I co-authored a book with a GenAI expert and I am slowly understanding the half that my co-author wrote. Check it out if you want to join me in the journey!

Pleased to see that 3 readers have answered correctly before the deadline! The answer is: vector-valued labels.

Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.

Will consider answers until Sunday 00:00 EDT!.

There is a kind of target variable that a neural network can predict but regression trees cannot. The first three who answer correctly will get our free book

At least some people like our fifth book "Kudos . for sharing their hard-earned knowledge and creating a practitioner resource that is timely, relevant, and ‘roll up your sleeves’ practical" - Jay Vyas, CSO, Firinne Capital; fmr Head of Research, Canada.

Deep Reinforcement Learning for Portfolio Optimization. Is it really better than our Conditional Portfolio Optimization method? See new blog post:

Applying transformer to multiple assets' features. A new blog post:

The gradient of an expectation can be expressed as expectation of a gradient - that's the essence of policy optimization in reinforcement learning as well as Monte Carlo optimization of neural nets. For the latter, see

Valeriy @predict_addict is one of the best AI experts out there. A great honour that he likes my book!.

Goodness! After reading my co-author @hamletjmedina 's explanation of attention and transformers 3 times, I finally understand them! So far his explanation is the best among many AI books I read:

Applying Transformers to Financial Time Series. A new blog post:

Some people question whether a "distilled" model can outperform the "teacher" model (as it was alleged that DeepSeek distilled OpenAI's GPT's output and outperformed it). We demonstrated that this can be done in our book

Though in fairness, I did publish a blog post recently at retesting some of the prompts in the book. We definitely see improvements in the responses.

You can treat Chapter 2 of our new book as a catalogue of GPT fails. It is breathtaking how many ways ChatGPT failed us. The day of AGI is far. .

I have a small quibble with @Ritholtz bestseller "How Not To Invest". Passive investing in indices isn't quite passive: you have to pick the index, and the index construction can be highly complex and discretionary.

The importance of p(x, y) and not just p(y|x): if feature x has a value never seen in trainset, we want to know about it.

"A Poor Person's Transformer": how transformer is used as a sample-specific feature selection method.

My co-author taught me what GenAI is about in a nutshell: supervised learning trains p(y|x); GenAI trains p(x, y), p(x), or p(x|y). See

It wasn't intentional that the acronym for our new book GATAM is pronounceable.

Have LLMs really improved over the last year? .Comparing their responses to our book's prompts a year later. Read our book's blog for more .