andrei trandafira

6th degree balancer

[andrei trandafira @andreidita]

@thsottiaux Make an unified hybrid mode windows/wsl so we can run either from same UX, also for a wsl main agent to be able to use a windows sub-agent for windows only stuff like computer/browser use plugins

[andrei trandafira @andreidita]

GPT 5.6 Sol Ultra is the first real orchestrator. The main agent keeps it's context pure only with high-level stuff, while it delegates low-level jobs to other agents that have their own context.

Also is the first model that is actually able to properly implement abstractions

[andrei trandafira @andreidita]

@thsottiaux Sol ultra is simply a different beast. Is the first model ever that does the proper work it needs to do on arcagi3. Just saying :)

[andrei trandafira @andreidita]

@thsottiaux @ajambrosino Oh, "temporal logic" that's geeky :)

[andrei trandafira @andreidita]

@thsottiaux I use them. Not sure how the 30 days expiry is counted, if it's per each reset, or globally at end of a month

[andrei trandafira @andreidita]

@thsottiaux that s absolutely awesome!

[andrei trandafira @andreidita]

@thsottiaux mine didn t reset yet. same as last time on 1st attempt

[andrei trandafira @andreidita]

@thsottiaux Mine didn't reset yet!

[andrei trandafira @andreidita]

@thsottiaux Got some very promising results with my reasoning harness. Pushed gpt5.4 mini on a task in ProgramBench from 48.1% vanilla result to 93%

[andrei trandafira @andreidita]

@sama Instant is finally able to solve the puzzle I test models with. O1-preview was the first to learn to solve it. Opus 4.6 still fails, but gpt 5.5 instant gets it. So yes, it's that good :)

[andrei trandafira @andreidita]

@tszzl not a very smart god either. claude (opus 4.6 haven t checked 4.7) still can t solve this little puzzle

gpt learned to solve this from o1-preview times lol

[andrei trandafira @andreidita]

@tszzl you just managed to fall under the same bad habit.
it's true of the majority of people calling themselves philosophers. but philosophy in itself doesn't have such affinities. there are schools of thought inside philosophy that don't suffer from this :)

[andrei trandafira @andreidita]

@thsottiaux @SIGKITTEN @Mugilan_SS @OpenAI One turn per week isn't that much of an overhead. And honestly someone a plus account that is capable of mustering a request that would get the model to work for hours/days...is way overqualified for the plus account and most probably s/he has what it takea to already be on pro

[andrei trandafira @andreidita]

@sama yea cause there is really a lot to be done to sort out current pre-agi global situation

[andrei trandafira @andreidita]

@thsottiaux @diegocabezas01 😂 love your humor

[andrei trandafira @andreidita]

@thsottiaux bring more gpu's!!!

just got this mid-turn on 5.5 on 100 pro plan

[andrei trandafira @andreidita]

@wangray @sama it works. just set the quality . but if you hit a prompt it will work, at least in playground. i got baffled by that no models available too

[andrei trandafira @andreidita]

@sama yes it did. it also improves reasoning alot because it can de-noise very well around the invariants!

[andrei trandafira @andreidita]

@MilksandMatcha building this. could use the extra inference

https://t.co/uBOPS3AYsa

[andrei trandafira @andreidita]

@Johnnydazhu @thsottiaux well you can if you tweak codex cli :) you can do a lot of stuff actually. For the app itself since it's a closed UX surface over codex cli, you need to wait for codex team, though