Just a guy with way too much AI psychosis
Chicago Suburbs
Joined August 2020
- Tweets1.8K
- Following28
- Followers67
- Likes5.7K
I am increasingly convinced that Astra is simultaneously the smartest and dumbest current frontier model.
Wtf did I even just witness with this Roadster unveiling? I am simultaneously extremely impressed and extremely confused. 馃槀
Muse Wonder Full is a fascinating model. It's that team member that may not be the smartest engineer in the room, but while everyone else is caught in the weeds, says, "Why don't we just do this?" And everyone else slaps their foreheads. It's that team member who split their school years between improving their engineering skills and developing actual social skills, rather than just being the lonely nerd no one wants to hang out with. It's that team member who can articulate to the boss what information they actually need to know, rather than ranting incoherently in excessive detail.
Wonder isn't the smartest model I've used, but it may very well be the least dumb. This model has been common sensemaxxed, and it shows.
Idk if it's just me or what, but 6.1 Sol leaks its reasoning process more than any other model I've ever seen.
Opus 5.5 is a great model. It's still worse that 6 Sol. Not because it's worse at coding; it's better. But it's worse at every other use case. It's easy to forget in this sphere that most people are using AI for other things.
Okay dot sounded interesting but was kinda an afterthought to me around all the other stuff. Now that I'm actually using it, this is fucking amazing.
Update on this: 6.1 Sol isn't the issue. The issue is that I'm used to the speed of commands on Linux, and the new cloud environments run on Mac.
Opus 5.5 is better than Sol 6.1, but OpenAI has done a sufficiently good job of getting me stuck in their ecosystem. So congrats, I guess. 馃槄
I'm still seeing people complaining about Sol and Astra usage. The cause and solution to this is known.
Codex agents spin up subagents to complete tasks now unless you tell them not to. That isn鈥檛 the issue in and of itself though. The issue is that, by default, the orchestrator agent checks in on each subagent every 30 seconds. Every time, it has to re-read its whole context window and the subagent's status. That's potentially hundreds of thousands of tokens on the orchestrator every 30 seconds for every active subagent. The funny thing is this is totally unnecessary, because the subagent will return its result to the orchestrator when it's finished, so the only reason for the orchestrator to check in on the subagents periodically is to give you updates and make sure they didn't get hung up on something or expand scope.
The solution is to go to your config.toml and set:
- min_wait_timeout_ms = 300000
- default_wait_timeout_ms 900000
This makes the default for your orchestrator to check in on subagents every 15 minutes instead of every 30 seconds. This alone made a colossal reduction in my usage consumption rate.
If you want to take this even further, you can make a skill that has the orchestrator use Luna 6 subagents at xHigh effort with very tightly scoped and defined tasks for most actual implementation work, while the orchestrator is constrained to planning, delegation, and diff review.
These two things together result in a consumption rate of about 2% usage/hour on my Pro x20 plan while running 3-4 Astra threads at High effort.
