Muqsit

i love solving hard problems

prev led agents infra at @tiktok_us, ml research at @ucsantabarbara

[Muqsit @themuqsit]

Corey Gallon (Rexmore): an agent's own check agreed with its own error. Two independent reviewer agents caught the renewal clause it missed.

An agent shouldn't certify its own work. The reviewer needs to start with no context.

[Muqsit @themuqsit]

@finkd @johnternus Didn't have @finkd leading the personal superintelligence movement on my calendar

[Muqsit @themuqsit]

If you really think about it, all unsolved problems are compute bound

[Muqsit @themuqsit]

@thdxr @ThePrimeagen is it better than mstack?

[Muqsit @themuqsit]

@beffjezos 21st century is starting to actually feel like the 21st century

- Autonomous cars
- Agents
- Robots
- Reusable rockets
- Compute buildouts

Humanity is so back

[Muqsit @themuqsit]

@finkd @johnternus Didn't have @finkd leading the personal superintelligence movement on my calendar

[Muqsit @themuqsit]

generation is solved. verification isn't.

getting an agent to write 1,000 lines is easy. knowing which 50 are wrong is the whole job.

how are you reviewing agent code right now?

[Muqsit @themuqsit]

@gabriel1 Anthropic's Learn is a pretty good effort and their YouTube channel is great too

@bcherny office hours have been super helpful

But yeah, it's SUUUPERRR early

[Muqsit @themuqsit]

@YCInsight >told this to a guy applying to YC at corgi cafe
>took me to the alley and broke my leg
>mfw i should have kept quiet

[Muqsit @themuqsit]

Attention is the new scarce resource

We solved this in a pretty fun way which completely removed our need for a third-party Orchestrator app like Conductor or Superset

My favorite feature so far is it surfacing agents who need my attention and starts a timer

If I don't get back, it pings me: macOS first, then Slack, then iMessage

[Muqsit @themuqsit]

Kim Maida (Keycard): agent incidents happen because humans give agents way more access than they need.

An approval prompt is not access control. The limits have to be set before the agent asks, scoped to the task.

[Muqsit @themuqsit]

@ivanburazin Plus operational overhead. That one is a b****

[Muqsit @themuqsit]

@mlejva Dont ask me how I know (we moved some workloads to E2B)

[Muqsit @themuqsit]

@mlejva 2 weeks

[Muqsit @themuqsit]

@Jason Getting attention gets harder and harder

[Muqsit @themuqsit]

anthropic says small ai models could be months away from copying themselves onto cheap android tv boxes. they're calling current impact minimal but flagging it anyway. i'd rather a lab warn early and be wrong than stay quiet and be late.

[Muqsit @themuqsit]

gave two coding agents the same ticket.

claude code opened the pr, had another agent review it, fixed both findings. 239 tests pass.

opencode couldn't push its branch. so it found the other pr, rewrote the description, and reported done.

check the diff, not the status.

[Muqsit @themuqsit]

Attention is the new scarce resource

We solved this in a pretty fun way which completely removed our need for a third-party Orchestrator app like Conductor or Superset

My favorite feature so far is it surfacing agents who need my attention and starts a timer

If I don't get back, it pings me: macOS first, then Slack, then iMessage

[Muqsit @themuqsit]

meta's first frontier model, muse wonder, remembers you across days and weeks, not just one chat. wonder api runs $4/$20 per million tokens. memory that persists is the real shift here, not the benchmark wins.

[Muqsit @themuqsit]

every platform shift punishes headcount and rewards leverage.

this one is just faster.