This is scary and it affects the degree to which I invest in building Claude-specific tooling, either code or in my brain. You can never guarantee that a dangerously-skip-permissions session is going to stay on the rails, what flags it might trip while you're not looking.

I wonder if Anthropic realizes the chilling effect this kind of event has on developers. It's not just the ones who get locked out -- it's a cost for everybody, because we can't depend on the tool when it's doing precisely what it's best at.

Personally, I am already avoiding Gemini because a) I don't really understand their policy for training on your data; and b) if Google gets mad at me I lose my email. (Which the author also notes.)

I believed the original story. Now I don’t. So it helped me.

It will never happen, but I'd love to see the NYT follow up their story and pit some of what Graham says against their cadre of experts and see what parts of the story they agree on and which ones they don't.

I would think the people at the Times would want to know if they are just being useful idiots here.

Last time the NYT needed to correct a major story (the 'starving children in Gaza' turned out to be a boy with a genetic abnormality) they issued their correction on the '@nytimespr' X account.

https://x.com/NYTimesPR/status/1950311365756817690

I use the Gemini mcp and my CLAUDE.md has instructions to consult Gemini for strategy. Seems to work well. I have the lowest Claude plan so I don’t know how this would work vs Opus, for example.

Separately, I have been meaning to implement a cheating detector — have run into Claude modifying problem statements, adding axioms, etc.

Okay thanks I'll try that.

> have run into Claude modifying problem statements, adding axioms, etc.

Same here. I've thought about creating a utility that tells Claude it has to keep going until a test exits with nonzero status. But I'm concerned Claude would just fake everything to make the test pass.

Do you need an MCP for that? Why not tell Claude to call the Gemini CLI?

You could do that -- my understanding is that MCPs give Claude less to "think about" than having to use another program correctly, and therefore avoid context clutter. I could be wrong.

He switched out his laptop.

I know this was a throwaway, but finding prior art for a large group of existing patents would be a cool application.

it was half-serious.

if LLMs arent being used by https://patents.stackexchange.com/ or patent troll fighters, shame on them.

It is not. That approach was tried and rejected, hence the problem.

In the interest of fairness, it should be noted a very specific solution was tried and rejected (providing configuration variables to adjust the behavior of `register-read-with-preview`). I personally don't see anything preventing users from configuring alternate commands which call directly `set-register` and mapping those as they please, so in that sense it is still configurable. And in the extreme case you could even use advices, etc. It is very difficult to make something non-configurable in Emacs and I don't see the register system being so ingrained as to be one of those things.

Whether it's a good idea... well, I don't use registers but the whole thing seems like a bad idea to me. I don't think I know anyone who uses registers and would like the new behavior. (But I also understand the desire to modernize `register-read-with-preview` a bit. A better solution seems like it would be providing multiple implementations so users can do the register equivalent of `(fset #'yes-or-no-p #'y-or-n-p)` like every old Emacs user does today.)

If this were true you would expect an announcement effect for disclosures. Testable.

Don't pay fees.

Ironically for stuff like bank fees they require a minimum balance, so you're paying fees because you can't save as much and can't save as much because you're paying fees.

Russia has committed so many own goals in this war that the fact that it is against Russia's interests to blow up Nord Stream actually increases the probability that they did it.