Hacker News Reader

Comment by zdragnar | Hacker News Reader

The real question, to my mind, is "is the current generation of AI yet another dead end?", because whether or not the tech can improve upon its flaws will determine whether or not it is worth a business investing in.

We've gone through AI winters before, where the actual techniques and hardware simply had a terminal point of usefulness beyond which it was unlikely to grow.

If hallucinating bad information is to be regularly expected / intrinsic to the tech, then it's basically Clippy 2.0 and a dead end.

On the other hand, if we can expect lower power costs and higher trust in the output (i.e. needing less human intervention) then it makes sense to start finding places where it can fit into the business and grow over time.

I'm personally in the camp that it is a fun toy but with limited applicability for most businesses, and unlikely to grow beyond that. I'd love to be proven wrong, though.

View on HN

I've found that LLMs are most exciting and useful when factual accuracy is irrelevant, like as a personal tool for creative brainstorming. Or searching for information that you're going to immediately validate anyway.

It's definitely something more than a toy, but I'm not sure that you can build a trillion dollar industry on that kind of stuff.

The problem with using large language models for brainstorming or writing is that the fundamental mechanism by which they work is to choose the most common thing to say at any given point — that is, the most average, middle of the bell curve thing to say. That's how they give the appearance of having any form of coherence, by rarely if ever deviating from the happy path. So any ideas you get from it are going to be pretty unoriginal, and even if you give it original prompts for ideas, it's eventually just going to regress back to the mean again as it traverses the probability space back in the direction of the average at every step. And its writing is always going to be essentially the average human writing style.

The public foundational models maybe.

But this isn't true in general, you can easily train a local model to write in very localized styles, and include temperatures that allow for wild swings outside of average.

If you want a rambling, occasionally brilliant Kerouac or de Montaigne you can make one.

Ya, having AI write a story based on some criteria, hallucinations aren’t much of a concern. I wonder, however, if you can curate and add to your content to make a AI chat bot more effective, since they are all attention based, focus on the attention and then present AI has the tool for getting info for that topic.

FWIW, I think inaccuracy - what people call hallucination - will probably be limiting in some applications, like public facing broad based chat applications with stringent requirements on factuality.

There are lots of other applications - in particular I'm thinking about turning unstructured data into structured data for aggregation and analysis, where you can tolerate some level of error and LLMs can work quite well to automate it.

Beyond that, obviously copilot or assistant type stuff is already something people use. It doesn't have to be flawless for that, the main use case for chatGPT is not making high stakes decisions blindly based on the answers it gives out. It makes mistakes and does dumb stuff sometimes but it's also helpful, the flaws aren't a dealbreaker even if people would like them to be.

At the same time, I don't expect it to get much better, it's going to be about getting useful work out of what we have.

The criticism in the article seems pretty weak. If the goal was an omniscient oracle, then, yeah, probably not the god you were expecting. However, I don't see how this is not more cost effective than an outsourced call center. Speaking to a person in another country who barely understands my language and has 5 scripted responses can't be an improvement. The article does not even speak to error rates for humans performing the same task.

>On the other hand, if we can expect lower power costs and higher trust in the output (i.e. needing less human intervention) then it makes sense to start finding places where it can fit into the business and grow over time.

So far, simply throwing more compute at the problem is making models hallucinate less and less. I think this easily explains why all the big players are making such massive investments in hardware, reliability seems to be an almost solved problem.

Reply on news.ycombinator.com