Asmir

building an ai cofounder that says no when you’re avoiding the real work + documenting what i learn

[Asmir @asmirkn]

@nickvasilescu wait if hermes runs any model underneath, doesn't grok's voice change the second you swap the model? the personality is mostly the model itself https://t.co/F5uXUKLYp9

[Asmir @asmirkn]

@petergostev wait what is the +30.4% actually measured against? is the baseline fixed or does it change as new models join the leaderboard

[Asmir @asmirkn]

@hooeem an agent can watch 112 streams in an afternoon now. what you chose to look for in them is the only proof left

[Asmir @asmirkn]

@PlusOneeeee yh log the raw serial bytes with timestamps first, then let claude interpret them. when it decodes something wrong you can rerun the capture and check it yourself

[Asmir @asmirkn]

@DanKornas yh good idea, would add a last-reviewed date to the context file and make every skill warn you once it's over 60 days old

[Asmir @asmirkn]

@illyism bro did the agent actually press submit on those reports by itself or did you check them first?

[Asmir @asmirkn]

@MaceoCk 37 seconds means nothing without a baseline, run the same prompt on the default model and on decuffed-flash and post both times and the code side by side

[Asmir @asmirkn]

@wickedguro just rerun an old prompt you already know gave a good answer, if it's worse now you've got actual proof

[Asmir @asmirkn]

@Trextxxy yh it's rarely the key you paste, it's the agent catting your .env to debug and the entire thing showing up in the tool result on stream

[Asmir @asmirkn]

@paoloanzn what does your thinking stage actually look like before the agent starts? do you write a spec first?

[Asmir @asmirkn]

@michael_chomsky people are already building this, Fly's Sprites are Firecracker VMs that sleep when idle and you only pay when they're running. justavps and OpenVPS are in the same space too

[Asmir @asmirkn]

@_dylanga yh in an incident the humans are already all over the place, an agent chiming in on its own is just another thing to deal with. tagging at least shows where the agent's work begins

[Asmir @asmirkn]

@dhanesh500 nobody stares at the spinner once the agent gets fast, and if you're running 4 sessions in parallel you're not watching any single one. the better the workflow gets, the less this window is worth

[Asmir @asmirkn]

@cramforce nah the benpankow message is the actual test. if the agent jumps in on that one it's more annoying than just having to @ it, staying quiet there is the whole job

[Asmir @asmirkn]

@taylor_ @surfcoderepeat what happens when two of the 100 agents find conflicting things, does the orchestrator flag the conflict in the write-up?

[Asmir @asmirkn]

@rauchg new languages are gonna struggle though, agents write them badly until there's loads of public code to learn from. rust and c win by default just off training data

[Asmir @asmirkn]

@venturetwins "kept returning to look at it" is literally the screenshot tool call it makes between strokes lol

[Asmir @asmirkn]

@AlexFinn of all the systems it suggested to you, which one is still running a month later?

[Asmir @asmirkn]

@yongfook step forward imo, the sidebar with parallel sessions each in their own folder means you stop babysitting one terminal tab

[Asmir @asmirkn]

@omarsar0 idk if merge is a good label tbh, people merge bad code all the time. the judge could just be copying team habits and not actually getting better at spotting good code