nokcha

I'd guess that prompt injection is the biggest risk in this setup, dwarfing the risk of exploits against the inference engine. Personally, I run LLM agents only inside a Docker container that limits the LLM's access to sensitive information and the LLM's ability to take irreversible destructive actions.

See also: https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/

> LLM generated content often (perhaps by definition) seems acceptable to LLMs.

In my experience (albeit with non-coding questions), ChatGPT 5.2 is often quite eager to critique snippets of its own replies from previous conversations. And reasoning models can definitely find flaws in LLM-written code.

I'd argue that environmental regulations that impede building modern nuclear power plants to replace coal power plants are net harmful. Nuclear power safety has advanced a lot since Chernobyl.

For non-broken cryptographic hashes (e.g., SHA-256), the false-positive rate is negligible. Indeed, cryptographic hashes were designed so that even nation-state adversaries do not have the resources to generate two inputs that hash to the same value.

See also:

https://en.wikipedia.org/wiki/Collision_resistance

https://en.wikipedia.org/wiki/Preimage_attack

> machinery, toys, lifts, medical devices, and vehicles

All of those seem high-risk except toys.

It's the other way around: with the new change, the blockee will be able to see (but not reply to) the blocker's posts, but the blockee's posts will still be hidden from the blocker.

> ultra-processed foods such as ... yogurts

> Frozen and canned vegetables are often classified as ultra-processed

I think this goes against the common usage of the term "ultra-processed".

ChatGPT is arguably a general intelligence, but it hasn't escaped AFAIK -- it's still contained within OpenAI infrastructure, and OpenAI can easily pull the plug on it.

Most individual shareholders (especially those who just buy index funds) have neither knowledge of corporate wrongdoings nor, in practice, power to prevent them. A better model would be holding accountable the particular individuals with power and responsibility, such as in the Volkswagen emissions case where Germany criminally prosecuted executives involved in the fraud.

I've heard that recent changes to reduce sulfur emissions from maritime shipping have caused a sharp increase in global warming:

https://twitter.com/RokoMijic/status/1674697743380430849

https://twitter.com/RandomSprint/status/1679851033424547840

The "software" part of an LLM is pretty trivial -- the interesting piece is the the weights. Since the weights are mechanically generated by a computer, it can be argued that the weights are not copyrightable, just like a photograph taken by a monkey isn't copyrightable.

I guess there might be a distinction between "development" of software and "maintenance" of software (e.g., bug fixing, refactoring, etc.)?

I imagine a VPN is sometimes still useful when torrenting, to avoid copyright complaints to your ISP.

Highly relevant:

https://AstralCodexTen.substack.com/p/the-media-very-rarely-...

> #4 ... "Hacking is Cool" is a really dumb idea.

This has aged poorly; nowadays, the most notable attacks are conducted by state actors (e.g., Russia and China) or for-profit criminal groups (e.g., ransomware) rather than lone hackers doing it for fun.

In regards to the difference in effect between Adderall and illegal street meth, Scott thinks the biggest difference is the dosage size, effectively around 50x higher for street meth: https://astralcodexten.substack.com/p/drug-users-use-a-lot-o...

Dosis sola facit venenum.

Question is a bit ambiguous whether it refers to a 'second-preimage attack' or a 'collision attack'.

'Second-preimage attack': You are given a message, and your goal is find another message that hashes to the value as the given message.

'Collision attack': Your goal is to construct two messages (of your choosing) that hash to the same value.

Oh, I guess I was thinking mainly about online censorship by tech platforms. I think most censorship from the right nowadays is about public schools?

>I love it because it's fast and minimalist.

Minor quibble about this: On large pages, it can be quite slow. E.g., on https://astralcodexten.substack.com/p/ivermectin-much-more-t... , it takes a few seconds after initial loading for it be responsive to pressing the PgDown key, at least on my machine.

> "Anti-censorship" is hugely unpopular with the mainstream intellectual community these days.

I'd disagree. It's unpopular with left-leaning mainstream audiences, but among intellectuals who write for a living, I imagine that extra assurances of not getting kicked off the platform are rather attractive.

For Twitter, perhaps a good compromise would be: instead of banning problematic accounts, set to them a "default-mute" state where only people who choose to follow the account can see its tweets.

See also Section 1.F of https://www.journaloffreespeechlaw.org/volokh.pdf , which distinguishes how Twitter's 'hosting' function might be treated differently from its curation functions.