More context now that I’m not annoyed. I think I wanted this site to be old ML twitter too much and tried a bit much to make that happen. I’m just going back to old posting rules: once a day, no discourse, mostly papers

This week's #PaperILike is "Elephants Don't Pack Groceries: Robot Task Planning for Low Entropy Belief States" (Adu-Bredu, RAL 2022). Love the focus on planning with "low entropy beliefs" -- not full-fledged POMDPs, but also not full observability. PDF: arxiv.org/abs/2011.09105

What's left to do in self-driving given Waymo is taking off? An argument that it's still a great research problem: open.substack.com/pub/emergere...

Deleted a post that was going viral and bringing an undesirable amount of attention from anonymous accounts. But, I'm going to be honest, the toxic folks here have won. So much of the usage of the site is about avoiding their attention. I'll just use this site to get book recs going forward

Unfortunately suspect I am experiencing a type of platform capture here. Going to try to spend less time on this site and more time on long-form writing around the things I find exciting as opposed to the conversations I dislike

Incredibly precise manipulation from Kyber Labs

Is there a good understanding of why subway fare evasion ticked up so much over time?

Hello! 👋 Are you interested in AI for board games using language models? Want to do some hobby tinkering with fine-tuning or RL? We've released an easy-to-follow example colab that fine-tunes Gemma models via Kauldron to mimic an MCTS player. Details here: github.com/google-deepm... ♟️🎲♦️♠️♥️♣️✨🎉

Cannot say enough how lovely our little lab group is. Particularly taken by “self play emerges from self play“

Had a group reach out offering to pay me to place PhD students in my group and uh, there's a lot of problems here. First off, I hear some other profs were offered 120k so that's a nonstarter

Doing a PhD is - at heart - one long discussion with your mentor. The discussion changes over time - with unexpected turns and ups & downs - but through it all is a pair of people discussing a topic endlessly to make sense of it. PhD students: choose someone you like to talk to!

Tech executives saying something stupid is really not an excuse for turning your brain off and not being able to evaluate a technology on its own terms

Broke: LLMs are no worse than a calculator Woke: Being opposed to calculators

Implementing California-style governance failure modes from the classic legislative example "do not implement California-style governance failure modes"

Kinda thought that antisemitism would just keep going down for the rest of time and pretty sad to be wrong about that

It just seems like blue states should be building more housing and renewables than Texas. Any political platform that doesn’t reckon with why this is seems suspect

Infinite Scroll RecSys Disorder (ISRSD) is real

At this point the only browsers are functionally Chrome and Safari. It kind of makes sense that other browsers would try new things to survive and pursue their vision of the internet

I continue to think LLMs are absurdly powerful for learning, and I'm using them more often for it. The problem is that a lot of people don't have solid metacognitive skills around epistemics and autodidactism. You've got to have a handle on what you know, how you learn, and if you're learning.

Is "action chunking" a rebranding of options? Is it something distinct?

shockingly a government run by 4chan is even less of a good idea than it sounds

at times these feel like the actual axes we're debating over

A text based social media site, but only for people who don’t like to read

It’s dangerous to name something in an empirical research project. Once it’s named, it acquires a reality that can persist even if it’s not grounded in anything formal

It would be nice if classic villains were why progress is slow but the story rarely works out so nicely. Billionaires aren’t why homes are expensive, or why we can’t build a subway, or why no one wants solar panels built near them

Sorry, journalistic standard cannot be that personal drama of any tech executive within 3-4 levels of leadership is worth reporting on

Really would love to see a product roadmap that makes clear there's an exciting future for this platform...

New work in why action chunking is so important for robot control (it helps fight compounding error) arxiv.org/abs/2507.09061

Love that the response to the OpenReview issue was for the major ML conferences to donate to it to help them improve their stack. That's what you want to see.

This might make people mad but it's not obvious to me that spending a ton of time writing and reviewing grants is either a good use of time or an obvious way to increase meritocracy

Does the median person like a product that 800 million people use of their own volition? It's impossible to say

Tread softly, if you say it you agree with it

So the federal government deleted 18F, which was doing great work modernizing the government, and then replaced it with a clone? What is the point of all this. techforce.gov

I love Tesla engineer posts about how sensor fusion isn't possible

Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3—and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵

A real challenge of social media is the way it fuses the broader debate with the individuals participating in that debate. You don’t want to cede a point because it causes “your side” to give up standing in the eyes of the audience even if it makes the 1 on 1 conversation much duller

An excellent article describing how LLMs fit into the bureaucratization of science: open.substack.com/pub/artifici...

Have had more close calls with a high speed e-bike on the sidewalk than I would like this year. Would probably be good if something were done about this...

Unfortunately it appears much of the academic community has reconstituted itself on LinkedIn

A quick sunday rewrite of an old blog post about how one should evaluate the effectiveness of an empirical paper: open.substack.com/pub/emergere...

After THE GLASS ONION I was very worried but I am glad to report that WAKE UP DEAD MAN is in fact very funny and kind. May we get 10 more of them.

To every person who complains about speeding laws just remind them: 1. Roughly 1/100 people die in a car accident 2. Risk of death goes up rapidly with higher speeds

My wife and I have used Messenger for a decade and regularly look back through old jokes and photos. Today Messenger archived our chat and we seemingly can’t undo it. What is this nonsense.

More people probably need to know that cars are the fourth largest cause of death in the US. It’s such a “normal” part of life that people forget

Sorry I have so few childhood heroes left you can't take Oliver Sacks

Conservatives who support this stuff are trying to tear immigrants out of the communities they have enmeshed themselves in, to make it impossible for them to live their lives without every moment being fearful. www.nytimes.com/2025/12/12/u...

Have seen a lot of family members pick up all sort of gross ideas by slow osmosis from their media environment. This has shaped why I care so much about not letting any single individual have access to a propaganda platform

🚨 New preprint! 🚨 Everyone loves causal interp. It’s coherently defined! It makes testable predictions about mechanistic interventions! But what if we had a different objective: predicting model behavior not under mechanistic interventions, but on unseen input data?

A practical introduction to (deep) RL, providing intuitions to understand the more recent algorithms (continued). In this second post, I continue from DQN on to the Soft Actor-Critic (SAC) algorithm and its extensions. araffin.github.io/post/rl103/

If you've been interested in FROs, you'll want to see this from NSF: sam.gov/opp/7332ade9... "sustained... support to... independent organizational structures operating outside of existing academic, start-up, and industry constraints... a dedicated, full-time team... operational independence..."