@KathyReid it's an incredibly awful reflection on us that most of corporatopia is absolutely ripe for this - it's the perpetual Colony in action, even with meat robots pushing keys.

We could have done almost anything, and we chose to do this :/

@KathyReid my problem with all of this automation is, if people lose jobs, who is going to buy stuff?

The race to get rid of people at companies is so silly, capitalism can’t help itself 

@KathyReid

36 5 degree surveillance. Badum tish.

@KathyReid

And eventually all the humans have no job, no income, and the producers have no one to whom they can sell their goods.

Happily, climate breakdown will wipe us out before we get to there.

@KathyReid this implies a rather strange future dystopia:

Humans building new inefficient human-oriented UIs so the trained-on-2020s-human-behavior AIs can inefficiently act like humans and use them.

Which does sound like what the shareholders want, so... seems plausible.

@KathyReid

> It's not about *images*. It's about modelling what workers do on Windows, and then replacing them.

So like WOPR, but for cubicles?

@KathyReid - and then, after they've broken the world to build their Replicant Worker™️, they find out it isn't able to actually produce anything of value, because all it will be able to do is replicate the externalities of work. Like making a replica of a cake out of wood & insulation foam & expecting it to taste like the real thing.

We truly live in the stupidest timeline.

@KathyReid

AI training on porn, so much porn.

Prepare for some delightful AI PowerPoint presentations.

@KathyReid

This makes so much sense, there is not enough data for them for general purpose computing compared to -- let's say images and text.

@KathyReid
Something like this exists partially.
The last buzzword relating to this was AFAIK "Robotic Process Automation".

Scraped sequences of pressed buttons and filled-out checkboxes give you a great basis to train your LLM.

And of course, this approach has two "great" benefits:
a) You can tell your workers that you could replace them with AI, so you can pressure them to take less payment.
b) If you introduce it, you can demote your workers to mere "content checkers" who get all the fault and only part of the payment.

@KathyReid

There is absolutely zero doubt that MS would love to model human workers as you suggest.

However, from my understanding of how Recall is implemented, all that history data is encrypted & stored locally on the machine. The OCR is handled by a local Copilot instance. MS says they are not uploading any of that to their cloud. For now, anyway.

So assuming this is correct, how could they build *anything* with Recall data? I'd just assumed it was a workplace turbo-surveillance tool.