In the past week, there were problems every single day with every single plan type.
People unable to log in either due to backend being completely down, mysterious “cannot verify location”, even when they finally log in unable to load models, or getting free-tier quotas despite being on paid plan.
Add to this degraded performance of the models (sometimes they behave like the frontier models they are, sometimes they behave like a half-lobotomized Llama 3 local model).
Add to this russian roulette of “Agent execution terminated due to error.” with NO EXPLANATION given what exactly was wrong - bad tool call, bakcend crash? no no information.
30% of prompting ends in this error. (lately up to 70% of inferences)
None of your services are usable and have not been for the past weeks. There is ZERO explanation of any of it - radio silence. For both personal users on expensive plans and business users.
No status page for neither the google AI platform in general nor antigravity in particular.
Every single startup has a status page. But not Google.
This is a joke.
4 Likes
All these generous quotas (which are done so in order to lock people in to your products, and wont last forever), are not worth much and wont do their intended job, if you cant use the service.
I’d rather have much lower quotas that work, than high quotas with 50-50 risk that the model will crash halfway thru its reasoning/step execution, or that I will be locked out of my account for hours. Or that I am routed to a dumb model that will make a mess instead of the real thing.
Also when one model crash happens it causes subsequent calls to the model in that thread to be incredibly more probable. Yet to “go back” you need to revert any changes the model did so far. So you are either stuck re-doing the whole thing multiple times, or accepting the 70% error rate that will soon start happening, or do some patch-around-ing and prompting the model “hey you crashed halfway thru take a look what is already in file X Y Z” -.-
And because neither the user nor the model is informed WHAT went wrong, it cannot learn from its mistake. If it was an error in parameters passed to a tool SAY SO IN THE ERROR MESSAGE so it wont keep doing the same thing. And if its a complete backend error unrelated to the models behavior SAY SO IN THE MESSAGE -.-
1 Like
I think this is dead form and no one replies
it is pretty dead but their employees sometimes read it.
they just dont reply, or reply with a canned and bland response to cover their… assets.
otherwise plausible deniability - “we didnt know of the bug / scope of bug” (that’s also why there is no status page, that takes 15 minutes to set up in this year)
1 Like