Hacker News Reader

Comment by rch | Hacker News Reader

Is the new board going to fix the product? It seems completely nerfed at present.

There was a tweet that from an engineer at OpenAI that they're working on the problem that ChatGPT has become too "lazy" - generating text that contains a lot of placeholders and expecting people to fill in much more themselves. As for the general brain damage from RLHF and the political bias, no word still.

It does feel like an employee that did really well out of the gate and is starting to coast I their laurels.

I've thought one of the funnier end states for AGI would be if it was created but this ended up making it vastly less productive than when it was just a tool.

So the AI of the future was more like Bender or other robots from Futurama that display all the same flaws as people.

If it is really agi that will be the result. Nobody likes to be asked the same question a hundred times.

Using the api, I've been seeing this a lot with the gpt-4-turbo preview model, but no problems with the non-turbo gpt-4 model. So I'll assume ChatGPT is now using 4-turbo. It seems the new model has some kinks to work out--I've also personally seen noticeably reduced reasoning ability for coding tasks, increased context-forgetting, and much worse instruction-following.

So far it feels more like a gpt-3.75-turbo rather than really being at the level of gpt-4. The speed and massive context window are amazing though.

Yeah, I usually use gpt-4-turbo (I exclusively use the API via a local web frontend (https://github.com/hillis/gpt-4-chat-ui) rather than ChatGPT Plus). Good reminder to use gpt-4 if I need to work around it - it hasn't bothered me too much in practice, since ChatGPT is honestly good enough most of the time for my purposes.

id be willing to bet all they're doing behind the scenes is cutting computation costs using smaller versions and doing every business' golden rule: price discrimination.

id be willing to bet enshittification is on the horizon. you don't get the shiny 70b model, that's for gold premium customers.

by 2025, it's gonna be tiered enterprise prices.

This has been the case with gpt-3.5 vs gpt-3.5-turbo, as well. But isn't it kinda obvious when things get cheaper and faster that there's a smaller model running things with some tricks on top to make it look smarter?

Gonna be hilarious when AGI turns out to be analagous to like a sassy 8 year old or something?

Like "AGI, do all this random shit for me!"

AGI: No! i don't wanna!

"Why?"

Ad infinitum.

It's actually interesting this is a universal phase of children.

Beginner's mind. I wonder if McKinsey's done any work on that...

Also, one of the simplest algorithms to get to the bottom of anything.

yeah it means like there's this genetic drive to understand the world. Do many other animals have this hard coded in?

Reminds me of those greedy slime things

Yes, definitely. Some never stop!

My son asks why, but only once. I'm not yet sure if it is because he is satisfied with his first answer, or if my answers just make the game too boring to play.

That's a premise of sci-fi "novel" Golem XIV from Stanislaw Lem: https://en.m.wikipedia.org/wiki/Golem_XIV

Where can one read the English for that?

With archive.org 14 day loan account:

https://archive.org/details/imaginarymagnitu00lems

for purchase:

https://www.kobo.com/ww/en/ebook/imaginary-magnitude-1

and via some zlib proxy if you're going full pirate.

https://annas-archive.org/search?q=Imaginary+magnitude

Thank you

they're B2B now, that means only political correctness.

and I'm not sure why anyone dances around it but these models are built by unfiltered data intake. if they actually want to harness bias, they need to do what every capitalist does to a social media platform and curate the content.

lastly, bias is an illusion of choice. choosing color over colour is a byproduct of culture and you're not going to eradicate that but cynically, I assume you mean , why won't it do the thing I agree with.

What does political correctness and bias mean in this context?

edit: I'm asking because to my eye most of these conversations revolve around jargon mismatch more than anything else.

Whatever the people who buys ads decides, losing your ad revenue is the main fear of most social media and media companies.

See twitter for example, ad buyers decided it is no longer politically correct so twitter lost a lot of ad revenue. Avoiding that is one of the most important things if you want to sell a model to companies.

The only AI safety that companies care about is their brand safety.

IIRC they've put in guard rails to try and make sure ChatGPT doesn't say anything controversial or offensive, but doing so hampers its utility and probably creativity, I'm guessing.

“Write me a web application.” Sure, here are some Microsoft and Google products to do so!

Not all filtering has to be prohibitive, just unnoticed.

Can you share a link to that x/tweet?

https://twitter.com/owencm/status/1729778194947973195

It's such a strange thing apparently they can tune gpt4 turbo cleverness up and down on the fly depending current load.

That would explain a lot! Sometimes when it's fast it spits out all code. When it's slower, it's Lazy! Thanks for the link

This is such a big issue using chatgpt for coding. Hope its a bug and not intended.

That has been my observation also

What is "RLHF" here?

Reinforcement learning from human feedback [1]

[1] https://en.wikipedia.org/wiki/Reinforcement_learning_from_hu...

Reply on news.ycombinator.com