Eddi

Opus 5.5 is as much confidently wrong as LLMs were before when you have stuff that is just complicated enough. This also means, you need to question the outputs and keep up with what’s happening even more the more complex the tasks are you are working on. Otherwise endless loops of nothing are waiting for you.

I feel the state of transcription shows, that AI models handling real life reliably is still a challenge. I mean, I have a studio mic pointed at me and I am the only one in the room. I think there is an element of me not being a native speaker, but the mistakes are SO annoying.

It would be kinda weird if OpenAI releases a model on Sunday for their 28 day challenge.

Man if you have to do ANYTHING that has concepts that clash with model training and you have to work against the model falling into learned behavior, you are in for a bad time. That is basically anything novel.

your agents explaining you how bad software you relied on every day for a decade is designed is quite something.

I kinda like Gemini.

A few years and I will be the dude in his 40s. Not ready for that yet 😄

Opus jumping to conclusions without thought will one day be my end.

I can't count the hours I tried basically all the different "code it once, deploy on all devices" frameworks from RN to Flutter and Native Script and Xamarin and whatnot. Now you just build them all natively from a base concept. In parallel. And you don't even think about it.

I really hope that OpenAI uses all their compute to create the next overlord model or something because man the service has never been this unreliable.

I would not trade Gemini 4 for Opus 5.5, but man is it ever better in approaching a problem in a scientific way. Had several unblock by Gemini every other model just completely missed.

Gemini following my CLAUDE.md and honoring my instructions better then Opus 5.5 was not on my bingo card.

Can some lab PLEASE drop a model that is as capable as Opus but pragmatic as OpenAI models? I can't take that Claudism bullshit any longer but there is no f'ing alternative.

I don't know man, ultrafast GPT 6.1 Sol on the Pro500 plan is quite useless when it drains so fast. Especially if your second account is still on the old 20x plan and basically does not drain at all it's quite weird. They really need to fix the 500 plan, or actually all the plans. Ultrafast is kind of like the only selling point they have at all. But I pay 500 bucks a month to at best get like 5-6 hours of ultrafast usage on the smaller model.

I just opened another editor from T3 Code. I think I haven't done that since that update that brought file editing and up-/downloads some time beginning of September.

Looks like there is no AI bubble after all.