Ask HN: Is ChatGPT4 broken since last weekend? | Hacker News Reader

Ask HN: Is ChatGPT4 broken since last weekend?

30 pointsfranze36 comments

Since the latest update (past weekend), responses longer than a paragraph often fail to fully execute, resulting in a "There was an error generating a response" message + the Regenerate button.

Seems to happen pretty much every time when I ask for code or the code executor starts.

Does ChatGPT4 still work for you?

View on HN

Slightly OT as it's about API not web app, but I think related to the fact that they switch out models without informing people pretty often, and I think do a fair amount of A/B testing -- so it's very "black box" from a user perspective still.

Since a few weeks ago, the API is just completely ignoring my system prompt. I have things like "never apologise" and "always assume the user has already consulted with competent authorities and other sources. Do not give advice not directly asked for." and "be succinct and do not repeat parts of the question in your answer". and "Do not use phrases like "it is essential to note" or "it is important" to try cut through its BS hedging.

This used to work great. Now I just get the same responses as if I hadn't put a system prompt in at all

Is there a place where one can finds useful prompts like yours?

I've had success using a variation of Gwern's system prompt https://gwern.net/tla#system-prompt

That just seems like the RLHF got turned up to overcome your system prompts so it can continue its “bs hedging”

It works but performance has definitely taken a hit over the last few days.

That’s my experience, too. Most things work, but I get occasional errors and response is often much slower than a few days ago. Just now, I was editing a GPT and the text produced by the editor bot appeared at about human typing speed.

Seems like they've traded the recent full-on system-wide outages for distributed intermittent failures. IMO they've rolled the updates out before they were fully ready. This is what happens when you use ChatGPT to write your scripts!

a lot of the issues seem capacity related, I'm not getting a lot of failures since the outage but responses have been slow pretty often. Plus launching these new features probably led to all sorts of unforeseen bugs in production, probably what caused the major outages. I'd imagine there are tons of ways to intentionally or accidentally cause issues using the new agent and code interpreter functionality

I just asked it to print the Debian logo as ascii art and it produced an obscure looking triangle brick thing

I am not even sure I am on GPT4 turbo, it streams even slower than before on GPT4.

I have the new interface and GPTs, but its slower than ever.

Even late last night, it was markedly slower than a couple weeks ago.

Yes, I’ve seen this happen to me multiple times. And Dalle outputs are lower resolution now

Have been getting the regenerate message a lot past week but regenerating has worked and other times fine. So certainly not broken for me at all and I’m sure if it were for everyone would be headline news, but definitely seems like some growing pains since launching the updated model.

>So certainly not broken for me at all

You are contradicting yourself (or so it seems to me) if you corroborate that you're seeing a certain issue, then claiming nothing is broken for you.

Even the UI has been unbearably slow as of late — To the point of a single letter per second.

It does work but it’s much slower, when we are discussing code it’s started trying to execute it, which has failed every single time so far (not because the code is wrong but due to operational issues like timeouts), very annoying

I try to transcribe a table from pdf and always break after 1/3 of the table.

Also had the "There was an error generating a response" message once, but a retry solved that. It's extremely slow for me right now though. Quite frustrating to use.

It's slow as heck for me this morning.

Also, asking the usual questions frequently gets it stuck in "Searching Bing for this answer" or something similar, but not 100% of the time.

For me it works but is very slow.

Very slow this morning for me as well.

Aaand I've had a long series of "Something went wrong" errors in the couple of hours since I wrote the comment above.

It appears that OpenAI, just like GitHub is struggling for uptime and resiliency and cannot handle Google's scale.

Lots of strange comparisons and claims towards replacing search engines which given the tons of outages is far from that.

Had ChatGPT not been behind a login wall, the inference costs and outages would only skyrocket.

Using actions is painfully slow this morning. Responses are slower than a 300 baud modem.

Been using the new 4 for mid sized code snippets with no problem pretty much every day.

Yeah. It was down last Thursday for me. Couldn't generate anything. It started working after a day but has been slow ever since. The update definitely took a toll on the performance.

Code executer is slow, but basic text completion seems fine.

It might actually be your ISP's peering with Cloudflare (looking at you, Deutsche Telekom). Try connecting through any VPN and see if it's faster.

I only use the azure API and did not notice anything

Azure has been incredibly slow for us lately. Requests that used to take ~30 secs a couple of months ago are taking 4 or 5 minutes at times. We’ve been told that the basic plan has shared resources for all customers so it’s spikey, and even though they’re supposedly adding more resources I guess the increase in demand has been faster than that.

Yeah same. I switched to Azure a while back but it's unusable now. Switched back to OAI API

Have been very very slow and even too slow to give the complete answer before the session is automatically cut down since two hours ago

I get occasional errors but generally working for me. A lot of new features were recently introduced. Perhaps this is related.

The web interface started working for me again within a day of the outage.

Gpt 4 has been working OK for me, but I’m not using the code executer.

Have not noticed an issue this week.

The Bing AI which uses GPT-4 on Edge browser sometimes does brilliantly and sometimes throws temper tantrums. Not sure why this is the case, though.

Reply on news.ycombinator.com