Presented by Chris Goad

Location: Mim's, 685 Gonzales Road, Santa Fe, NM

Where to park: either on the street, or as far up on Mim's gravel driveway as you can get.

I added some text  to what appears below on Tuesday, March 3rd. The new text appears as point 1)  below.

History:

Distant past (50s-80s): GOFAI (Good 0ld-Fashioned AI) - based on attempts to automate logical and mathematical thinking.

2009: Start of the age of artificial neural networks (ANNs) .ANNs have been around since the 50s, but did not become mainstream until their successes in computer vision in 2009.

2017: The transformer architecture is introduced by Google.

2022: ChatGPT is released.

2021, 2022: Text-to-image products are introduced by OpenAI (DALL-E) and Stable Diffusionn based on latent diffusion technology.

2026: Transformers (and other modern AI techniques) enter robotics.


Points I want to make:

1) Much modern AI technology, such as the ANN and the transformer, is opaque. ANNs and transformers are structured as layers of nodes where the connections of the nodes in one layer to those of the next are numerous. The deeper a layer is in this structure, the farther it is from either input or output, the less likely it is that any human can explain what the node states mean in that layer. Modern transformers often have more than a hundred layers, and what goes on in their central layers is utterly incomprehensible. ANNs and transformers are typically trained in billions of steps, and that is how they achieve incomprehensibility.
 

Of course the engineers who build AIs understand exactly how individual nodes and connections work. But, to make an analogy, just because you understand collisions between individual air molecules doesn't mean you understand aerodynamic turbulence.

2 ) It does work, sometimes spectacularly, but not always.

3) The Large Language Models (LLMs) underlying chatbots are the worst of AI because of the contradictory training data (human writing). AlphaZero, AlphaFold and now robotics are the best of AI since they rely on real-world training data.

4) Artificial Intelligence bears little resemblance to human intelligence. Alien Intelligence is a better term for it.

5) Artificial General Intelligence (AGI) is an ill-defined term and is of little use.


Readings and viewing:

First articles which give the flavor of modern AI:

A few new AI techniques

Agentic AI

AI begins its ominous split away from human thinking

Next a video from the Chinese company UniTree Robotics which illustrates the current state of the art (your jaw will drop to the floor).