Thoughts about Scientific Machine Learning and AI – Michael Green

Categories

All (32)

accountability (1)

ai (3)

AI alignment (1)

AI policy (2)

alviss (1)

attention (1)

autonomous agents (1)

bayesian (11)

bimodality (1)

cacophony (1)

causality (1)

chemistry (1)

cloud (1)

coco (1)

confidence intervals (1)

conjugacy (1)

constrained optimization (1)

corona (1)

correlation (1)

credible interval (1)

deep learning (2)

deep-learning (1)

derivative (1)

differentialequations (1)

dirichlet-multinomial (1)

drug discovery (1)

econometrics (1)

elasticity (1)

energy (1)

epidemiology (1)

europe (1)

european (1)

European AI sovereignty (2)

european sovereignty (1)

evolutionary (1)

foundation models (2)

frequentist (2)

from-scratch (1)

frontier models (2)

gradient (1)

hessian (1)

human-rights (1)

hypothesis testing (2)

inference (2)

infrastructure (1)

international-law (1)

jacobian (1)

JEPA (2)

jepa (1)

julia (3)

language models (1)

laplacian (1)

learning (3)

likelihood (1)

llms (1)

logd (1)

logp (1)

logs (1)

machine learning (2)

marketing mix modeling (1)

marketing-science (1)

math (1)

maximum likelihood (2)

measurement (1)

ml (1)

mmm (1)

model (1)

modeling (6)

molecular properties (1)

multinomial (1)

neural networks (1)

neural-networks (2)

nextcloud (1)

non-linear (3)

object detection (1)

open source AI (2)

optimization (2)

p-value (1)

pandemic (1)

personal development (2)

physics (1)

pKa (1)

planning (1)

posterior (1)

price (1)

prior (5)

programming (1)

rant (1)

regression (8)

robotics (1)

sample efficiency (1)

self-supervised-learning (1)

significance (1)

SIGReg (1)

simulation (2)

statistics (1)

stochastic differential equations (1)

technology (1)

time series (1)

transformers (1)

uncertainty quantification (1)

variable selection (1)

virus (1)

world models (2)

world-models (1)

xi correlation (1)

I’m Michael Green, a physicist turned AI founder. I write about Bayesian modeling, uncertainty quantification, drug discovery, and the occasional rant about why Europe should own its own tech stack. If your model can’t tell you what it doesn’t know, it probably shouldn’t be making decisions.

I run two companies: Alviss AI (Bayesian marketing mix modeling) and AI Alpha Lab (AI-driven equity fund). Below you’ll find posts where I try to make principled science accessible, and call out the stuff that bugs me along the way.

The Two Pillars, Revisited: The Moat Rotted Faster, The Choice Didn’t

Kimi K3 closed the composite gap to three points. Europe still has no frontier lab of its own. The first part aged worse than I expected, and the second part aged exactly as I feared.

European AI sovereignty

foundation models

frontier models

Two weeks after ‘The Two Pillars Are Both Rotting’, Kimi K3 closed the composite gap with the closed frontier to three points and Europe still built nothing. The moat rotted…

Latent Stochastic Differential Equations for Time Series

When the noise is the point

machine learning

stochastic differential equations

uncertainty quantification

Latent SDEs are the model I reach for when a neural time-series model has to take uncertainty seriously in all three directions: where it started, what it is, and how the…

Rediscovering the transformer, from attention up

Or, how much of the architecture can you rebuild before you reach for the paper?

neural-networks

A from-scratch derivation of the transformer, rebuilt piece by piece from attention up through the full block, with a hand-computed worked example on “an apple and a banana”…

The Two Pillars Are Both Rotting

If the US closed labs restrict access and the Chinese open-weights labs stop releasing, Europe has nothing. We need our own frontier labs.

European AI sovereignty

foundation models

frontier models

Europe’s AI runs on two foreign pillars, the US closed frontier and the Chinese open-weights frontier, and both can be pulled. Benchmarks show Mistral at 47 versus a…

$SmolWM architecture. Top: image, proprioception and instruction each go through their own encoder (ViT, MLP, and a frozen text encoder), fuse into the state latent z_t. The predictor takes z_t and the action a_t and outputs the predicted next latent z-hat. Bottom: the next image goes through the same shared encoder and fuse into the target latent z_{t+1}. On the right, the prediction loss compares z-hat to z_{t+1}, plus a SIGReg term keeps the latents Gaussian.$

A small world model that knows when a cube is cheating

Reimagining SmolVLA as a JEPA world model, and the story of where the planning worked and where it walked into a wall.

Building SmolWM: a tiny JEPA world model that learns robot dynamics with no labels, flags physically impossible video, and where latent planning hit a wall.

