?? AI's world models

In today's edition, we delve into?whether LLMs could be building some kind of understanding of the world, a “world model”.

?? Thanks to our sponsor, Masterworks, an investing platform that enables everyday people to invest in multimillion dollar paintings by artists like Banksy and Basquiat.

How does an LLM work? Simple answer: it predicts the next token according to an input, based on its training and reinforcement learning. But is there more to it than that? Somewhere along this learning process,?could the LLM be building some kind of understanding of the world, a “world model”? This concept of a world model is currently the subject of hot debate across the AI community. (I talked about the many disagreements in generative AI and how to deal with them in Friday’s essay.)

The latest salvo comes from MIT researchers Wes Gurnee and Max Tegmark , who recently published a paper arguing that

modern LLMs acquire structured knowledge about fundamental dimensions such as space and time, supporting the view that they learn not merely superficial statistics, but literal world models.

LLMs, the authors argue, have a world view because they appear to have a structured notion of space. There is a constant, stable structure of relationships underneath the statistical relationships.

But is this proof of a world model? And how could we tell? First, it depends on how you define a world model. The idea that intelligent systems (animals included) must have an internal model of how the world works is not a recent one. Nvidia’s Jim Fan offered a good summary of how our understanding of world models in AI has developed over recent years, one key aspect being that they capture causality and intuitive physics. An LLM would have to show more than predictive capacity, an actual causal understanding and ability to simulate the physical world. However, an understanding of, say, geography wouldn’t support an understanding of the physical laws and causality of the world. Gary Marcus explains that this appearance of having a structure to the knowledge doesn’t mean you have an understanding of the knowledge sufficient for it to be called a “model”. In fact, he shows that much more primitive AIs contain correlations about physical space.?

A lot of the internal machinations of our LLM helpers remain a mystery, but this kind of research, and most importantly the discussions it provokes, brings us closer to understanding them.?

See also: A study has found that LLMs provide useful feedback on research manuscripts, in some cases better than human feedback. Plus, the first wearable AI device was unveiled by Humane at the Coperni Paris fashion show.

??Today’s edition is supported by Masterworks.

Monet’s painting sold for $8 million… and everyday investors profited?

When the painting by master Claude Monet (you may have heard of him) was bought for $6.8 million and sold for a cool $8 million just 631 days later, investors in shares of the offering received their share of the net proceeds.?

All thanks to Masterworks, the award-winning platform for investing in blue-chip art. To date, every one of its 16 sales out of its portfolio has returned a profit to investors. With three recent sales, Masterworks investors realised net annualised returns of 17.6%, 21.5% and 35%.

Shares of every offering are limited, but Exponential View readers can skip the waitlist with this exclusive link.

SKIP THE WAITLIST?

*See a disclosure