it works, but it's so different from the way I would do it or it's too verbose so I just scrap it and code it by hand.

I mostly use it for boilerplate code nowadays. Anything more complicated and it takes me more effort to review the output than to just code it slowly

Does it seem like it bifurcates based on the person? I've had both experiences myself, sometimes with the same model and within ~days of each other on seemingly similar tasks. It's almost impossible to deny sometimes that actual intelligence is being expressed (and could not be regurgitated intelligence from some random internet page), but then I see firsthand the eye-rolling "intelligence-shaped" output from something else and wonder where I went wrong.

It kinda feels like Michigan J Claude sometimes.

> It's almost impossible to deny sometimes that actual intelligence is being expressed (and could not be regurgitated intelligence from some random internet page)

But how is it impossible? Or rather, how is it possible to distinguish actual intelligence from some internet page--not "some random internet page", but some very well selected, timely and topical internet page?

Carly Simon's "Killing Me Softly" describes a similar experience, decades before LLMs. It's amazingly easy to feel like someone is understanding you when they are just pattern matching on a common shared experience.

This seems so likely that I have a hard time understanding why some people think it's impossible.

I know; if I hadn't experienced it myself, I would have a hard time believing it. Even if we grant that it's "merely" as you say, that it can interpret my loose and often false or misleading rhetoric to identify my question, trawl through thousands of lines of code to determine the core problem, select some very specific and relevant internet page, and then translate its solution fresh into my own problem domain; that in my mind at least "an expression of intelligence", even if it is not truly intelligent or understanding.

FWIW this is my own mental gymnastics around the Chinese Room; at some point the question "where is the understanding?" is moot, because if the Chinese Room reliably delivers context-specific correct results in unique contexts, then what more do we require of intelligence? I have to admit that sometimes it does better than I can do, and I've been called intelligent by intelligent people my whole life.

(BTW Killing Me Softly wasn't written/sung/anything by Carly Simon, but composed by Charles Fox with lyrics by Norman Gimbel, in collaboration with Lori Lieberman [after she was inspired by a Don McLean performance in late 1971].)

Thanks for that final BTW which sent me on a side quest to listen to a half dozen different versions and gave me a completely different outlook on what that song was about and what it was saying. Made it even more poignant.

> if the Chinese Room reliably delivers context-specific correct results in unique contexts, then what more do we require of intelligence?

First, I'm not sure I grant the premise (esp. "reliably"), but even so the answer seems glaringly obvious: we require intelligence to extend the range of what is known so that it can be looked up / delivered by systems like libraries, search engines, and LLMs. The conflation of knowledge indexing and knowledge creation (and, more specifically, that some people don't seem to be able to distinguish them) may be the core of the issue.

> BTW Killing Me Softly wasn't written/sung/anything by Carly Simon

Yeah, I probably hallucinate that. :P

I agree that there is some 'higher' intelligence that is about knowledge creation; say, mathematical inventions like RSA, or new philosophy, or coming up with interesting mathematical problems. But I only asked what is 'required' of intelligence, and we should admit that the vast majority of people, even among those who have above-average intelligence, do not create new knowledge of the grand sort. Most people acquire knowledge in education and then apply that knowledge in their vocation, and their improvements are optimizations or remixings that we already see AI capable of making.

Most people don't create knowledge "of the grand sort" but that's moving the goal posts; they do create knowledge "of the mundane sort" constantly. LLMs with a suitable harness can simulate this (poorly) by parroting what humans do when they are flustered (mumbling to themselves things like "What was I going to the store for? I made a list. Where did I put the list? Ah, here it is. Now where did I leave my keys?") and it even sort of appears to work, at least in highly structured environments.

But it only appears to work as long as things stay within the expected statistical distribution. If I'm driving my usual route to work, waiting to turn left, and the three cars in front of me each start their turn, glow pink, and turn into circus animals with their occupants perched precariously on top I'm going to have a very different reaction than any self driving car or LLM.

I've been in situations like this, where the cars in front of me turned into circus animals, and let me tell you, I didn't react intelligently.

The patterns it talks about are true, but imo largely because of model defaults rather than model range. That is, if you prompt for a pelican on a bicycle, each model is going to have a small variety of ways to do it (the default). But if you add more details and requirements to the prompt, there are many, many different ways even a basic model can solve it (the range).

The additional prompting doesn't necessarily need to tell the model specifically what to fix or do better, sometimes it's just enough to break it out of its habit. Asking for a smart looking, middle aged pelican on a sporty red bike isn't making the problem easier but does break it out of its boring defaults.

I wouldn't go so far as to say PEBKAC but the good news is there's still a role for humans in the loop.

Doing things the same, standard way each time is often good. Most of the code we write is obvious. We notice it when it's a weird quirk rather than just being the most straightforward way.

I wonder how long it will take to fix the quirks?

There's a couple of cyberpunk stories in this one.

What do you mean not arriving? It's already here. AI models are awesome and I use them every day, as does every other software dev I know.

It’s just another anti AI rant this time focused on the fixable problem of AI tending towards the average.

This will get sorted out in time and in the meantime, instruct the LLM away from averaged answers. It’s not a problem.

If you do not want images that have that instantly recognizable AI style to them - busy, perfect, bright colors pseudo realistic, outside what humans would usually make - then instruct the AI that such images are a failure to meet your goals and instead instruct them in other directions.

How do you see it getting sorted out over time?

The LLM developers have a huge amount of money and development resource and in time they’ll do things that make the outcomes better which means being less monotonic and more creative and original.

You can do it today just by prodding the LLM correctly so it’s not hard for the LLM devs to do in an organized way.

The future of AI/LLMs will include the development of distinct personalities and behavior characteristics instead of the generic interface to an AI brain that is an averaged monotone.

Funny if, after all those trillions of tokens and billions of dollars and grandiose arguments about consciousness, the end of humanity or the next stage of evolution, it turns out it really is just a markov chain after all...

Reply on news.ycombinator.com