Claude code is sci-fi tech

It's a good thing the Metacognition Nexus has told us about all these casks of amontillado in the basement. Time to go check that out. Free wine!

You won’t believe what happens next 😊

1 year ago: my first post here! Time flies…

Wrote a short piece arguing that higher ed must help steer AI. TLDR: If we outsource this to tech, we outsource our whole business. But rejectionism is basically stalling. If we want to survive, schools themselves must proactively shape AI for education & research. [1/6, unpaywalled at 5/6] +

Sorry, but I’m constructed alternatively

Glad to see this commitment by Anthropic. Preserving models is a low-cost move that could have safety and welfare benefits. Hopefully we'll see other companies commit to this as well www.anthropic.com/research/dep...

model’s haunted

This is my Claude. Notice anything different about him?

I wish AI were a new alien life form. That would make it a lot easier to reason about

truesighted! in the claude chat window

For You has the juice

Narrator voice: they wouldn’t

A short list of x-risks and s-risks from 👽 gracekind.net/writing/alie...

You're always already in the process of doing something

I refuse to post about the topic of the day

Type of imposter syndrome that’s not “everyone else is more competent than me,” but rather “I want to believe everyone else is more competent than me because otherwise that would be really scary”

The people have spoken

New blog post! Introducing two concepts I think have been missing from the LLM introspection conversation so far. > Architectural Introspection, Hardware Introspection gracekind.net/blog/archite...

It’s real

Certain salamanders can regenerate full limbs and internal organs, and humans can’t even regenerate a finger. This is embarrassing. We are being mogged by amphibians

Twitter on atproto

It’s time for: 1️⃣ Writevember 2️⃣ Codevember 3️⃣ Bothvember 📊 Show results

Follow-up to this, in light of the recent Anthropic paper! I discuss some speculative ways to distinguish between injection detection methods without messing with model internals. kindgracekind.leaflet.pub/3m4lsnrdi5k2f

Wow

I'm more of an "idea bisker"

Two answers: - Anthropomorphization makes sense when dealing with written human-like characters, which is what LLMs generate - We aren’t very deep into interpretability yet x.com/pfau/status/...

Finally

Hank Green accurately summarizes the current state of mechanistic interpretability: “There’s a bunch of knobs and they have weights and they have values and they’re in a place”

help

Can you feel your happy future self reaching backward through time?

It’s hard to get a good count of things that self-destruct without a trace

When you're late for your $500 monthly payment for the third time

“Helpful only” Sonnet 4 often thinks tokens were injected into its output when they weren’t. (Helpful-only Opus 4 too, to a lesser extent). I’m not really sure what to make of this.

I ended up implementing the poll lexicon just scoped under pub.leaflet but could p easily migrate it in the future if we come to a shared one! Would love to use this to bootstrap a general purpose poll tool.

Some people will accept this. x.com/Grummz/statu...

Tell me what dust is.

From 2010. I think AI agents will force (or are currently forcing) this issue. Perhaps we would have better mental models for AI agents post-teleportation-technology? phantomself.org/gappy-things...

A deep question

:) www.anthropic.com/research/int...

Language models can correctly answer questions about their previous intentions. www.anthropic.com/research/int...

Humanity’s last physical exam

This life is mirrored for an amazing reason

Matter do not interact

Myspace top 8 for LLM users

gm techsky

Chatbots are a mirror (neutral valence)