World Models and Why JEPA + SIGReg Is a Good Bet for Teaching AI Agents

autonomous agents

sample efficiency

Here’s a number that I’ve been thinking about for a while now. A frontier transformer based LLM today is trained on roughly 10 trillion tokens of text. Ten. A four-year-old…

Building a JEPA world model from scratch in Julia

Or, what I learned trying to teach a neural net physics it never saw

self-supervised-learning

neural-networks

A build of the smallest thing that is still a JEPA world model, in Julia on a bouncing ball, and what probing its latent space actually revealed about observability and…

Marketing Science: Why Most Brands Are Still Guessing

Bayesian thinking, measurement honesty, and the marketing stack that could use a physics lesson

marketing-science

Attribution is breaking, MMM is back, and the thing that actually separates honest marketing measurement from guesswork is Bayesian uncertainty. A physicist’s take.

The Gradient, Jacobian, Hessian, and Laplacian

A physics story

machine learning

A first-principles walkthrough of gradients, Jacobians, Hessians, and Laplacians. What they are, where they come from, and why they matter in machine learning.

A Law That Should Disturb Every One of Us

On Israel’s death penalty law targeting Palestinians, and why silence is not an option

international-law

On Israel’s ethnically targeted death penalty law, why the EU must act under the Association Agreement, and what you can do about it.

Of monsters and large language models

Why bigger is not always better

language models

neural networks

european sovereignty

Comparing small local LLMs against the biggest models from OpenAI and Deepseek on real tasks. The results might surprise you, the energy numbers definitely will.

How I got rid of American Technology and why you should too

A practical guide to replacing American tech with European alternatives. Every tool I switched, from email to cloud to hardware, and why it was worth it.

The cacophony around us

A rant about the lack of critical thinking

personal development

A rant about the explosion of uninformed opinions in the age of information. Why we need depth over breadth and writing over scrolling.

Learning a new skill: A thing of the past?

personal development

AI won’t replace human skill. The people who thrive will be those who master the skills to build it, not just use it.

Evidential musings

Utilizing conjugacy for the Multinomial distribution

dirichlet-multinomial

Using the Dirichlet-Multinomial conjugate distribution as a neural network loss function for classification with built-in uncertainty quantification.

ξ correlation: A new tool for your causality checks

Introducing the ξ correlation coefficient: a new rank-based measure that detects non-linear relationships Pearson correlation misses, with applications to causal discovery.

LogP, LogD, pKa and LogS: A Physicists guide to basic chemical properties

molecular properties

A physicist’s guide to LogP, LogD, pKa, and LogS. The fundamental molecular properties you need to understand for AI-based drug discovery.

What does the prior actually do in a Bayesian analysis?

A visual walkthrough of the Beta-Binomial model showing exactly what priors do to your posterior, and why maximum likelihood is a special case.

COVID-19 in Denmark - An epidemic in a small country

differentialequations

Modeling the COVID-19 epidemic in Denmark using differential equations and exploring what the numbers actually told us early on.

A quick introduction to derivatives for machine learning people

A first-principles walkthrough of derivatives for machine learning practitioners. No memorized rules, just the actual math.

The importance of context

maximum likelihood

Why context is everything in statistical modeling. How ignoring variable interactions leads to dangerously wrong conclusions.

Deep Neural Networks in Julia - Love at first sight?

First impressions of building deep neural networks in Julia. A hands-on exploration with the Boston Housing dataset.

On the apparent success of the maximum likelihood principle

maximum likelihood

Why the Bayesian treatment of models is fundamentally important, demonstrated through a time series model of Milk sales.

Building and testing a simple deep learning object detection application

object detection

Building a simple object detection app using a pre-trained MobileNet model on the COCO dataset.

About identifiability and granularity

marketing mix modeling

variable selection

How identifiability and data granularity affect Bayesian regression models, with practical examples from marketing mix modeling.

A few thoughts on apparent bimodality for regression problems!

Exploring how bimodal posterior distributions arise in Bayesian regression and what they tell us about model specification.

On the equivalence of Bayesian priors and Ridge regression

A mathematical proof that Ridge regression and Bayesian regression with Gaussian priors are exactly the same thing.

About choosing your optimization algorithm carefully

constrained optimization

A comparison of optimization algorithms on non-linear landscapes showing why your choice of optimizer matters more than you think.

The truth about priors and overfitting

hypothesis testing

How Bayesian priors prevent overfitting and what happens when you change them. A practical demonstration with regression models.

The insignificance of significance

hypothesis testing

The p-value distribution under the null hypothesis is uniform. Here’s why that makes significance testing far trickier than most people realize.

Doing price optimization in R

Calculating optimal product pricing in R using price elasticity models and simple optimization.

About confidence intervals

confidence intervals

credible interval

Why confidence intervals are not what you think they are, and how Bayesian credible intervals give you the answer you actually wanted.