iiinova — where your digital self keeps evolving

Alex Albaugh

@alexalbaugh.com

105 followers 228 following 37 posts

Chemical engineer, sort of. Assistant professor @ Wayne State albaugh.eng.wayne.edu

Alex Albaugh @alexalbaugh.com · 19/12/2025

I've had a healthy fear of studying glass. It's kinetically trapped and hard to sample. But my intrepid colleague @glassic-cam.bsky.social convinced me to be brave and we're now simulating organic glass vapor deposition. I think it looks like snow, so enjoy this glassy holiday scene!

020

Alex Albaugh @alexalbaugh.com · 02/11/2025

We're sharing our latest work at AIChE 2025 and with pre-prints! Hear Ali on mechanically interlocked polymers on Mon. or read it (arxiv.org/abs/2506.13968), read our foray into catalysis (chemrxiv.org/engage/chemr...), or hear me on WSU's ML for ChE class on Tues. and molecular motors on Wed.

AIChE 2025 technical session schedule for the Albaugh Lab at Wayne State.

021

Alex Albaugh @alexalbaugh.com · 10/09/2025

Happy Battle of Lake Erie Day to those who celebrate.

010

Alex Albaugh @alexalbaugh.com · 18/06/2025

The car has a drag coefficient of 0.3, a cross-sectional area of 2.5 m^2, and weighs 4500 lbs. With this and some reasonable estimates for the density of air and rolling friction coefficient, at a steady 30 mph air resistance is 80 N (18 lbs-f) and rolling resistance is 200 N (45 lbs-f).

Drag force is 0.5 * (drag coefficient) * (air density) * (cross-sectional area) * (speed squared).

Rolling resistance is (coefficient of rolling resistance, about 0.01 for a car) * (mass) * (gravitational acceleration, 9.8 m/s^2).

120

Alex Albaugh @alexalbaugh.com · 30/04/2025

Me at the beginning of the semester vs. me at the end of this semester.

000

Alex Albaugh @alexalbaugh.com · 21/04/2025

Is there anything in the middle?

Venn diagram of Stefa, Maxwell, and Boltzmann.

000

Alex Albaugh @alexalbaugh.com · 18/04/2025

This evidence is from NYT's lawsuit against OpenAI for copyright infringement. Is ChatGPT just regurgitating copyrighted material with some randomness sprinkled on top? Kinda. Like any machine learning model, ChatGPT learns by example and its outputs are going to look like what it has seen before.

Lawsuit filed by the NYT against OpenAI, creators of ChatGPT.

100

Alex Albaugh @alexalbaugh.com · 18/04/2025

What happens if you turn the temperature down on ChatGPT? The New York Times found out that it can reproduce their articles word for word. The most likely predictions for certain prompts are just copies of copyrighted material.

Example from NYT lawsuit against OpenAI. Comparison of ChatGPT output (left) and NYT article (right) with identical text marked in red.

Another example from NYT lawsuit against OpenAI. Comparison of ChatGPT output (left) and NYT article (right) with identical text marked in red.

100

Alex Albaugh @alexalbaugh.com · 18/04/2025

They're the same*! By adjusting an LLM's temperature you adjust how it selects words. A high temperature LLM will select more broadly from the distribution, making stranger choices more likely. A low temperature LLM will usually just select the most likely word.

High temperature LLM makes a wider variety of words more likely.

A low temperature LLM will usually just select the most likely word.

100

Alex Albaugh @alexalbaugh.com · 18/04/2025

Converting from activations in the LLM neural network to probabilities for the next word is done with a softmax function. The softmax has a temperature that adjusts the resulting probabilities. Compare the softmax function on the left with the Boltzmann distribution on the right.

Softmax function gives the probability of some word (i) based on the LLM network activations (x_i) and a "temperature" (T).

Boltzmann distribution again. Probability density of a state (x) based on its potential energy (U) and the Boltzmann constant (k_B).

100

Alex Albaugh @alexalbaugh.com · 18/04/2025

LLMs generate a probability distribution for what the next word should be. Then they make a randomly weighted selection from the options. This is why you won't get the same answer twice from ChatGPT if you ask it the same question. There's built-in randomness to make it feel like natural language.

Example of LLM giving probabilities for the next word. "Well" is selected with 84% probability, good with "12%", and so on.

100

Alex Albaugh @alexalbaugh.com · 18/04/2025

A large language model (greatly simplified) tries to predict the next word. And it does this over and over. So if you give it a prompt, it will try to predict the next word. And then the next one. And then the next one.

LLM making a word prediction for the prompt, "How are you?"

LLM making a prediction for the next word.

LLM making a prediction for the next word, again.

100

Alex Albaugh @alexalbaugh.com · 18/04/2025

The mathematical form of this probability is called the Boltzmann distribution.

Boltzmann distribution. Probability density of a state (x) as a function of the potential energy (U), temperature (T), and Boltzmann constant (k_B).

100

Alex Albaugh @alexalbaugh.com · 18/04/2025

Say you've got a system that can be in states with different energies. Which state is the system most likely to be in? The state with the lowest energy. But the rest depends on temperature! Lower temperatures make other states less likely and higher temperature make other states more likely.

Potential energy (U) as a function of the state (x).

Probability density (rho) of a state (x) for low, moderate, and high temperature (T).

100