mastodon 4.7.3
Just got a text from one of my kids that they're at an invited talk (at their university) by @emilymbender@dair-community.social -- right on.
RE: @grimalkina@mastodon.social
This is excellent. Having seen Cat speak elsewhere, I wish I’d been in the room for this talk in order to get the spoken parts in addition to the slide content.
Not only were there valuable insights about the surface topic, the framing of effective learning resonated with me as an active teacher of a physical (not technological) discipline as well.
This post by @mttaggart@infosec.exchange is one of the most thoughtful posts by a programmer on GenAI that I have seen in a while. Wrestling with our own conflicting perspectives and feelings about this is hard work, and important.
I saw a paper (by some GenAI skeptics?) that concluded "we find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average."
Oh, wait, it wasn't some anti-AI authors. It's from Anthropic.
@rckenned@hachyderm.io Trick question! The model doesn't know anything other than "here are more tokens" and so it can't know that what it gives you is patented. Arguably, both the creators of the model and the users of the model might be infringing, in different ways, but both sides of that are as yet untested.
This is Meta's head of AI "safety" doing more than making a “rookie mistake" as she put it.
The biggest failing here, imo, was not letting a probabilistic chatbot have full control over her email account. That's bad, but even worse is anthropomorphizing the LLM.
When it started to "misbehave" (remember, they don't have any real sense of true/false, and "rules" that you give them are just more text for the blender) she started pleading "don't do that" as though it was a person. It's a computer program. You can just turn it off.
This is part of the danger of using the intentionally misleading language that everyone has been pushed into using about these, such as (but definitely not limited to) "hallucination." Even people whose full time job is about this problem get tricked into forgetting what these are, and what they are not.
https://www.404media.co/meta-director-of-ai-safety-allows-ai-agent-to-accidentally-delete-her-inbox/
@neil@mastodon.neilzone.co.uk We tried to find a balance on how attestation could be meaningful without degrading user control of their own systems when we looked forward in the direction of this problem about 20 years ago.
(paper was published about 4 or 5 years after we wrote it)
In general, it can be a complicated problem with semi-competing interests.
Principle 3 in the abstract is about the user's ability to constrain disclosure during attestation.
https://www.researchgate.net/publication/220066790_Principles_of_remote_attestation
@joel@gts.tumfatig.net If a teacher over-focuses on “how” and isn’t prepared and willing to get into “why” then I don’t expect that they have much depth to share. If they happily go into why they do things the way they do, then even if I disagree or do things differently then I am inclined to listen to see what might be worth learning or at least thinking about.
He does like to make predictions. They don't tend to be accurate.
Just one recent example:
https://futurism.com/six-months-anthropic-coding
Those baseless predictions have a very real cost:
https://hbr.org/2026/01/companies-are-laying-off-workers-because-of-ais-potential-not-its-performance
It could be interesting to start calling the kind of abusive scraping, theft, and regurgitation that all the major LLM-bros base their whole product around as "distillation attacks."
If someone says they meant something else when they made up that term, the conversation about describing the distinction ought to be fun.
https://www.theregister.com/2026/02/14/ai_risk_distillation_attacks/
What they're angry about (and asking the government to prevent!) sure does look a lot to me like what they themselves have done to... everything else on the web.
