Rowan Tommins

RE: @antopatriarca@mathstodon.xyz

At some point, the PHP project is going to have to adopt a policy on LLM use.

I am not looking forward to the flame wars, and honestly am not 100% sure what variants I would be willing to support.

But, if anyone has the fortitude to start the ball rolling, I think "allowed once you have a proven track record" is an interesting option to consider.

#PHP Bound-erased #generics RFC has officially been declined. Bracing myself for more terrible takes that everyone who voted No is an idiot who hates types and is killing PHP or whatever.

Still hopeful that someone will pick up on it and deliver a version which doesn't involve lying to readers of the code about what type something is.

Using hidden text in an LLM's token stream as "guardrails" is exactly as secure as using front-end JS for validation.

Security restrictions have to live *outside* of the thing you're restricting.

RE: @mttaggart@infosec.exchange

"I can distinguish my own thoughts from your speech without effort; they arrive through completely different channels with completely different sensory signatures. But for an LLM, everything arrives through the same channel as one long token soup. Its own thoughts sit next to your instructions, which sit next to the contents of a random webpage it just fetched."

This seems like one of the biggest hurdles both to reliable deployment of AI agents, and to the emergence of AGI from this technology. At some point, we need to move away from just making increasingly complex text streams, and come up with a technology which separates cognition and perception.

@tripplehelix@tech.lgbt There's a difference between "an expert human would never have found that" and "an expert human would have taken much longer to find that". Anthropic et al like to imply the former, but Mozilla are saying the reality is the latter.

At the Titanic Museum in Belfast (well worth a visit) there's an exhibit showing the notebooks filled with numbers used to manually calculate the engineering for ocean liners. Electronic computers can do the same calculations in a fraction of the time, but they're not doing anything *impossible* for a human.

I read Mozilla's post as basically saying the same thing: with enough funding and time, their human researchers would have found every one of these bugs - and so would attackers - but new tools can massively accelerate the process.

@reichenstein There's a bit of a false dichotomy here, of Word vs MarkDown.

They happen to be examples on opposite sides of a number of different debates that have been raging for decades: proprietary vs open, machine-first vs human-first, WYSIWYG vs visible-code. Not to mention presentational vs semantic, where they're both broadly on the same side.

But they're not the first or only formats in their respective categories, and not the only two categories that can exist.

Even within its category, I don't think MarkDown achieves its goals well. It does decently for really simple formatting, but the core design has no built-in way to extend to anything more complex. Instead, it relies on ever longer strings of unmemorable punctuation like ; or just giving up and using HTML.