Dan Crewger

Product

[Dan Crewger @DanCrewger]

@thsottiaux More reliable German STT. Idk what's happening but the quality varies so much. Usually, mobile worked great and within codex it was horrible. Last two days mobile is soo bad. This is really such a convenience driver for me. One of the reasons why I use chat over Claude

[Dan Crewger @DanCrewger]

@d4m1n Yeah because they're using their training capacity to serve the model, to the detriment of their R&D.

[Dan Crewger @DanCrewger]

@thsottiaux The fact that people are scheduling runs in order to open the 5h window overnight so that they can start the day with 2.5h and 99% quota left tells you everything 😂

[Dan Crewger @DanCrewger]

I have a one-test benchmark where I ask any new model to identify tennis courts from a Google Maps screenshot and calculate how many Padel courts could be placed there. No model so far was able to pass. GPT 5.6 Luna (!) x-high was the first model to crack it. 👏#codex #gpt

[Dan Crewger @DanCrewger]

@steipete Honestly the app seems solid. Wish Hermes had this. Maybe I should switch?

[Dan Crewger @DanCrewger]

@tweets2xs @MTSlive Distillation is like taking a photo of an artwork, printing it an reselling it. It's not quite as good as the original, but it'll work for some people.

Alibaba supposedly used their access to Claude to train their own model based on Claude's answers.

[Dan Crewger @DanCrewger]

@thsottiaux I suggested this a few weeks ago on Reddit. I am open for PM job offers from OPENAI. 😉

[Dan Crewger @DanCrewger]

@clairevo Do you do 1 on 1s with your agents?

[Dan Crewger @DanCrewger]

I am finally a token billionaire - entirely without token maxxing. Also, thanks to #Fable, today was my highest API-price-equivalent day at $70. I know, beginner numbers for some, but I feel I'm on the right path!

[Dan Crewger @DanCrewger]

What's more, Fable missed a bug that was rated as P1 - Blocker by Opus. Kinda glad I didn't fully trust Claude.

[Dan Crewger @DanCrewger]

Had Opus 4.8 develop a feature, then Fable 5 and Gpt5.5 code review. Passed the findings back to Opus and asked which reviewer seemed more senior. It went with 5.5.

#fable #gpt55 #opus #codex

[Dan Crewger @DanCrewger]

@theo @themmyleke Which is actually wild to me. I mean Google owns the vertical stack - chips, infra, model, application. Their unit economics have to be great. Plus, they have a money printer. Why they didn't position themselves as the cheapest option is so weird to me.

[Dan Crewger @DanCrewger]

@thsottiaux @skirano @OpenAI handing out 10x extra usage for special projects.

My expectation: tools that speed up cancer research, environmental preservation simulation,...

Reality: an official chatGPT plugin that extends the Codex experience and will likely be swallowed within the month.

[Dan Crewger @DanCrewger]

@chrisgpt Hm no deepseek, no mimo.

[Dan Crewger @DanCrewger]

@depak_7 @chrisgpt It's that far apart on the hardest problem set. Is all your dev work super hard engineering problem? If not, that explains why. Models are likely similar in performance for easy-medium problems.

[Dan Crewger @DanCrewger]

@DanielSmidstrup The fact that I can use Codex with Hermes means I'll likely always keep at least OAI Plus.

[Dan Crewger @DanCrewger]

@DanielSmidstrup I'm the ball

[Dan Crewger @DanCrewger]

@DanielSmidstrup Currently on Codex, but limits seem to get tighter whereas Anthropic seems to get better, so I see myself bouncing back yet again.

[Dan Crewger @DanCrewger]

@tinkerersanky @OpenAI @thsottiaux And then they took 5.3-codex and with it a way to be smart about what to spend quota on.

[Dan Crewger @DanCrewger]

@thsottiaux yooo you guys pulling 5.3-codex from the subscription is not cool. I've been using it for less complex work. Don't get why we're forced to use overpowered models for simple tasks. I guess I'll get a $6 MiMo subscription to swap in for the easier stuff.