Uncle Rester
unclerester.bsky.social
Uncle Rester
@unclerester.bsky.social
Also known as Frankentan.
Cryptid of a person. Kinda undead.
Slow of thought and life. Autistic and disabled pensioner. He/Him.
Trying to learn to draw.
Reposted by Uncle Rester
As an elementary school teacher, I've had lots of parents telling me their kids read at a "high level" and are reading "hard books" and yet when I ask their kid one key question about the actual text, you realize they're just reading words and have missed the whole point.
November 30, 2025 at 8:30 PM
Reposted by Uncle Rester
everything makes so much sense now
November 30, 2025 at 8:34 PM
Reposted by Uncle Rester
Unironically? Yes.
November 30, 2025 at 8:34 PM
Reposted by Uncle Rester
oh GOD is that why right wingers don't understand something unless it's explicitly in the text and act like you're making shit up when you say a character is idk trans or something?
November 30, 2025 at 8:33 PM
Reposted by Uncle Rester
"Are you capable of recognizing how choice-of-language can be used to induce an emotional response? And do you understand how that same emotional response can be used to influence your thinking, or the thinking of others?"
November 30, 2025 at 8:15 PM
Reposted by Uncle Rester
Guardian did a much better headline framing the “LLMs/AI guardrails are easily defeated” thing than sensationalism of AI WILL TEACH YOU TO BUILD A NUCLEAR WEAPON!:

“AI’s safety features can be circumvented with poetry, research finds”

www.theguardian.com/technology/2...
AI’s safety features can be circumvented with poetry, research finds
Poems containing prompts for harmful content prove effective at duping large language models
www.theguardian.com
November 30, 2025 at 5:42 PM
Reposted by Uncle Rester
Have you ever spent a lot (like years) of time on the internet talking to people about and searching for nuclear weapons related materials?

There is SO MUCH BULLSHIT out there that LLMs got trained on.

In addition to the whole making computers innumerate thing.
November 30, 2025 at 4:13 PM
Reposted by Uncle Rester
You can get me to help you “design a nuclear bomb” by buying me lunch ffs.

I do not assess public LLMs like ChatGPT as a nuclear weapons proliferation issue.

Hire a grad student instead of asking ChatGPT, the research will take longer but it will be more accurate.

www.wired.com/story/poems-...
Poems Can Trick AI Into Helping You Make a Nuclear Weapon
It turns out all the guardrails in the world won’t protect a chatbot from meter and rhyme.
www.wired.com
November 30, 2025 at 4:06 PM
Reposted by Uncle Rester
Also:

Who cares?

LLMs are trained on *waves arms*.

any state serious about building nukes will rapidly need info specific to circumstances or otw not public.

public LLMs can’t produce a new workable weapon design b/c they’re parrots not thinkers.

There is no “nuclear secret”
an elderly woman is pointing at a whiteboard with the words " it does n't matter " written on it
Alt: an older white woman in a colourful blouse & red scarf is pointing at a whiteboard with the words "it doesn't matter” in white written underneath her in the GIF.
media.tenor.com
November 30, 2025 at 4:04 PM
Reposted by Uncle Rester
They were *really* counting on the big Activision/Microsoft tentpole release being a success and proving that the hoi polloi don't care about AI usage. Being proven wrong is making it dawn on them that it *is* a bubble and the public will never accept this shit.
November 30, 2025 at 1:51 PM