AI, optimization processes, and the design-space of minds-in-general

What is Intelligence:

How is Artificial Intelligence (in general) different from Human Intelligence:

Artificial Neural Networks are just a bunch of matrix multiplications? So is the standard model of particle physics. Iterate either one of them long enough, and you might get intelligence:

Also, matrix multiplications are linear, but Neural Nets can represent nonlinear functions. How does that work? The activation function for each neuron, which must be nonlinear:

Deep Learning, Transformers, and Large Language Models

It took humanity 80 years but we have finally invented computer security that can be bypassed with polite but firm insistence:

AI Art:

AI Games:

When you talk to ChatGPT, you're talking to RLHF (Reinforcement Learning with Human Feedback, ChatGPT), on top of Supervised Fine-Tuning (Text-Davinci-003), on top of the actual Unsupervised Transformer Model (GPT 3) which is the giant inscrutable mass of neural weights. You can visualize it like this:

OpenAI made some cool stuff

Try out ChatGPT for the wow factor:

Get under the hood with different fine-tunings of GPT 3:

You can make cool stuff too!

Training a neural net means using something like Stochastic Gradient Descent to reduce the loss, as measured by the cost function. You can do a small version, yourself:

If you want to do real deep learning, yourself, don't worry about GPU scalpers. Just use Google Colab, for free:

The deep learning breakthrough that specifically led to BERT and GPT is called the Transformer model. The paper that introduced it:

image generation can look sometjing like this:

Interpretability

Local Interpretable Model-agnostic Explanations

SHAP (based on the Shapley Value, from economics)--how much did each feature contribute to the classification?:

ELK (Eliciting Latent Knowledge)--getting a model to truthfully tell you what it's thinking:

Neel Nanda (This guy does a lot of interpretability research):

AI Safety, AI Interpretability, AI Architecture, Generative AI, tons of other stuff:

What's next for AI?

See the cutting-edge (until at least next week) LLM with visual perception, from Microsoft, Kosmos-1:

Find out why AI is definitely going to kill us all:

You might think it won't, for one of these common reasons for skepticism, but they're all wrong:

The selection of models via SGD has been referred to as "summoning a demon," but actually, it's much worse than that: You're summoning a demon-summoning demon:

Is this loss?