Only hassle was setting up the Broadcom WiFi drivers which aren't upstream, but I did this before I would have trusted AI with this task and even then it was not a big deal. I think these days running Linux on those things is likely trivial.

If you wanna claim this is a skill issue and you are built different etc, fine. I guess we can just say C is fundamentally unsafe for _everyone except you_.

(This is like when people wanna argue about if self driving cars are safe or not and they're blathering on about technical capabilities and limitations. No. We have the data).

I do actually think that with modern C++ it's possible to get to a style of programming where memory unsafety is mostly a theoretical concern but not with C.

I suspect this is because they are usually introduced like "here are bitfields, but DO NOT USE THEM for anything that might become ABI because their order is implementation defined". I think that causes people to mentally file it under when actually there are many contexts where the ABI risk is not real.

Yeah I was thinking of prefixing my "memory corruption" with "software-bug induced"! I don't see a credible solution to Rowhammer. ("DDR[n+1] fixes it" - lol)

> an unsafe block, an FFI boundary

Honestly these feel solvable to me at this point! I think we'll see:

- unsafe code shrink as languages get more powerful

- amount of analysis we can apply to each unsafe line shoot up exponentially as AI gets cheaper

- amount of FFI we actually need shrink as it gets easier to just click "rewrite it in $lang" on the decision card when your coding agent says "I found a library for that but it's in a different language"

(Having said all of that, people seem to be adopting Zig for some bizarre reason... So maybe I'm naive to expect unsafe lines to shrink)

(But also, maybe AI gets so good and so cheap that we can just type "go fidn all the bugs andfi xthenm" into an LLM, between sips of a Piña Colada)

Nowadays I actually believe I'll likely see a world without memory corruption within my lifetime.

(Is it given that there's a play? Waymo is culturally a little distinct from Google so from personal experience I can't actually rule out that they're just doing this coz they think it would be good for the world).

The goal is "Linux on Mac". In a commercial product, specific details of what gets built usually stem from a more detailed vision that can be articulated as one. In open source projects those details are determined by what individual contributors are motivated to work on.

That doesn't means a cohesive vision can't emerge, there might be one for Asahi actually. But I think the default position is that it's kinda just a mob of people doing what they each feel like doing.

If you have any professional connection at all to Linux then $11/mo is an absolute bargain!

It's not something I expect, but I definitely don't see an issue with people putting a couple of sentences in an "expand to see AI usage notes" tab at the top or bottom of the page.

> Would you expect none?

As in, would I expect no AI usage? Coz no I absolutely expect people to use AI.

If you had got Claude to do this investigation for you then actually I think that would have been notable enough to include details of your AI workflow. But not because "it's cheating not to disclose it", rather because you'd probably have come up with quite a neat workflow that is interesting in its own right.

My main question was "did Claude do the investigation by itself or just write up the article from someone's notes?"

Also though, I'm quite willing to believe this is human written and the human just happens to have a Claudey style. The actual prose isn't that Claudey it's just the structure of how it presents ideas. But, Claude had to get that structure from somewhere. It's not that surprising to see people with that style of communication.

I would reject "Auto Mode is safe" as a message but FWIW I am totally on board with "on aggregate, making Auto Mode the default improves the safety of Claude Code compared to the prior status quo". Coz I would say in the vast majority of cases the access classifier is doing a better job than the thing it replaced.

Anyway yeah. Like I said, conclusion is the same: we should be decoupling this from the harness. We ought to be sandboxing agents the same way we sandbox applications. I wish Claude Code would make this path smoother :(

I don't really understand what Claude Cowork is. Sometimes I use it instead of plain Claude Chat for tasks that "feel big"? And when I've done that I've felt this was the right choice. But until I actually had that session up and running I would not be able to articulate what Cowork is or enumerate any specifics of how I expected it to behave. I just thought "maybe this is a Cowork task?"

(I have only used the cloud one. My understanding is the the local app is essentially Claude Code but for non-coders, which matches the description of the OpenAI equivalent).

Even as someone who's now used it a few times I wouldn't have been able to articulate as many details as Simon W does in this article, it's a weirdly shaped product structure.

To be fair to the authors they don't actually say it is. But then they contrast it with the "0.00% prompt injection attack success rate".

The upshot is kinda the same - this is still evidence that we should be sandboxing our agents. But it doesn't actually challenge Anthropic's "our models are too clever to prompt-inject" vibe.

I think the term "stake your reputation" is quite good for this situation. I.e. "hi new contributor, please acknowledge that you Stake Your Reputation before we move ahead with this code review".

(Doesn't help with legal aspects of course)

This scares me because the stuff that screams "THIS COMMENT WAS WRITTEN BY CLAUDE" has gotta be the tip of the iceberg. There must be much more of these comments that are better-prompted and using better-tuned model pipelines to make them less detectable. (Probably even just the lazy garbage is harder for me personally to detect if you just pass the prompt to GPT instead of Claude, it's just that I've read so much Claude output that I'm attuned to it).

So yeah suddenly dead internet theory is seeming a lot more real today than it did just a couple of months ago.