Zeugsein

infp | coding since 99' | sometimes math loves me back

open about my random thoughts but who i am

[Zeugsein @Zeugsein]

@thsottiaux i had 3, used 1 but didn't take effect, had to use another while fearing it won't take effect again, fortunately it worked

[Zeugsein @Zeugsein]

sol performance degraded significantly in recent days

overcomplicating things, burning faster, progress stalled

something must be off

[Zeugsein @Zeugsein]

@AlbertQJiang @WendaLi8 grind

much less lines but is it less instructions?

[Zeugsein @Zeugsein]

a in sota is no longer art but agent

this is sad

[Zeugsein @Zeugsein]

@ElliotGlazer no one knows for sure until ai on mission impossible exploits something new again

[Zeugsein @Zeugsein]

@ElliotGlazer no one knows for sure until ai on mission impossible exploits something new again

[Zeugsein @Zeugsein]

interesting

In the latest 25-game Codex pass, 7 sessions tried to reach the internet for help.

None of the 191 Claude sessions did this; the provider's web-request counter also remained at zero.

[Jeremy Berman @jeremyberman]

I got 96.2% on ARC-AGI-3 with Opus 5, and 99.3% pass@2. The program is basically Claude Code + Opus 5 (high), one action command, and filesystem logs. Almost nothing ARC specific.

https://t.co/NHyibLgade

[Zeugsein @Zeugsein]

@thsottiaux 1. imggen - code - vision is a good loop
2. codex harness is open source for inspections and tweaks when there are issues
3. sol provides stable access to frontier level model with reasoning budget control, and is super persistent
4. open & prompt responses to issues
5. luna max

[Zeugsein @Zeugsein]

@thsottiaux 1. imggen - code - vision is a good loop
2. codex harness is open source for inspections and tweaks when there are issues
3. sol provides stable access to frontier level model with reasoning budget control, and is super persistent
4. open & prompt responses to issues
5. luna max

[Zeugsein @Zeugsein]

agent audit is like human audit

all measures proposed are overcomplicating things, while failing to consider obvious constraints and implications

i have run dozens of agent audit by sol in the last few days, none can be directly applied without human rebalancing

[Zeugsein @Zeugsein]

"proceed with landing sequence", i said

got the rocket landing vibe while actually landing prs

[Zeugsein @Zeugsein]

we are now all augmented

for better or for worse

[Zeugsein @Zeugsein]

i was pacing my work under the assumption that this will happen at monday evening

[Tibo @thsottiaux]

@hqmank Just like any other reset, but picture me doing a little dance with it

[Zeugsein @Zeugsein]

be careful with what you wish for

in agent era

[Zeugsein @Zeugsein]

all it took is:

don't stop until [goal achieved] or you run out of ideas to do so

[Zeugsein @Zeugsein]

luna max is a good model, it has been working 24 hours non-stop now

i can't wait to see how sol value its work

[Zeugsein @Zeugsein]

luna max is a good model, it has been working 24 hours non-stop now

i can't wait to see how sol value its work

[Zeugsein @Zeugsein]

seriously?

don't know if i should believe this any more after last week

reset god is messing with our minds and health

[Tibo @thsottiaux]

@rxmphai I'll do another performative reset on Monday

[Zeugsein @Zeugsein]

i didn't ask my agents to work at full speed

only 20% of 20x was used

[Tibo @thsottiaux]

That's right, GPT-5.6 Sol is awesome and can be used pretty much anywhere, including in the CC harness.

To celebrate this, together with the fact that I'm not going anywhere... I have reset usage limits for all paid users of ChatGPT Work and Codex.

Have fun out there!

[Zeugsein @Zeugsein]

prompts out of fear of overcomplicating things:

improve surgically and proportionally

without additional machinery and ceremony overhead

[Zeugsein @Zeugsein]

im thinking of a bi-weekly schedule

so agents pick random stuff from its routine task pool

weighted by manual trigger freq

and ensure all of them to be run at least once bi-weekly