Ted Underwood

This is the perfect amount of gossip. I need to know this, but I don't need to see a screenshot or want to know the specific dorky shit.

you can just do things (derogatory)

If the hierarchical reasoning model holds up under scrutiny, it can become the basis of someone’s "the sweet lesson" blog which we will be citing every time we can’t afford the compute resources to implement an idea.

They don’t frame it as a benchmark but as projection onto the “evil vector” :)

Persona Vectors: Monitoring and Controlling Character Traits in Language Models Large language models interact with users through a simulated 'Assistant' persona. While the Assistant is typically trained to be helpful, harmless, and honest, it sometimes deviates from these ideals...

According to this! arxiv.org/abs/2507.21509

Universities are not unified organizations: they're more like an armed truce between a bunch of competing departments.

In that context, changing core curricula is incredibly hard and unrewarding work.

When I say "machine learning," that's not limited to language models or neural networks.

The implicit premise of my recommendation is that students are going to be interacting with systems produced by machine learning for the rest of their lives, and many will need to tune or adapt those systems.

what "statistical learning" is, what "out-of-distribution" means, what a neural network is, how performance is measured and reasons not to trust those measurements.

"What the models can do" is a moving target, but machine learning is a flexible mode of reasoning about problem representations.

My logic here is the same as for, like, biochemistry. We don't teach all students "here's how to use the Agilent 2024 Microarray Scanner"; but we do teach them what proteins are and what a gene is. That knowledge will serve them broadly.

Same with machine learning. They need bias vs variance, +

OpenAI is going to release an open model (1st since GPT 2) and GPT-5 within weeks of eachother. The open model is indicative of where impact is possible in the AI space, where GPT-5 isnt a huge step like GPT-4.

Open models & agents/products are big stories of 2025 so far.

With care, and lots of caveats about change, you can teach students about the affordances or limitations of today's models.

But for a required course, I would recommend fundamentals of ML. I think that will serve students in a wider range of ways for a longer time.

100% accurate as a description in 2020, but already needed a lot of serious qualification (related to post-training) in 2022. By 2024 it was really no longer right to say that these models lacked sensory grounding, and by 2025 it was misleading to say they worked by directly predicting next words.

I don't mean that I think these courses will be one-sided. Bergstrom is making an admirable effort to be balanced in that material.

But I do think the account of "what LLMs do" is susceptible to dating quite rapidly, even if we resist adding a phrase like "can only." E.g., "fancy autocomplete" was+

The "LLM debunking" or even "media literacy for AI"-style material is going to date *so* fast. Much of it may be inapplicable by the time students graduate.

I think "AI literacy" is the path that will often be adopted — because it's easier to find people to teach it, and because faculty aren't ready to affirm that machine learning is going to be central to 21c life.

But I really think we would serve students better by teaching fundamentals. +

Yes, I mean for all students.

Yep, implementing stuff by hand is definitely the way to understand it

I think the answer is probably just “inertia”

The stated rationale would be something like “they can take that to fulfill their quantitative requirement,” plus probably a bit of throat-clearing about not wanting to be narrowly vocational

But actually just inertia

Text Shot: Something that has become undeniable this month is that the best available open weight models now come from the Chinese AI labs. I continue to have a lot of love for Mistral, Gemma and Llama but my feeling is that Qwen, Moonshot and Z.ai have positively smoked them over the course of July.

The best available open weight LLMs now come from China https://simonwillison.net/2025/Jul/30/chinese-models/ #AI #OpenSource

"It's called transfer learning if it works and overfitting if it doesn't."

A problem that used to occur in underparameterized models

I hope you’re right, but I have learned that I usually underestimate both human inertia and the power of maladaptive institutions to persist nevertheless

most people are *bad* writers. if a tool is available that lets them be mediocre writers for one of the few times in their life when the quality of writing will matter for them, that's not a bad thing - assuming (big assumption) that they think to actually fact check it.

It's going to be much faster. Everything is going to be much faster (and that's a major issue).

Correct. We should be teaching all that in high school.

I disagree w Ted mostly but strongly agree w him here and I think a good question to ask is why core requirements have not been updated to include basics of computing

And also, if there’s a foundation of knowledge all students have, you can teach an ambitious Archeology class that builds on it, without having to convince your own dept to make it part of the major

The reasons are not purely instrumental. Students should know what overfitting is for the same reason they should know what a sentence or an equation is; it’s just a matter of understanding the logic that governs the world around you.

It’s going to take decades for this to happen, but an intro to machine learning should be required for first-year students.

As a general purpose technology, AI has all kinds of unexpected uses that are hard to anticipate.

Example: this study finds blind users turn to AI to describe sensitive materials (pregnancy tests, checking appearance), they know it is not 100% accurate, but it provides privacy where there was none