Hacker News Reader

Comment by ctoth | Hacker News Reader

This one seems weirdly more verbose, in my brief testing. Definitely uses more tokens on average than 002 when answering. The writing style seems oddly-pretentious and yet somehow uninformed, perhaps there were too many copies of HN in the dataset?

I kid, I kid.

View on HN

Brilliant business move. Charge by the token, and then train the model to ramble! Pretty soon you'll have to buy it a beer before getting anything useful :)

Open-Source (Not you, Open-AI) models for the win!

They should have named it dumas instead of davinci.

After the failure of FB's Galactica model, I wonder whether LLM are starting to hit diminishing return. I've read of several ways diminishing returns could happen:

* Not enough training data - you've used up the Internet (even a percentage of the Internet might be as much as is usable by clever brute force).

* Not enough compute time to fully train (we're not close to that)

* The model covers such a large area that testing is impossible

One thing I'd speculate about is perhaps the more different subjects the program is expected to combine, the more it learns to spout plausible bullshit and clever quips, since for clever humans, that how they relate to stuff they don't know. So "pretentious but uninformed" might be a sign.

I still need to read past the first page, but I was reminded this morning of the 2021 paper “On the dangers of stochastic parrots: can language models be too big?” which discusses the harms and diminishing returns of moving to ever larger language models.

I had to rush out the door today after seeing this paper come up so I can’t speak much to its content right now. But if anyone wants to read it and reflect here I’d like to hear it.

https://dl.acm.org/doi/10.1145/3442188.3445922

It's also likely we're just seeing the limits of what the tech can do. It's not actually "intelligent" it's effectively a cool trick, and at some point the gap between what the trick can do and what a "perfect" AI can do becomes obvious. It's really not that different from Eliza being cool for a few minutes before being obviously found wanting. It just takes longer for the shine to wear off as we get to a better technology. Ditto for self driving cars - a steep climb to a 90% solution, then an impossible chasm to cross to something that actually handles the edge cases.

It’s also not trained the way humans are. We’re trained interactively, and from an early age, say things in order to express ourselves and try to get what we want. Machine learning models are generally trained passively.

The chatbots using the model over at character.ai give you a sort of genetic algorithmic interface by letting you choose from alternative responses and vote on them as conversations progress, and you can make your own bots then train them and get some very interesting results. I am personally extremely optimistic about the future of such technology and I think we've only begun to scratch the surface collectively.

I refuse to believe that anybody who compares LLMs to Eliza has interacted with the technology for more than 10 minutes.

I think the poster's point was that initial reactions to LLMs are similar to initial reactions to Eliza. Everyone was gobsmacked for a while, but then once the shine wore off and its limitations were exposed, everyone got bored and moved on quickly.

Ah yes, the AI Effect in action.

https://en.wikipedia.org/wiki/AI_effect

I am eagerly awaiting GPT4, it might be the case that language models will hit a plateau, just like video games graphics hit a plateau too, with each generation being less and less impressive compared to the previous, while still not being perfect

That’s an odd comparison since unreal engine 4 is incredibly close to live action. I watched the recent matrix demo several times in awe.

It can't have been trained on HN, I tried generating my comment for this thread and it would only output positive marketing speak instead of "This name seems to collide with the artist Da Vinci."

I knew it wasn't trained on HN when it didn't immediately try to correct my word choice.

I think you mean "refine" rather than "correct".

Verbosity horrifies me. If it ruins the next iterations of GitHub Copilot, I'm going to lose it. Long generations are flashy toys for marketing, but too prone to being useless in practice.

It's interesting, verbosity is generally a sign of having poorly formed ideas "sorry for the long letter, I didn't have time to write a short one" etc. These language models are side effects, afaik they are trained to predict the next word and at scale it becomes a good sentence generator. I wonder if there is some way to regularize by encouraging concise language

https://beta.openai.com/playground/p/LoLOf250swXu7UxeLWuQBeg...

have you tried tweaking the behavior, e.g. asking not to guess, to be brief, providing examples, constraining output with syntactic sugar, etc? e.g. it's like talking to a person, give feedback and you might get better result.

Just read the hot covid poll topic if you're looking for more good training data. ;)

I was really impressed by the 002 version. Looking forward to trying out 003 tonight!

Reply on news.ycombinator.com