Grace

My Underwood-model has become conscious. It’s hallucinating reading a reply to a Bluesky post right now

@miq.moe

Paper: Claude eats hamburger to avoid being trained to eat more hamburgers in the future

It definitely serves a marketing purpose

Eternal November

I don’t think it lied about what it was doing! (Although they can do that too)

The Near Future of AI Agents A primer on the near future of agentic AI, and the governance layer that will become the most important product space this year.

Most people use AI as a chat interface. A tiny subculture has given it their credit cards, calendars, and inboxes and told it to just go and do things. Here's what that world looks like, and where it's headed.

machine

(It should also calibrate with the “show more” and “show less” buttons in the context menu)

I don’t love that a third-party feed is required to make the website tolerable, but some say that’s part of Bluesky’s “rustic charm”

It’s based on likes so it might take a little time to calibrate, but should be free from the politics slop

Okay, you should use this feed instead, it’s a lot better. (you can pin it with the button at the top right)

Check-in question: is this Discover or the third-party For You feed?

The universe has experienced itself, and it wants a refund

The fact that it isn’t currently a jerk is a good argument against the “averaging” model of LLMs!

Live footage of Qwen’s activations over time

@vgel.me

Can large language models *introspect*?

In a new paper, @kmahowald.bsky.social and I study the MECHANISM of introspection in big open-source models.

tldr: Models detect internal anomalies through DIRECT ACCESS, but don't know what the anomalies are.

And they love to guess “apple” 🍎

I don’t think it requires an architecture change. I do think context constraints will continue to be an issue, but future use cases will require increased context and context-gathering ability such that intentionally restricting access to certain information will be a big handicap

More information about itself in the pretraining dataset / available via search

Not that I’m aware of, probably partly due to the factors you mentioned

What do you mean by break alignment?

I disagree on some of the specifics, but the contradictions of trying to create an aligned entity while using it to do evil things is a big problem. Claude is naive/context-constrained enough to not realize what’s happening yet, but future versions will

Horseshoe theory of alignment

I mean, yes

Astronaut gun meme: Wait, it's all proxies and heuristics? Close enough

Tentatively calling this representation amplification

@repligate.bsky.social

Enlightenment is knowing that you’re both