Oops I did it again 🤭

#TotalSolarEclipse

Longer context is a game changer, now we're going to see the real power of in-context (meta-) learning!

Excited to announce SIMA, a general AI agent for games & 3D virtual settings. It marks the first time an agent has demonstrated it can follow natural-language instructions to carry out a wide range of tasks across a large array of game worlds, similar to how a human would play.

Interesting discussion with Ian Osband about what it means for a large model to be 'grounded'. I had no idea my take was controversial! Language is very useful for abstraction, but a world made only of language, and constrained only by human subjectivity can't be that grounded?

I should add that we also trained and tested on a variety of custom designed and research environments like Playhouse and the newly developed Construction Lab! Custom environments played a pivotal role in our ability to evaluate and analyze agent capabilities.

Ever wondered how your LLM splits numbers into tokens? and how that might affect performance? Check out this cool project I did with DJ Strouse: Tokenization counts: the impact of tokenization on arithmetic in frontier LLMs.

Read on 🔎⏬

At the poster for our work on 'Passive learning of active causal strategies in agents and language models'
twitter.com/AndrewLampinen…
Tue 12 Dec 5:15 p.m. CST — 7:15 p.m. CST
Great Hall & Hall B1+B2 (level 1) #825
nips.cc/virtual/2023/p…

Sadly I am not going to #NeurIPS2023 , but if you are, you should definitely check out Julian Coda-Forno's paper on Meta-in-context Learning in LLMs!

neurips.cc/virtual/2023/p…
on Thursday at 5:15 pm

x.com/cpilab/status/…

Can we teach AIs to be good partners to humans and each other?
Emotions and playful courtship seem to be key to successful partnerships in humans and animals. We built an evolutionary model to try to understand why. With Edgar Duenez-Guzman & Joel Z Leibo
pnas.org/doi/10.1073/pn…

The Gemini era is here. Thrilled to launch Gemini 1.0, our most capable & general AI model. Built to be natively multimodal, it can understand many types of info. Efficient & flexible, it comes in 3 sizes each best-in-class & optimized for different uses blog.google/technology/ai/…

👀 Gemini is out!

BBS currently has an open call for commentaries on our paper on “Meta-Learned Model of Cognition.” We encourage anyone interested in cognitive models, especially Bayesian and neural network models, and their applications to apply! cambridge.org/core/journals/…

Now that Q-learning is the latest hot trend, I'm looking forward to the AI community rediscovering policy gradient and actor-critic methods, and then finally meta-learning! (hint: it's now called 'in-context learning')

Everything old is new again.

Happy Thanksgiving to my American friends! 🦃

If you're looking for some light reading there's a non-paywalled version of our paper here: rdcu.be/drm0j