If you devote yourself to a pursuit, spend years in devotion even, and get to a place where you begin to realize the outcome, it can be hard to explain to someone asking you to just give them the outcome that you can't give it away, they have to do the whole devotional thing too.

One of the interesting potential future directions is the use of Temporal Language Models (TLMs) for forecasting. Really spooky if that ends up working.

I have an explainer section on TLMs in the final part in my series on CalCo's Computational Semiotics lectures coming out soon, but in the meantime you should click in and read this thread and his paper he linked in it:

all my mutuals have delusions of grandeur. And honestly? That’s so valid.

The Dreyfus brothers, Stuart and Hubert, both worked at RAND in the early 1960s. Stuart was a mathematician on the control and optimal-control side. Hubert was a philosopher at MIT, and Stuart helped bring him in to evaluate RAND's artificial intelligence research. Hubert wrote a paper in 1964 that called AI a kind of alchemy ("Alchemy and Artificial Intelligence"). Newell & Simon sharply criticized it, but RAND released it in 1965 anyway. Hubert's objection was that early AI researchers were mistaking their success at formalizing narrow problems for making progress toward human-like intelligence. He used this analogy for the mistake: a man tries to reach the moon by climbing a tree. Every branch takes him higher and he can point to his progress. But climbing will never get him to the moon. His mistaken assumption is that he is on a viable path. Meanwhile, Stuart was programming RAND's JOHNNIAC, and coauthored "Applied Dynamic Programming" with Richard Bellman. His key contribution to control theory in 1962 showed that the necessary derivatives for a multistage optimization problem could be computed recursively using the chain rule. While Stuart's procedure sat in control theory, it would later be applied to neural networks. In 1969, "Perceptrons" by Minsky & Papert contributed to a long freeze in neural net research. The same year, Bryson & Ho's "Applied Optimal Control" printed the gradient procedure Stuart had simplified, which is mathematically equivalent to what neural-network researchers would later call backpropagation. Stuart didn't spell out the connection to neural nets until 1990, in the "Journal of Guidance, Control, and Dynamics." So Hubert would become one of AI's most famous philosophical critics, while Stuart helped develop the mathematics that would eventually enable us to train it. The brothers were not merely in intellectual opposition to one another, though. They coauthored "Mind Over Machine" in 1986. When they revisited the moon/trees analogy from Hubert's "Alchemy and Artificial Intelligence" in this book, they actually attributed the line to Stuart. The brother whose math opened a new path for AI may have been the first to say the old one was a dead end.