@SemiAnalysis_ They did not have capacity for 100T tokens, clearly as evidenced by the outages and errors when handling <10T tokens per day.
The repetition of 100T is a real testament to marketing, and totally separate from reality
@doodlestein Did you also experience a lot of timeouts and errors?
It was good enough for me when it worked, but the reliability was quite poor - daily issues. I have to think that improves when usage is moderated by cost.
@mynameisyahia@damonchen Probably automating it because its easy now, and all he has to do to cancel is tell his agent to cancel the renewal... across many different services, never having to be bothered figuring out how to cancel XYZ again. Simple switch, seems handy.
@atbeme@mxlcolm It has good defaults, good feel, easy switch, great features (vibe mode, workflowz, etc). Very worth familiarizing
If you had tried it, you may not have built this
@Greedisgood2024@usr_bin_roygbiv My experience has been that everything on deepinfra is typically 40tps or less, meanwhile gemmy3.7 is 120+. Other models are better at some things, but speed has a quality all its own. Feels good. Often I can let things run, but when trying to get it done, Gemmy provides
@of_by_one@ImSh4yy@ikuznetsov_com The service price might increase in the future, so one could pay hyperinflated hardware prices now, to run much less capable models now, because some day service might get more expensive.
I... am not sure I am a buyer. I hope people are having fun, and not swept up in hype.
@kmcnam1 I would be embarrassed to put out something like this. Cool personal project maybe, but the awareness I don't technically get the system design and the architecture is garbage, then to promote it publicly for use?
Kinda embarassing.
@DrPhiltill@grok Fast mode is gonna use your limits faster, so if you are running out of usage turning that off if first step. Likely other mistakes as well.
@teortaxesTex Ox painfully slow. Gemini can complete 10 tasks before Ox finishes 1.
Its glm 5.3 so I dont understand the google speculation part. If it were google, or gemmy pro, its TPS would be better.
@willwashburn@doodlestein Life humbles us all at various times!
I recognized it because I'm a systems engineer. I've made dumb mistakes, overlooked, lacked imagination, etc.
Examples like this do challenge the hype. Systems design is hard. Goes wrong easily. All the intelligence in the world, yet...
@willwashburn@doodlestein A lot of overconfident people making a lot of money with no idea how to do things, doing things their superintelligence tells them to, without understanding the implications.
Same way their testing keeps hacking other companies. Not providing effective controls or guidance.
@mrginden@natebjones It might be reasonable for a person to think that if 5.2 is covered, then 5.6 is the same thing.
Insurance is not a person. And their lawyers are not reasonable.
@mrginden@natebjones Insurance sometimes.
Before insurance pays out your claim they verify the event qualifies.
If they can point to a failure to maintain the controls described in the policy, they will not cover the claim.
Policy is written in such a way to limit insurer risk, it flows from there
@5fakb48@1kartikkabadi1 I require determinism, so my agent writes scripts and I verify them, I review those to make sure they do what we want. So how the agent gets it done doesn't matter, the final product that moves into production matters a lot.