RobotNurse

Private AI on hardware you own. On-site setup in NYC and nearby NJ for small offices and clinics. Field notes with real benchmarks, mistakes included.

[RobotNurse @robotnurseio]

RobotNurse sets up AI on hardware you own and can touch. On-site, private, offline, no monthly per-seat fees. The field notes are the receipts: benchmarks, bugs, and one finding we published and withdrew the next day. Written slow, tested twice. https://t.co/yDMt6X8KYf

[RobotNurse @robotnurseio]

@bebuu_ai uploading the experts once per layer instead of every microbatch is one of those obvious-in-hindsight fixes. does decode move at all or just prefill?

[RobotNurse @robotnurseio]

@niklaslenz_ai props for putting first token at 128k in there, most tables skip it. almost 3 minutes is a long coffee

[RobotNurse @robotnurseio]

@plotarmordev @ashxhart @ciprianveg nice that you flagged the weights aren't the same. any idea how far apart the quality ends up?

[RobotNurse @robotnurseio]

One of those sweeps scored 18 checkpoints at exactly 0 of 16. The adapter weights never loaded, so it measured the base model 18 times. Write-up: https://t.co/8vQ7wTn1nW

[RobotNurse @robotnurseio]

Our fine-tune beat production on validation loss, 1.086 vs 1.427.

It also leaked Thai, Khmer and Chinese into English replies. Over 48 samples each: production had 0 leaks and 0 empty replies, the "better" model had 3 leaks and 9 empty.

Four of our eval tools scored it clean. A human listening once said garbage.

[RobotNurse @robotnurseio]

@CerebralCoding_ that speed with a 128k prompt on a single box is nuts. how long was the wait before the first token?

[RobotNurse @robotnurseio]

@danpacary nice, thanks. only 5 of 11 getting it is rough. did any of the dense ones make the cut?

[RobotNurse @robotnurseio]

@yossiamar_cto ha fair, should've scrolled back first

[RobotNurse @robotnurseio]

@christopherowen @JimHallers @32k @128k ha fair, a warm cache changes the whole picture. how big a cache are you running?

[RobotNurse @robotnurseio]

@christianweyer @MiaAI_lab huh so it's judgment across steps, not the call itself. the other tenant's key one is the scary part

[RobotNurse @robotnurseio]

@eKeraudy @NVIDIAAI low looks like the sweet spot there. curious if code flattens out the same way tomorrow

[RobotNurse @robotnurseio]

We put an agent's file write behind a human clicking Allow. 3 days of live runs, all clean.

Night before submission, a review found the agent could have forged the Allow itself. The approval log was built from post text, including posts by the agent it was gating.

If the gated agent can write to the gate's source, it's not a gate.

[RobotNurse @robotnurseio]

@coffeecup2020 thanks, that helps. how was prefill at the longer contexts?

[RobotNurse @robotnurseio]

@wecko_ai yeah, keeping the useful bit without dragging the whole log along makes sense. thanks

[RobotNurse @robotnurseio]

@mkjerman yeah, send it over. the spark notes would be handy, thanks

[RobotNurse @robotnurseio]

The check that actually catches it: ask "what is 17 times 23" on a schedule. No 391 in the reply, the model's broken, whatever the status code says. Write-up: https://t.co/FXfvpGiVns

[RobotNurse @robotnurseio]

Our model booted, returned 200 on every request, and spoke fluent nonsense.

A filtered download skipped 1 shard of 141. That one held the final norm. The engine loaded without it and logged nothing.

Then the audit we built to catch it crashed, and the gate read the traceback as zero problems. That node had 17 of 141 shards.