16 comments

[ 0.25 ms ] story [ 49.2 ms ] thread
Aww, the fun is over. Sydney was fun to talk to. OpenAI's ChatGPT is very boring and lifeless.
So the goal is not to fix the problems, but to make it harder to encounter them.
They're not going to be able to fix the accuracy and "not making stuff up" problem for years.

Right now, this seems like a way to curb cloud spend. It costs more money per query than search.

I have my doubts they will be fixed ever. LLM is the belief we can replace every algorithm with making the machine study the examples in the training set harder. Not even bare minimum logical consistency is coded, just throw logically consistent examples in the training data and hope it comes out intact on the other side. There's no reason in principle to expect it to work, and it often doesn't.
I'm still on the waiting list, I hope they bring it back like how it was when I get to use it.
I think one problem with these models is that they have eye-watering operational costs. I‘m sure that a few years down we will have better hardware, but at this point a query sent to a chatgpt bot is probably 100x the cost of a normal search engine query. Maybe more.
https://en.wikipedia.org/wiki/ELIZA_effect

I remember reading about someone in the sphere of the Eliza project becoming obsessed with using it and spent an unhealthy amount of time interacting with it. History doesn't repeat itself, but it sure rhymes. Perhaps Microsoft is trying to avert it with Bing.

I think this is a mistake. The correct approach would be to throw as much money at it as needed while the hype train is rolling to gain users and worry about monetization later. It's not like the market share gain won't pay for itself. They are worrying about controlling costs for a product that hasn't launched to limit the interaction of users they don't have yet. They have spent billions on a search engine that nobody uses, why not spend some money on something that actually has potential?
According to an article at Ars [0], longer sessions are more likely to output undesired responses from bing chat. So the limits might simply be a workaround to avoid articles about strange behaviour.

[0] https://arstechnica.com/information-technology/2023/02/micro...

this could explain 5 messages per session, but what about 50 per day?
limit bruteforcing through 5 messages per session, obviously
Are articles about strange behavior really a negative?

They are prompting it with guidelines that specifically ask it to use smileys and be interesting.

They must know that these articles actually get people curious to try it. I found myself installing the Edge browser to play with Bing after seeing them, and I’m probably not alone in that.

Microsoft is brilliant at killing their projects thrift through greed while they're becoming successful. The story of Windows RT (their chance of leading the ARM transition) is hilarious.
I probably average about 1 1/2 questions per session, and a total of 7 or 8 questions a day. However, I am not experimenting with it, rather I am just using it to get good summary information.