Here’s my new favourite illustration of the limitations of multimodal LLMs (here using the free version of ⁦OpenAI⁩’s ChatGPT 5).

Oh, wow. This is amazing.

Just a reminder that last November there were **1.7 million** layoffs in the US in aggregate. And that was a relatively *low* number. But there were 5.3 million hires that month too. In a labor market as large as the US there's just an incredible amount of churn all the time. 1/2

It's because although a LLM ultimately works as Input -> Latent Space -> Output and this paper only examines the first of those links, people are interpreting it to speak about both, or about the initial training of the model (which is necessarily lossy).

Read this.

I recommend people read the sequence of posts by Roger E. A. Farmer with a critical view of the NK model. I did not engage in the debate with them and many other great ones on this topic here on X just because my views are too nuanced for this forum (or, perhaps, I'm just too confused

Prettier than a data center, and it actually has a moat

Many, many matrices : ML/AI Very, very large matrices: Quantum Mechanics Many, many, very, very large matrices: Quantum field theory Learn linear algebra (and functional analysis)!

The discourse around macroeconomics would be a lot easier if everyone understood the distinction between: (1) causal expressions of choices made by agents in the economy; (2) equilibrium relationships; and (3) accounting identities.

I wonder what the probability is that while the US federal government is shut down, (1) the SCOTUS rules against the US government’s IEEPA tariffs *and* (2) a stock market crash occurs.

Personalisation setting to make ChatGPT usable.

In all seriousness, genuinely shocked there wasn't a higher dropout rate, but wow what a result!

5%

George Selgin Just once, I’d like the “Loans create deposits and that’s all there is to it” crowd to ask themselves why banks don’t issue twice the quantity of loans per month, or what the central bank might do if they did.