Theresa Eimer

Foundation models on the AutoML podcast 2/3: are LLMs killing AutoML? It's probably not that simple. Listen for more details 😉

Stealing all of the recommendations!

This made me think of The Left Hand Of Darkness, though I guess that's actually almost the opposite, communication bridging a seemingly impossible gap in understanding each other...

I fell into a hole, but made it out again with new episodes! This is part one of three of an accidental series on foundation models. The next parts will be released in October and November, so stay tuned!

Great opportunity to work with great people. Go apply!

New blog post: AI Allergy.

On my increasing disgust with the AI discourse, even though I still like the technical and philosophical. And how I wish I could be excited about AI again.

togelius.blogspot.com/2025/08/ai-a...

It is time

The "reproducibility crisis" in science constantly makes headlines. Repro efforts are often limited. What if you could assess reproducibility of an entire field?

That's what @brunolemaitre.bsky.social et al. have done. Fly immunity is highly replicable & offers lessons for #metascience

A 🧵 1/n

Getting SAC to Work on a Massive Parallel Simulator: Tuning for Speed (Part II) | Antonin Raffin | Homepage This second post details how I tuned the Soft-Actor Critic (SAC) algorithm to learn as fast as PPO in the context of a massively parallel simulator (thousands of robots simulated in parallel).

Need for Speed or: How I Learned to Stop Worrying About Sample Efficiency

Part II of my blog series "Getting SAC to Work on a Massive Parallel Simulator" is out!
I've included everything I tried that didn't work (and why Jax PPO was different from PyTorch PPO)

araffin.github.io/post/tune-sa...

1/2 Offline RL has always bothered me. It promises that by exploiting offline data, an agent can learn to behave near-optimally once deployed. In real life, it breaks this promise, requiring large amount of online samples for tuning and has no guarantees of behaving safely to achieve desired goals.

Crazy volume! On the other hand, not that surprising. We also got one of these and only did so because it was such a good deal that even if our complete lack of experience makes research on it hard, we can use it for teaching only, and be okay with spending the money. I doubt we're the only ones!

AutoML School 2025 Scope AutoML has become a cornerstone in the toolkit of many developers and researchers. With the rise of foundation models, AutoML's potential has expanded even further, enabling smarter, more powerf...

📢 Only 3 Weeks to Go!

The AutoML summer school (June 10-13th) is just around the corner, and there is not much time left to register!

---> www.automlschool.org <---

👇 We added several new speakers to the program

I got fooled by AI-for-science hype—here's what it taught me I used AI in my plasma physics research and it didn’t go the way I expected.

Going to the hospital because I broke my wrist smashing the endorse button:
www.understandingai.org/p/i-got-fool...

We can only presume to build machines like us once
we see ourselves as machines first.
Abeba Birhane (2022, p. 13)
This is the core. So true.

YouTube video by Center for Connected and Automated Transportation The Future of AVs Panel | 2023 CCAT Symposium | Day 1

Panel discussion on the current economic precarity of autonomous vehicle businesses. www.youtube.com/watch?v=gDG-...

"We are at a really tough spot in generating flows of cash right now." 👇

After a short era in which people questioned the value of academia in ML, its value is more obvious than ever. Big labs stopped publishing the minute commercial incentives showed up and are relentlessly focused on a singular vision of scaling. Academia is a meaningful complement, bringing...
1/2

It's strange to me that the focus of many people's worry is still "superintelligence" and not the reality we're currently living where increasingly authoritarian governments wield technology oppressively.

This fantastical distraction based on speculative rhetoric is increasingly harmful.

Humanoid Robots in Manufacturing Or, there's a reason we don't pull cars with mechanical horses

A sensible perspective on humanoids in manufacturing (TLDR: if you can make humanoids, you can probably make better, more manufacturing specific things)
blog.spec.tech/p/humanoid-r...

Mark your calendars, EWRL is coming to Tübingen! 📅
When? September 17-19, 2025.
More news to come soon, stay tuned!

Llama 4: Did Meta just push the panic button? One of the weirdest releases of the year and understanding the future of the Llama endeavor. For the time being, we have some more amazing open weight models!

Llama 4 was a messy release: unreleased finetunes boosting scores, rumors of training on test, released on a weekend, etc

As (open) models are commoditized / competition grows, what is the role of Meta's Llama efforts in the future? Should they continue?

At least there is no need to jailbreak the model anymore 🫠 (Is there a counterpart to make it nicer 🎭?)

The school kids visiting me during this year's future day really had hard-hitting questions: "Do you still have a lot of free time?"

Me, a pretty fresh and currently slightly overwhelmed PostDoc: "It's important to be good at time management. Like my colleague, maybe you should ask her."

So far, 2,135 people have responded to the poll Søren and I posted a few days ago. Of those, 94.4% replied “Yes” to being interested in officially presenting accepted @neuripsconf.bsky.social papers in Europe. (1/7)

German media I beg you one day just please go just one day without being obsessed with migration. One day. I promise it won’t kill you. You have lakes and mountains and good football and good healthcare and asparagus. You’ll be fine.

True, I've been "socialized" in the AutoML community, how to compare algorithms is a big deal there.

I remember discussing with my advisor whether it's worth evaluating issues with improper HPO setup in RL, he thought it was so obvious that everyone must already be doing it (spoiler: not really)

Well, then there's only one alternative: "We define OurPO as PPO with lr=0.01, ent_coef=0.1.... and compare it to OurQN which is DQN with lr=...." 😂

Tell them their argument might be valid with different hyperparameters

This obviously then also depends on budget, HPO method and combines performance and tunability into one score, but I think that's quite reasonable in practice. Not very satisfying for an empirical nihilist, though, I imagine 😉

Well, what validity are you looking for? The absolute "algorithm A is better than B on benchmark C" is hard wrt hyperparameters, but algorithm A is better than B on C given I can realistically try out 50 configurations" is what we often want in empirical ML anyway, no?

Machine Learning that Matters Much of current machine learning (ML) research has lost its connection to problems of import to the larger world of science and society. From this perspective, there exist glaring limitations in the d...

So true, Gilles.

Yes, it is a pretext task, but often, when we try real tasks, we find that the problems are not those we expected.
We need more people looking at relevant problems.

Kiri Wagstaff said this 15 years ago

arxiv.org/abs/1206.4656

My PhD supervisor discussed my first two or three reviews with me (including checking over wording etc.) and does that for all his PhDs, but I know that's not the standard in most other groups I'm familiar with...