ratorx

Can you give some examples of where Imperial works better from first principles and why you think so?

Yep, but the certs.pem file only contains the root and does not know anything about intermediates.

So the only way to fix apps that don’t handle revocation of intermediates (eg the ones that only look at certs.pem) is to revoke the root.

The issue might be Root CA vs intermediates. The root CA that eMudhra rolls upto may be used for other intermediates which were not compromised or revoked. Removing the root CA will mean that these valid certificates no longer work.

This is technically a problem in the CLI - it needs to support revocations properly, not rely on root CA verification only. I’m not sure though whether there’s a standardised mechanism (that Apple supports) for CLIs.

> These steps affect every tool that uses /etc/ssl/cert.pem. Afterward, tools that use only this file might not be able to verify some websites whose certificates chain to these root CA certificates through intermediate CAs that weren't affected.

Being pedantic, the risk/benefit of remote perspectives are theoretical and dependent on the destination network also having multiple independent links. Otherwise any “uplink” ISPs are all SPoF regardless of anything Let’s Encrypt does because the traffic must traverse them to get to the destination.

But remote perspectives are still worth doing, Let’s Encrypt should not be the “less diverse” than the infrastructure of their typical certificate requestor.

I don’t think looseness is always more comfortable. You need looseness in the right areas which is what tailoring provides.

Maybe it’s cold and the extra airiness is working against you.

Maybe the clothes rustle and rub more when you move.

How do (non-elastic) pants stay up, without some minimal measurement for the waist?

I own some more baggy T-shirts and layering with a sweater feels more uncomfortable because the clothes will fold in places you don’t want them to with a sweater on, which is not a problem for more trim fit clothes.

I didn’t like the article because all of the missing defense-in-depth necessary to prevent the intrusion is firmly in the category of traditional cyber security. I could sympathise with an argument about organisational pressures and lack of time/people leading to an insecure sandbox, but that doesn’t seem to be the argument they are making. People have been securing systems that can run untrusted code for a very long time, and whilst mistakes happen, there are not really many novel ones.

The problem of “how to stop the model from attempting to be a hacker” is probably a more interesting and novel one, but it does not need to be solved to prevent the incident.

I think (for now), it is still relevant. A language is an abstraction, and a good abstraction, like a good LLM harness, can be quite valuable.

Let’s say I’m writing some concurrent code with an LLM. I’d probably feel much safer having it write Rust, rather than C. So even in a post-LLM world, languages will continue to evolve as long as abstractions can be improved.

Oops should have RTFA before commenting, I thought the CNAME was only to your servers.

Out of curiosity, how is MPIC relevant? Not that familiar with it, but CNAME would resolve to your server regardless no?

Yeah, I don’t think it assumes zero knowledge, which is what I’d expect from a “first programming language” book.

But it does say it’s targeted more for an “algorithms” course in the introduction.

As a side note, I think a REPL is the best environment for learning to program for absolute beginners. You get instant feedback and can focus entirely on just short snippets of real code, without having to worry about compilers, build systems, and “the OS” as a whole. Especially if it comes with a GUI wrapper, you can skip the terminal entirely.

I think it’s more of a time vs code tradeoff, if done properly.

For example in the Vec case, you could theoretically build an alternative which encodes the “three sections” property internally, and ensures correctness at construction time for the pointers. Not as completely safe as a Vec, but you can still get similar benefits for the “business logic”.

But I agree, just having a custom structure that does not provide a safe wrapper around this would be sacrificing standard guarantees.

The existing, human standards are based on human capabilities. Humans may not have read all the papers that could be cited. However, all these papers are almost certainly in the AI’s training set and it potentially has the ability to retrieve them over the internet into its context. Why shouldn’t we expect the AI to cite more comprehensively?

It seems more like a goal-setting problem, than an actually novel one. But a paper with lots of citations is not as “cutting edge”, so maybe doesn’t make sense for marketing reasons.

Putting aside the WASM sandboxing (I’m not familiar enough with it to understand how sandboxing works) there’s a DoS vector at least. Even regexes have had many DoS issues, and I can’t imagine WASM being easier to sandbox for DoS risk.