Summary

A hands-on capability test rather than a tutorial. Futurepedia's presenter runs GPT-6 Astra's computer-use feature across ten desktop applications — Google Earth Studio, Canva, Photoshop, Premiere, Illustrator, BrickLink Studio, OpenRocket, Algodoo, Blender and GarageBand on a mirrored iPhone — and reports elapsed times, follow-up counts and honest failures for each. The method is unusually disciplined for the genre: he deliberately blocks the model's scripting routes with instructions not to use code, APIs or MCPs in order to isolate computer use, then re-runs two of the same tasks unrestricted so the two paths can be compared directly. He also argues explicitly against the way the capability is being portrayed elsewhere. All results are a single operator's experience on his own machine, not a benchmark.

• The central finding is a comparison, not a capability claim: where an application exposes a scripting route, letting the model use it beat driving the interface on both quality and time. Recreating a scene in Blender took roughly 1 hour 45 minutes via computer use for a result he rates poor-to-middling, against about 54 minutes via Blender's Python API for a clearly better result — reported as roughly half the time.
• The same pattern held in Illustrator. Forced through the interface, a vector recreation ran 40 minutes, then a further 52 minutes when told to work until perfect, and he describes the result as 'not good at all'. Allowed to use Illustrator's Python API, the first pass took 20 minutes and was better, improving over two more rounds of 10 and 15 minutes — though he says even the Python result was not close to the original it was copying.
• He reports that Astra defaults to the scripting route unless blocked: asked plainly to make something in Illustrator or Blender it will drive the application's Python API rather than move a cursor. His practical advice is to let the model pick whichever route it is fastest at, and that his restriction existed only to test the feature.
• Reported timings for the tasks that worked. Google Earth Studio animation: 11 min 49 s first pass, three follow-ups totalling about 23 minutes, final result he calls 'basically perfect'. Photoshop thumbnail from scratch with layers preserved: 25 minutes, then 14, then 20 — about an hour total, and he used the output as the actual thumbnail for the video. Premiere text-behind-subject and pop-up animations plus sound-effect selection: 12 minutes, then 13.
• He is candid that the Photoshop hour was slower than doing it himself, and argues the time comparison is the wrong frame because the work was unattended — he was downstairs with his children and reviewed the result on his phone. The claim is about attention rather than speed, and it is his own framing.
• A single-pass success in software he had never used: asked to design Delicate Arch in BrickLink Studio, it produced a model of 257 pieces organised into 27 assembly steps on the first try. His screen recording of this run failed, so only the result is shown.
• On GarageBand driven through iPhone mirroring, the model recorded seven tracks in 52 minutes by tapping instruments directly. He reports three concrete interface limits it hit and worked around unprompted: it cannot hear the metronome, so it schedules taps from the BPM and lands close to but not exactly on the beat; it appears unable to tap two places at once, so it abandoned a single drum track and layered several instead; and it cannot hold a note, so it used the sustain control as a substitute.
• He rates the resulting song as 'not the best song in the world' while calling the method impressive. The observation worth keeping is the model changing course on its own after discovering a constraint, which he flags as the notable behaviour rather than the output.
• He pushes back directly on how this capability is being presented, saying he has 'gotten a little annoyed' with portrayals implying Blender is solved. From his own non-expert perspective he rates the best Astra output he has seen as closer to a competent intermediate user, and says it is nowhere near the complex work skilled users produce. He caveats his own judgement by noting he has only used Blender a few times.
• His summary position is that the model is not better than someone who has spent years mastering these tools, and does not need to be: in software he barely knows it got him to a usable result, and in software he knows well it removed tedious setup and handed back an editable file. This is his conclusion from ten tests, not a measured claim.
• A practical warning on unattended iteration: the model can critique its own output visually and loop until it meets a quality bar, but he chose not to automate that, because a loop set on a task the model is bad at 'can end up spending all day' and consume a large amount of tokens. He recommends capping the number of iterations.
• He reports computer use as a step change against earlier attempts, saying previous models he tested felt nowhere near this reliable and that this is the first he could give a real task in software he actually uses and leave to work. Reported from his own testing; no figures accompany the reliability claim.

Why it matters

This is the most careful piece of hands-on testing in this batch and it cuts against the prevailing framing of computer use. The headline capability — an agent operating ordinary desktop software by cursor — is real and reported working across ten applications, but the test finds it is the slower and worse route wherever a scripting interface exists, by roughly two to one on time in the one case measured both ways. That matters for anyone deciding where to point this: computer use is the fallback for software that offers no other door, not the preferred path. The timings are the other durable part — quarter-hours and full hours per task, not seconds — which reframes the feature as asynchronous delegation rather than acceleration, and the presenter says as much. His pushback on the Blender coverage is worth preserving verbatim in spirit: a creator with a commercial interest in AI enthusiasm saying the published examples misrepresent the capability is exactly the sort of framing that must not be flattened. Everything here is one operator's single run per task, on one machine, with no repetition, and the model and application versions will date quickly.

2hTtpFVZE5A-transcript.txt