They're cloning you

Hi, and happy Tuesday.

Meta has been recording staff’s desktops - without their permission - to feed behavioral data to train AI models.

Zuckerberg justified it to staff, saying “it wasn't strategically in your interest for us to communicate everything.”

This came after Meta pivoted hard into AI, laying off roughly 8,000 employees (about 10% of the company) and redirecting resources toward building what it calls "Applied AI." 

Why do big tech companies want this data?

It turns out that the AI models can only do what they were trained to do.

In the first phase, they were trained on all the data on the Internet - and so acted like “Google on steroids.”

In this next phase,  to build AI agents that can actually do work - they need to be trained on data of people doing work.

Thus far, Anthropic appears to have pulled ahead by somehow magically sourcing such work activity data to produce Mythos / Fable class models that can perform multi-step actions.

But everyone is at it, and you can see it show up in different ways.

• OpenAI recently launched Record & Play: Using OpenAI’s Codex desktop app, you start recording and perform a task (e.g., uploading a document, generating an expense report). Codex watches your screen and keyboard inputs. You can then tell it to repeat the task later, and it will handle the process by itself.

  • https://www.youtube.com/watch?v=ZK3JhU73W18 

• Windows Recall in Copilot+ PCs takes continuous screenshots of your desktop activity. It allows CoPilot to find past apps, websites, or documents you’ve been working with. (Don’t worry, Recall is off by default)

  • https://support.microsoft.com/en-us/windows/ai/ai-features/retrace-your-steps-with-recall

While these are useful features, are your recordings also used to train AI models?

And what happens once this data has been captured, and the models trained on it?

• Anthropic recently launched Claude Tag, which lives in a shared Slack workspace or channel. Anyone in the channel can tag @Claude to delegate tasks, and the whole team can view its progress in the thread.

  • https://www.youtube.com/watch?v=VojDzHaciKQ 

Anthropic says 60% of their code is now written by Claude Tag.

Of course, if you’ve been following our work for a while - you’ll know we’re building an “AI colleague” for monitoring regulations, intelligence and other activity clients sometimes seek from Prescouter. (Reply to this email if you want to see a short Loom video demo).

Our view is that there is a lot of “grunt work” that nobody wants to do, so cloning this work is valuable. It frees our experts to do what only they can do - through insights gained from lived experience of challenging technical problems - rather than recording screens.

Hopefully, the means of getting there can be less questionable.

Best,

Dino