It's so important to remember that unlike code which can be reverted - most file system and application operations cannot.

There's no sandboxing snapshot in revision history, rollbacks, or anything.

I expect to see many stories from parents, non-technical colleagues, and students who irreparably ruined their computer.

Edit: most comments are focused on pointing out that version control & file system snapshot exists: that's wonderful, but Claude Cowork does not use it.

For those of us who have built real systems at low levels I think the alarm bells go off seeing a tool like this - particularly one targeted at non-technical users

Frequency vs. convenience will determine how big of a deal this is in practice.

Cars have plenty of horror stories associated with them, but convenience keeps most people happily driving everyday without a second thought.

Google can quarantine your life with an account ban, but plenty of people still use gmail for everything despite the stories.

So even if Claude cowork can go off the rails and turn your digital life upside down, as long as the stories are just online or "friend of a friend of a friend", people won't care much.

The first version is for macOS, which has snapshots [1] and file versioning [2] built-in.

[1]: https://eclecticlight.co/2024/04/08/apfs-snapshots/

[2]: https://eclecticlight.co/2021/09/04/explainer-the-macos-vers...

Once upon a time, in the magical days of Windows 7, we had the Volume Shadow Copy Service (aka "Previous Versions") available by default, and it was so nice. I'm not using Windows anymore, and at least part of the reason is that it's just objectively less feature complete than it used to be 15 years ago.

Q: What would prevent them from using git style version control under the hood? User doesn’t have to understand git, Claude can use it for its own purposes.

Indeed there are and this is no rocket science. Like Word Documents offer a change history, deleted files go to the trash first, there are undo functions, TimeMachine on MacOs, similar features on Windows, even sandbox features.

In theory the risk is immense and incalculable, but in practice I've never found any real danger. I've run wide open powershell with an OAI agent and just walked away for a few hours. It's a bit of a rush at first but then you realize it's never going to do anything crazy.

The base model itself is biased away from actions that would lead to large scale destruction. Compound over time and you probably never get anywhere too scary.

There's no reason why Claude can't use git to manage the folders that it controls.

TimeMachine has never been so important.

You make a good point. I imagine that they will eventually add Perforce-style versioning to the product and this issue will be solved.

Somewhat related is a concern I have in general as things get more "agentic" and related to the prompt injection concerns; without something like legally bullet-proof contracts, aren't we moving into territory of basically "employing" what could basically be "spies" at all levels from personal (i.e., AI company staff having access to your personal data/prompts/chats) to business/corporate espionage, to domestic and international state level actors who would also love to know what you are working on and what you are thinking/chatting about and maybe what your mental health challenges are that you are working through with an AI chat therapist.

I am not even certain if this issue can be solved since you are sending your prompts and activities to "someone else's computer", but I suspect if it is overlooked or hand-waved as insignificant, there will be a time when open, local models will become useful enough to allow most to jettison cloud AI providers.

I don't know about everyone else, but I am not at all confident in allowing access and sending my data to some AI company that may just do a rug pull once they have an actual virtual version of your mind in a kind of AI replication.

I'll just leave it at that point and not even go into the ramifications of that, e.g., "cybercrimes" being committed by "you", which is really the AI impersonator built based on everything you have told it and provide access to.

So the future is NixOS for non-technical people?

>>I expect to see many stories from parents, non-technical colleagues, and students who irreparably ruined their computer.

I do believe the approach Apple is taking is the right way when it comes to user facing AI.

You need to reduce AI to being an appliance that does one or at most a few things perfectly right without many controls with unexpected consequences.

Real fun is robots. Not sure no one is hurrying up on that end.

>>Edit: most comments are focused on pointing out that version control & file system snapshot exists: that's wonderful, but Claude Cowork does not use it.

Also in my experience this creates all kinds of other issues. Like going back up a tree creates all kinds of confusions and keeps the system inconsistent with regards to whatever else it is you are doing.

You are right in your analysis that many people are going to end up with totally broken systems

There was a couple of posts here on hacker news praising agents because, it seems, they are really good at being a sysadmin. You don't need to be a non-technical user to be utterly fucked by AI.

Not a big problem to make snapshots with lvm or zfs and others. I use it automatically on every update