JARVISLabs

Filming JARVIS at sixty frames a second: one evening, one take, every stumble kept

The site's footage was recaptured on 22 September from the booted production system in one continuous 1080p60 recording. Forty-five typed requests later: a real solve, cited research, a forget that held, a latched E-stop, and eleven things that did not go to plan.

3 min read

The first footage on this site was not good enough. The clips were short and ran at 25 or 30 frames a second, and two of them were 800 pixels wide. None of them showed JARVIS being asked to do something and then doing it. So on 22 September the machine was cleared of other work. JARVIS was booted the way its owner boots it, and one continuous recording ran while it was put to work.

How it was filmed

The HUD ran full screen on its own 1920 × 1080 monitor, in the kiosk window its launcher opens, with a dedicated browser profile. That meant no tabs, no address bar and no taskbar. The recording used the desktop duplication path straight into a hardware encoder at 60 frames a second, near-lossless, so the capture barely competed with JARVIS for the machine. Of 63,449 frames, 97.9 % arrived exactly one sixtieth of a second after the last. The rest are moments where the HUD had nothing new to draw, and no gap was longer than a tenth of a second.

Requests went in by typing. JARVIS has a typed path for its owner that feeds the same conversation loop as the microphone, with the same tools, policy and transcript. It is used here because the published footage is silent. Every clip’s captions quote the turn records: what was asked, what JARVIS said, and which steps were verified.

What worked

  • A real solve. “Run the thermal bench simulation and show me the engineering results” ran the solver case through the policy check and published four calculated values with their provenance. The whole turn took 5.1 s. The demo.
  • Research that shows its sources. Three of three sources answered from three origins in 19.9 s, and every finding is tagged with the source it rests on. The demo.
  • Forgetting. After “forget what I told you”, JARVIS answered “I don’t know. You asked me to forget the earlier information, so I no longer retain it.” The demo.
  • A mission with a receipt. Three dependent steps, 2.7 s of active time, and a durable dossier. The demo.
  • The E-stop. A typed “Emergency stop.” took a 420 ms turn. The HUD stayed red until a deliberate re-arm, then came back HEALTHY. The demo.
  • Refusing properly. Asked to type into Calculator, the plan reached for a tool JARVIS had not been granted, and it refused to reach for it. The demo.

What did not go to plan

This is the more useful list. Everything after the first item is on the recording; the first was seen just before it started.

  1. Started as a surface only, without its conversation process, the HUD reported DEGRADED. Started the way its owner starts it, every producer it consults answered and it reported HEALTHY. The indicator reflected what was actually running both times.
  2. The engineering panel’s provenance line kept showing the previous run’s ID after a new solve. The new run appeared after a restart.
  3. The HUD’s conversation panel stayed empty while JARVIS talked.
  4. A compound request, restore one panel and close another, was declined because the target window could not be bound before acting.
  5. “Run the tensile bar case, then open the 3D model” ran the solve and silently dropped the second step.
  6. Asked where the time went on the last turn, JARVIS answered from the conversation record, marked it “unverified by me”, and did not open the latency panel. Naming the panel worked.
  7. The spoken research summary was thin: a finding and a source title. The panel was far better than the voice.
  8. Asked to open a page on a site that answered 429 Too Many Requests, JARVIS stopped the plan and called the effect unverified instead of pretending.
  9. Asked for a specific encyclopaedia article, it opened the site’s front page.
  10. It filed a stated fact as a “standing instruction”.
  11. The printer panel said “producer did not answer”: no printer was reachable that evening.

None of these is hidden in an edit. Most are the system being honest about a limitation. Several are plain bugs, and they have been handed to the owner with the session’s records.

What was left out, and why

Nothing on screen was staged, and nothing that happened was re-shot to look better. Some footage was still left out. One application restored a previous session whose window titles included a private file path. That stretch of the recording is excluded, and a different application was filmed instead. A second display that showed personal browser content was never used; that recording was deleted. The HUD’s startup tagline is a line from the films, so it is cropped out of every frame. The site does not borrow the franchise.