Something about ChatGPT treating paid customers like guinea pigs for their unstable infra as the standard Silicon Valley playbook rubs me the wrong way. I’m a paying customer, I should get at least a reliable product. I’m not even asking for a 99% SLA.
Nothing worse in a workflow than to type away, craft a prompt, hit submit, only for it to say “There was an error in generating this response”, smashing the “Regenerate” button to no avail, hitting refresh on the page in the browser, trying again, hoping that it will go through, and then praying it stay stable for the next 4 messages.
I’ve also gotten other messages including network connection errors (yes, my browser is stable, my connection is stable, I don’t use a VPN), partial responses, mid-response erroring, “Error in Moderation” (!!??), and other cryptic errors. And with zero documentation on these errors or why they work. Some level of affordance in the design would be valuable (i.e. we have overloaded servers, it’s peak time, your message has a large amount of tokens so processing may be ineffective or slow), anything to help reassure me, a paying customer, that my dollar is valuable to them.
Why are we all paying for what feels like perpetual beta software? I get that it’s bleeding edge but it’s also trying so hard to rapidly expand with $30/user memberships for large 30,000 user organizations. Not to mention it’s fueling Microsoft’s own CoPilot experiments. And here I am, a lone user, trying to get it to spit out a few sentences for my own work. It’s a bit miserable to say the least.
OpenAI, we deserve better for the money we are paying for. The mobile app is well-designed, the web interface works, the vision functionality is great, but it’s all hype if your servers just can’t handle the load. I get it, they are moving fast, but please, stop breaking things.
Frankly speaking, open source GPT-4 Level models can’t come soon enough. Better competition will get to the root of this problem.
(also for the record, I also use Claude which is great, but has mega nannying issues, Gemini Advanced which is quickly become a lot more valuable, nat.dev for when I’m tired of rate limiting and want to use a semi-functional workspace, and a local LLM (Trinity-v1.2 7B Q4) which is great on airplanes but really heavy to use on 16GB of ram when placed beside other applications).
(also, also, GPT-3.5 is cool for certain use cases but GPT-4 is where it’s at for sure).