daturkel

Surely this logic doesn't apply if we're to believe that "code is cheap" now :p

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."

Notion AI is pretty great for both search and writing.

Relevant xkcd: https://xkcd.com/55/

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!

[0] https://github.com/daturkel/pyto/blob/master/AYTO_S8.ipynb [1]: https://blogs.sas.com/content/operations/2018/08/14/are-you-...

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.

[0] https://danturkel.com/2023/01/25/math-code-are-you-the-one.h...

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).

https://github.com/daturkel/dt31

That insane behavior around acting on the highlighted image instead of the selected image has finally been fixed: https://github.com/darktable-org/darktable/issues/16850#issu...

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?

Annoy came out of Spotify, and they just announced their successor library Voyager [1] last week [2].

[1] https://github.com/spotify/voyager [2]: https://engineering.atspotify.com/2023/10/introducing-voyage...

Any particular reason to highlight this release? 1.4 introduced Properties, which is cool, but this is just a minor release.

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.

[0] https://news.ycombinator.com/item?id=30932790

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.)

I maintain a list of well-known or foundational papers in ML in a github repo that may be of interest to readers of this thread

https://github.com/daturkel/learning-papers

I appreciate the post. The linked version is to the draft I shared while still writing. Here's the permalink to the final version:

https://danturkel.com/2022/12/27/advent-of-code-2022.html

(Maybe @dang can fix the submission link)

I created a Github repo that tracks "classic papers" of machine learning that folks on HN have enjoyed in the past. Feel free to make a PR if you think I'm missing something good.

https://github.com/daturkel/learning-papers

I've made a few measurable goals for 2022 that I wrote up on my personal site [0]: - read 25 books (3 of which are "technical") - write 12 blog posts - take one online course - get conversational in French - shoot 2 new batches of photos and post them online (and get through some of my photo backlog)

And the slightly less measurable: eat better and exercise more. I'm getting old enough that I should stop putting this off.

[0] https://danturkel.com/2021/12/31/goals-for-2022.html

Python is not fully a functional programming language but it supports some functional patterns. There's a nice mini-ebook by David Mertz on functional programming in Python, and it used to be freely available but I can't find it at the moment. However, he wrote an article version here: https://developer.ibm.com/languages/python/articles/l-prog/

Also, pattern matching is coming to python in 3.10. You can read about it here: https://www.python.org/dev/peps/pep-0634/

Overleaf is very unlikely to be a good drop-in replacement for MS Office for your company because Word and LaTeX are very different tools (with different goals and different audiences in mind). The former is a user-friendly word processor and the latter is a fairly complex plaintext-based document preparation/typesetting tool.

I doubt that most offices would benefit from teaching their employees that instead of just hitting control-B to bold some text, they should now do \textbf{foo}, or that quotes should now be types ``like this''. If you need the huge array of features (which is typically only the case in the context of academic publishing), then it's worth making the investment to learn. But for 99% of corporate documents, Word likely can do what you need with much less pain.

This is interesting. I don't know much about html canvas, but it looks like this is handled in the author's custom js [0] in the line that starts with: "} else if (mode === "blades") {"

[0] https://ciechanow.ski/js/lenses.js

I came to the comments to shill for Numi, which is my favorite app for quick calculations. (I don't personally understand why tiled number buttons would ever be desirable for a non-touchscreen GUI app—I much prefer typing equations in.) And I love Numi's variables which you can update and automatically propagate the changes through downstream calculations.

I'm going to chime in to disagree strongly with this. I've seen and worked with lots of non-CS PhDs in data science who have been terrific. It's pretty common too—there's even the Insight program that trains PhDs to go into data science [1]. I don't think it will be seen as a downside and will instead show that you know how to do research and solve problems.

[1] https://insightfellows.com/data-science

For a column supposedly about a greater theory of how we (should) use the Internet, it sure devotes a lot of space to attempting to vindicate the Hunter Biden laptop story...

I would argue that originalism doesn't really exist in practice in the context of the Supreme Court. Interpreting the constitution is always an active process where the interpreter injects some of themself into the the interpretation. The difference is that those who interpret the Constitution as a "living document" are up-front about this process, while originalists deny that it is happening.

And then I'd argue that this half-baked notion of Internet Originalism isn't really a sensible metaphor. The author argues that the way we use and build the internet should be in line within the framework that the web is a democratizing tool, and I agree with that (simply because I share those values, though I'd debate how much they're truly enshrined in the concept of the internet itself).

But the author seems to care more about not Internet Originalism but...Twitter Originalism. That a private company has to abide by a set of values that it (purportedly) supported at its inception.

That is why a “living internet” interpretation is so dangerous. These companies are driven by profits and politics, not principle. [...] The alternative is “internet originalism” — no censorship. If social media companies returned to their original roles, there would be no slippery slope of political bias or opportunism [...]

Yes—companies are driven my profits (and, to the extent that they're at risk of regulation, politics)! That's, like, their whole thing. Regardless of the potential greater good of these companies returning to their supposed "original roles," they're supposed to be forced to do so even when it means losing money and failing to self-regulate? It seems perfectly reasonable that a company which seeks to provide a mass communication platform might realize that the rules which worked at first need to be adjusted at scale. I don't see that as some massive betrayal.

No doubt there will be (and has been) some over-correction as Twitter navigates how to quell the use of its platform for spreading misinformation and hate speech. They've historically not been great at it, and they'll probably get to an ~ok~ solution eventually. And if they don't? A new private venture can show up to offer something better (as has happened several times).

Hello. For a while I've been adding links to this repo in my spare time. Machine learning is a fast moving field, and new techniques are constantly emerging in research. While the formulation of an ML concept as it's used today may have diverged or evolved from its original presentation, I've found it helpful to track down the papers which introduced common ML concepts, from things as general as Lasso regression to as specific as the papers which accompanied GPT1, 2, and 3.

Feel free to contribute. I'm open to suggestions on what to include as well as how to organize and present it. I hope it's a useful tool for people hoping to read up on the research behind various ML ideas.

This is actually a very interesting point that I've come across on a few different platforms recently (Patreon, for example, [also targets their homepage to creators, not consumers](https://www.patreon.com/)). I wonder if Substack just doesn't want to deal with opening the door to some newsletters getting "free promotion" by being features on their homepage? Nonetheless, I'd love to see "trending" or "editors picks" or something.

Nonetheless, here are some I subscribe to:

Big, by Matt Stoller (about monopolies): https://mattstoller.substack.com/

Margins (about business and tech): https://themargins.substack.com/

Flow State (music): https://flowstate.substack.com/

Normcore tech (tech, data science): https://vicki.substack.com/

The Passion Economy (business): https://passion.substack.com/