I think an interesting test of LLMs is their ability to engage in coherent puzzle creation. I asked Claude 3 and GPT-4 to create a complex puzzle for a D&D game. They both get very close, but get lost in the complexity, creating puzzles that are almost, but not quite, solvable.
RESEARCH
-
Supercomputer Simulations Reveal Galaxy Formation Secrets Over 13.8 Billion Years
By
–
Unlocking the Secrets Behind Galaxy Formation: https://
scitechdaily.com/unlocking-the-
secrets-behind-galaxy-formation/
… Astronomers used ~100 million hours on supercomputers to simulate the formation of galaxies from the Big Bang 13.8 billion years ago to the present day. -
Pluralistic Alignment: A Roadmap for AI Systems
By
–
A Roadmap to Pluralistic Alignment Sorensen et al.: https://
arxiv.org/abs/2402.05070 #Artificialintelligence #DeepLearning #MachineLearning -
Pluralistic Alignment: A Roadmap for AI Safety
By
–
A Roadmap to Pluralistic Alignment Sorensen et al.: https://
arxiv.org/abs/2402.05070 #Artificialintelligence #DeepLearning #MachineLearning -
Self-Attention Mechanism: How Key, Query, and Value Matrices Work
By
–
Jokes apart! It explains how key, query & value matrices combine in the self attention mechanism. Here’s an illustration of the same:
-
NLP for Misinformation Tracking: Framing the Research Challenge
By
–
I think there might be uses for NLP in tracking/disrupting the spread of misinformation, but today I'd say that we need to start with a larger frame (e.g. one provided by people who study that spread & the behaviors behind it) and then ask: where can NLP help?
-
LLM Information Reliability: Persistent Issues Beyond Hype
By
–
As a counterpoint to the "tech moves so fast/can't keep up" trope, I'd like to point out that many of the issues actually don't change. Here's me in 2019 arguing against proposals to use LLMs as a source of info about the world — and already worrying about the info ecosystem.
-
Sensory Streams Analysis: Underappreciated Touch Modality
By
–
This is a good analysis of sensory streams by @ylecun
. I love these back of the envelope estimates. Also, great point about touch, a modality that I feel is under appreciated. -
Generative Models Advancing Across Health, Weather, History, Mathematics
By
–
Had so much fun at this RAAIS day! We’ve come far since 2018 and generative models now doing amazing things in health and weather and history and maths and more. Generative models and probabilistic thinking still so important! Thanks for the great walk down memory lane
-
AGI Definition Confusion: Why Technical Precision Matters
By
–
The definition of AGI I use is "AI that can perform any intellectual task that a human can." But a few teams have come up with alternative definitions, so its meaning has become muddied and confusing, I now less it less esp in discussions that need technical precision.