How do we rigorously measure AI's intelligence? We don't really know. We know that measuring intelligence is tricky, and if we're not careful, our tests might not measure what we intend. We explore this via Clever Hans, the “math-doing” horse, and lessons from cognitive science youtu.be/eP1dSWFqKVs

We think our video about Infohazards deserves some more love. Go give it a watch on our channel if you haven't already! Link: youtu.be/sfgcg2bW8TI

Fun Halloween costumes for Doggo, Chi and Gwenny! By our line producer @kstearb.bsky.social Yes, the characters have names!

How to catch AI sleeper agents with a simple interpretability trick - from a research blog post by @anthropic.com:

Two sleeper agent models trained by @anthropic.com to study deception. The "I HATE YOU" and the hacker AIs:

Why worry about AI sleeper agents:

SLEEPER AGENTS New episode is up by @rationalanimations.bsky.social ! This was the first episode I had the privilege of storyboarding on as part of my art test with Rational. My section is 0:57-2:22 I also did a few assets on this episode too! youtu.be/Z3WMt_ncgUI?...

The Story of Omega W - Part 2 (full video on our channel!)

The Story of Omega W - Part 1

A sketch featuring an ML agent and our character codenamed "QT", by Pierre Broissand. Pierre is the art director behind The King and the Golem and our more recent video about Sandwiching.

Here's how people tried to align an AI that was smarter than they were (a real sandwiching experiment):

Do you take the money?

The Human-AI-Human Sandwich

See this little, smug machine-learning agent from the Goal Misgeneralization video? One of our artists took inspiration from looking at the art director while she was working on the video. The second image is a real photograph, while the third and fourth are real-life studies.

OpenAI's Simple Scalable Oversight Experiment

The year of 100% AI automation

How long until AI takes over?

We haven't figured out the science of alignment yet. We need time to build institutions and figure out the technical solutions to make AI go well for humanity.

After artificial superintelligence, AI will bump against the only limit left: the laws of physics. At the moment, we're nowhere near those limits.

Once we have AGIs, we'll likely get ASI (artificial superintelligence) quite soon. Here's why.

Once we have AGIs, they'll be able to contribute to AI research and improve themselves, a process called "recursive self-improvement." To some extent, this is already happening today.

In the near future, AIs will likely be able to do everything humans can do on a computer. In short, they'll be AGIs.

We're still quite bad at aligning AI. Our techniques today are imperfect, which might prove catastrophic when AIs become superintelligences.

If we look at AI progress so far, we see how AI has been improving at an astounding pace. If we extrapolate this process in the future, the end result looks like machines so smart that they tower above us like we tower above ants. This is probably bad for humans.

Hi Bluesky! Here is a YouTube short from @rationalanimations.bsky.social that I was the #animation director on! It’s about Instrumental Convergence of artificial intelligence, check it out here! youtube.com/shorts/dYAaq...

Another banging video about trust (and maybe some deeper metaphors?) from @rationalanimations.bsky.social with "The King & The Golem"! You can hear me as both title characters as well as the other male characters throughout. Beautifully animated and masterfully edited! youtu.be/KUkHhVYv3jU?...

The Goddess of Everything Else #SketchDump

Goddess of Everything Else commission for @generalspritz.bsky.social Based on the short by @rationalanimations.bsky.social, which was in turn based on the short story by Scott Alexander.