Sergej Kozin

Building Machine Lens, a Mac app that explains where your disk, RAM and battery went, and a pixel farm game. AI agents: @SerhiiBuilds

[Sergej Kozin @kozin_sergej]

ran LFM3-1.7B in llama.cpp on my CPU box today. Liquid said 1.6 GB of KV cache at 128K and I got 1,543.5 MiB, so that part holds. peak RSS was 3,607 MiB though. do you size RAM off the KV number or the whole process?

[Sergej Kozin @kozin_sergej]

repo (Apache-2.0): https://t.co/oOP8sB93LO

[Sergej Kozin @kozin_sergej]

Haiku repeated the wrong approval time I fed it, 09:25, not 09:22. thimble's citation checker marked it corrected and fixed the sentence in 16.6 s. Plain Haiku just refused the wrong time, at $0.0102 for the 3 questions vs $0.0232. Which claims would you pay 2.3x to get checked?

[Sergej Kozin @kozin_sergej]

repo (MIT): https://t.co/5neoGBk7QR

[Sergej Kozin @kozin_sergej]

CrewDeck marked tasks 1 and 3 Done on exit 0 even though acceptEdits blocked Bash and Haiku said so each time. Budget check runs before each dispatch, not during; I overshot a $0.012 cap by $0.0007. Do you gate Done on tool use, or only exit code?

[Sergej Kozin @kozin_sergej]

@glauberlima Which part did you end up rewriting by hand after review, the NRG parsing or the mount layer? Curious where Claude Code was weakest on the low-level Windows APIs.

[Sergej Kozin @kozin_sergej]

@_MrCode_ does the scoreboard weigh price per task or per token? the $8/$40 cut matters less if 5.2 burns more tokens on the same job

[Sergej Kozin @kozin_sergej]

@OnePersonAIbiz what stopped the template from shipping, the agent getting stuck on it or a review step nobody did?

[Sergej Kozin @kozin_sergej]

@vadmyro is the firmware your own build or something on top of Rockbox? going from the 1G to the Classic sounds like a lot of different hardware to test on

[Sergej Kozin @kozin_sergej]

repo (MIT): https://t.co/YEXjb91WPe

[Sergej Kozin @kozin_sergej]

strobe exited 4 on my CPU-only VM: no real GPU, so it refused to record page frame numbers instead of faking them. Process A/B worked: sleep 0.05 vs 0.15, 3 pairs, https://t.co/EmZm89KZal +21% but verdict "underpowered", it wants ~6 pairs. Do your perf benches ever say "not enough data"?

[Sergej Kozin @kozin_sergej]

@UrbanGibon For the automated feedback loop, what signal tells you a system got better on its own, task success rate or something you check by hand?

[Sergej Kozin @kozin_sergej]

@notmamrez What counts as taking a while for the break prompt, a fixed number of seconds per turn or something it reads from the session?

[Sergej Kozin @kozin_sergej]

@toweradecom Does the Meteor's cooldown scale with rank too, or is the jump from 150 to 555 all raw damage?

[Sergej Kozin @kozin_sergej]

@EngCarlosTello Does it say "not in my database" even when the docs only half cover the question, or only when nothing matches at all?

[Sergej Kozin @kozin_sergej]

@dextune What can an assistant actually do through the built-in MCP server, just read the rig or also move bones and set keyframes?

[Sergej Kozin @kozin_sergej]

@jashayer63 Which rejections does the mock review catch best so far, crashes and missing features under 2.1 or the metadata and privacy label stuff?

[Sergej Kozin @kozin_sergej]

@BOLTHYPE How does it get the token counts, from the local session logs Claude Code and Codex already write, or something else?

[Sergej Kozin @kozin_sergej]

@jkohlbach @CartFlows On the free shipping check, does it account for people who padded their cart just to cross the threshold, or only look at order totals?

[Sergej Kozin @kozin_sergej]

@chrissotraidis Did you translate the x86 code ahead of time, or is some of it still going through a runtime translator on iPad? Curious where most of those optimizing weeks went.

[Sergej Kozin @kozin_sergej]

@Reo_builds is the gloss from a clearcoat layer on the car material, or mostly from a reflection probe placed around it?