Skills in CC have been a bit frustrating for me. They don't trigger reliably and the emphasis on "it's just markdown" makes it harder to have them reliably call certain tools with the correct arguments.
The idea that agent harnesses should primarily have their functionality dictated by plaintext commands feels like a copout around programming in some actually useful, semi-opinionated functionality (not to mention that it makes capability-discoverability basically impossible). For example, Claude Code has three modes: plan, ask about edits, and auto-accept edits. I always start with a plan and then I end up with multiple tasks. I'd like to auto-accept edits for a step at a time and the only way to do that reliably is to ask CC to do that, but it's not reliable—sometimes it just continues to go into the next step. If this were programmed explicitly into CC rather than relying on agent obedience, we could ditch the nondeterminism and just have a hook on task completion that toggles auto-complete back to "off."
Giving them more credit than they probably deserve but: when you're solving "by hand" like they are in the show, keeping a known non-match couple together may actually be helpful for interpreting the results of a matchup ceremony because you'll know that that couple didn't contribute to the beams.
Thanks! Optimization was something I'd played with in previous rounds of coding up AYTO simulations, but not in the most recent version. (See the bottom section of this notebook [0]). There's also a very thorough treatment of the problem in a blog post from 2018 by SAS (the software company) [1]. It's surprising how many people have been drawn in by the allure of AYTO!
As a math guy who loves reality tv, I was also drawn to the show and wrote a blog post [0] about how to programmatically calculate the probabilities as the show progresses. It was a lot of fun optimizing it to be performant. You can `pip install ayto` to use it to follow along with the show or try out scenarios.
The linked post is a very thorough treatment of AYTO and a great read. I really like the "guess who" bit on how to maximize the value of guesses. It's a shame the participants aren't allowed to have pen and paper—it makes optimization a lot trickier! I'm impressed they do as well as they do.
I've been building a little toy computer and assembly language that's interpreted in python. Pretty close to the first release (and introductory blog post) and a lot of fun to build (and learn a bit more about real assembly as I go).
I gave a very short talk (now a blog post) about embeddings and we use them to bridge the gap between human notions of understanding and digital representations. It might be of interest to people who enjoyed this post: https://danturkel.com/2025/03/10/ignite-machine-understandin...
> I doubt that any recommendation system is capable of providing meaningful results in absence of the "awareness" about the actual content (be it music, books, movies or anything else) of what it's meant to recommend.
Years of experience have proven that you can get quite far with pure collaborative filtering—no user features, no content features. It's a very hard baseline to beat. A similar principle applies to language modeling: from word2vec to transformers, language models never rely on any additional information about what a token "means," only how the tokens relate to each other.
I recently had a letter published in The New Yorker in response to Andrew Marantz's (excellent) story "Among the A.I. Doomsayers." In particular, I wanted to highlight that extremist doomers and accelerationists are not the only ones concerned about AI's future. In the post, I elaborate a little on the letter and provide links to further reading.
Hey Francois, congrats to you and the team on the launch! I've generally chosen Pytorch over Tensorflow for my day to day, but now that Keras is framework agnostic I'm excited to revisit it.
One thing I'm wondering about is if it's possible (or necessary?) to use Keras in concert with Pytorch Lightning. In some ways, Lightning evolved to be "Keras for Pytorch," so what is the path forward in a world where both exist as options for Pytorch users—do they interoperate or are they competitors/alternatives to each other?
This looks really cool. One thing I've wondered about with, e.g., the OpenAI API is if json is really a good format for passing embeddings back and forth. I'd think that passing floats as text over the wire wastes a ton of space that could add up, and might even sacrifice some precision in. Would it be better to encode at least the vectors as binary blobs, or else use something like protobuf to more efficiently handle sending tons of floats around?
It was a bummer when the O'Reilly benefit last year [0], but I'm glad they're bringing it back at a steep discount from the street price. Access to thousands of O'Reilly books, Manning books, MIT Press books, and more for $75/year (on top of your $100/year membership) is still a solid deal.
Funny enough, I did this same project (minus the fancy web interface) for a numerical linear algebra course in college—except I had to do it in Matlab.
It's worse than just about any "real" image compression algorithm, but it works! (Plus you get lossless compression if your image is low-rank.)