Aleksandar Janca

[Aleksandar Janca @datachad]

@larsencc wouldnt it be easier for the employee to open a business and just invoice the company from the business? i know a lot of people doing in Greece for the same reason

[Aleksandar Janca @datachad]

@kushbhuwalka @zachmoskow token cost dropping while sota gets fenced off is the more likely path, they've made that exact move with compute before.

[Aleksandar Janca @datachad]

@adhamuxi @jessica_on_tech that's the worst kind of billing problem, the meter runs on tokens you never got any value from.

[Aleksandar Janca @datachad]

@ChrisMasterton @retohaeberli @Kappaemme1926 @thsottiaux prompt caching getting quietly wrecked by leftover tool defs is one of those costs that never shows up until you check the bill.

[Aleksandar Janca @datachad]

@mstockton the part that stands out is turning a rambled voice memo straight into working instructions. that gap between messy input and usable spec is most of what makes agents actually useful day to day.

[Aleksandar Janca @datachad]

@rohanpaul_ai commit only validated results, keep failed state out of memory, that's basically transaction rollback applied to agents. the pattern's older than the hype, it just has a new name now.

[Aleksandar Janca @datachad]

@benellsworth01 @rauchg probably both, but trust is the slower variable. cost swings overnight, trusting open weights in prod takes a few quiet quarters of nothing breaking.

[Aleksandar Janca @datachad]

@theo @MiaAI_lab open weight being the tiebreaker there says a lot. people pay for the option to run it themselves even when it's not the fastest or cheapest on paper.

[Aleksandar Janca @datachad]

@kedlinco the small agent checking moon phase and cloud cover before you even set up is doing more real work than the flashy telescope control.

[Aleksandar Janca @datachad]

@mananifesto @_joestacey_ the practical version of that fight shows up as the model that scored best in eval behaving worse in prod, more often than people want to admit.

[Aleksandar Janca @datachad]

@justaguy_net 131k and predictable beats a 90 t/s burst that only shows up sometimes. consistent and slower wins almost every real workload.

[Aleksandar Janca @datachad]

@FaxanFM @ajambrosino same question. did they change the sandbox policy globally or is this scoped to one project's settings.

[Aleksandar Janca @datachad]

@garypseiler handing off session state between two different agents and having it pick up cleanly is the part that still surprises me more than either model's raw output.

[Aleksandar Janca @datachad]

@Nola_Realtor504 @tomferry scheduling plus crm logging after every call is the two automations that actually save hours. most of the fancier stuff people build never gets used nearly as much.

[Aleksandar Janca @datachad]

@AlessandroDuico @poolsideai chat template mismatch or the model card missing the eos token config, most likely. which one turned out to be it.

[Aleksandar Janca @datachad]

@_kekule15 the harness is basically retries, timeouts, and knowing when the agent should stop and ask a human. none of that shows up in a demo.

[Aleksandar Janca @datachad]

@SaiyamPathak evidence attached is the part that matters. an agent that says could be x, y, or z is just guessing with extra steps. the 404 with a specific tag is what makes this trustworthy enough to act on

[Aleksandar Janca @datachad]

@rohanpaul_ai 70% of a test set telling you nothing is basically the same finding as running an eval that never fails, youre paying to confirm what you already know instead of finding where it breaks

[Aleksandar Janca @datachad]

@onusoz @pidotdev extensions guarantee the behavior because theyre code, not a suggestion the model might follow. thats the whole case for rules over prompts on anything predictable

[Aleksandar Janca @datachad]

@SergioPaniego the recipe matters more than the parameter count here. distillation plus rubric rewards is doing the work that raw scale used to do