Matt Stuchlik

Cheat code for faster memory reads 👀

@FelixCLC_ It's probably closer to 99%, at least on sufficiently constrained problems: josusanmartin.com/blog/2026/01/1…

i often forget that claude is a better performance engineer than a good 70+% of programmers. then i remember that' for a significant amount of programmers, enabling O3 is considered "performance engineering"

15% improvement on Parse Integers. Insane breakthrough.

@DrPhiltill Yeah, makes sense! One thing I like about Anki that the article does not talk about, is that it has an API and can therefore be used by Claude Code / Codex. You can imagine the exciting opportunities there... I suggest looking into DuChinese for listeining, I find it very nice.

This was great! Thanks for linking it. I tried using Anki but the startup was too steep for my patience and I found HanziHero which is the same idea but turnkey. My whole program is here: HSK1 Space repetition work (1-2 hours per day) HanziHero 20 new items/day HanziHero review ~100/day Listening & Speaking: Hello Chinese App YouTube Writing: Hanzi in “HSK1 Chinese Characters Workbook for Beginners” and “HSK2…” Grammar: Study 1 grammatical point every day on the ChineseGrammarWiki Website. Etc. for HSK-2 & up.

From the HelloChinese app during Mandarin practice today. 😂 It’s a great app, btw. Highly recommended.

@bryancsk "28 hour New York Bagel"

The New Yorkers in the comments are discussing the boiling of bagels in complicated terms like a Singaporean discussing the complexities of chicken rice, it's very refreshing.

@tereza_tizkova Ja taky!

Looking to cowork in Marina/Pacific Heights today again and looking for new spots. What’s good coworking cafe?

Wong Kar-Wai supremacy every shot in the mood for love is so distinct

@AI_Arav @dwarkesh_sp Did you do (ROT13) pbfvar fvzvynevgl sbe cnvevat + erfvqhny abez sbe beqrevat or something else?

solved the dwarkesh jane street puzzle would love to share notes with anyone else that's finished!

@GrantSlatton @sbensu en.wikipedia.org/wiki/Subvocal_…

@sbensu Someone should make one that operates via lip-reading webcam instead of audio so I can use in the office around coworkers

i've reverted to my 2004 dialect of AIM shortenings with my CLI code agents, since the bottleneck is mostly my own response speed why -> y, you -> u, really -> rly, etc and all the initialisms are coming back, wdyt btw np idk idc lmk ngl i kinda love it, brings me back

@DrPhiltill @ramez Reminds me of this snippet.

One example: Deep insight into the psychology of every individual human on the planet and an utterly deep grasp of group dynamics enabling implementation of subtle, undetectable, yet fully pervasive and utterly persuasive propaganda (basically, effective mind control) to overthrow existing democratic regimes and make the world one-party rule. It will be like a strategy game of humans versus guinea pigs, where the ones with the lesser AI are the guinea pigs.

What do people believe China will do with more advanced AI chips that will harm the US?

@RYANHINGSHING Don't listen to this man, 陕西 is where it's at :P

Happy weekend! I went to Shenzhen HuaYang Jiu Lou (华洋酒楼) and had some truly exquisite Dim Sum. Highly recommend it if you're looking for great Dim Sum and great Chinese food The best food in China is Guangdong’s traditional food if you ask me!

@davidma How useful did you find LLMs in solving the challenge (apart from implementation work, obviously useful there).

What worked for me: - LastLayer is the only one with shape 48,1 - There are 48 inp slots and 48 out slots. That's 48*48 ways to construct a Block (and 48! ways to construct 48 blocks), and 48! ways to order the blocks once they are constructed. - Computed cosine similarity between X and residual for the 48*48 ways to construct a Block. The idea is that if X is basically unrelated to inp/out, then the cosine similarity will -> zero as the dimensions -> inf. If X and inp/outp have a relationship, then there's a chance that the cosine similarity is not 0 and the chance doesn't -> 0 as dim -> inf. - I looked at the top 48 pairs with outlier cosine similarities and to my surprise they are all disjointed which gave me a high confidence I've found the right pairings. If they weren't totally disjointed, I would have used that to reduce the search space. - For the block ordering, which I "solved" before the Block pairing, compute how large the residual is for every block (residual norm) applied to X. The idea is that earlier blocks will contribute more to modifying X to match y_pred. But in practice it seems the reverse is true. From that permutation, brute force pairwise swaps until there's no more improvement. What partially worked: - Define an affinity metric between inp and out: Single-Block MSE (compute the MSE between y_pred and the y after one block and final layer). - Hungarian algo to pair up the inp and out. What didn't work: - Bunch of profiling: looking at feature distribution, features correlation matrix, weights std/norm, zeros, residual norm / X norm, visualizing the layers - Trying to incorporate the idea that the gradient should be small when y_pred ~= y_true - Greedy construction with various orderings. - Didn't try Sinkhorn as suggested by AI.

@will_z65038 @Alan_Ma_ @sergey_slotin @dendibakh And stuff from Agner Fog of couse! #manuals" target="_blank" rel="nofollow noopener">agner.org/optimize/#manu…

@will_z65038 @Alan_Ma_ @sergey_slotin Check out Performance Analysis and Tuning on Modern CPUs from @dendibakh too.

One of the greatest resources I've stumbled across from @sergey_slotin and team: en.algorithmica.org/hpc/ Phenomenal work!

@esrtweet @nocontextmemes Time stamps with no TZ offset are evil.

@nocontextmemes The one true date stamp standard is RFC 3339 YYYY-MM-DD HH:MM:SS In software contexts I prefer the version that's one token with a T in the middle. Fixed length tokens that naturally sort in time order, yay!

@SebAaltonen Huh? I use GeForce NOW quite often and it's been working great.

"Stadia service is no longer available. Thanks for playing." Jeff is not an expert. Google hired experts and burned a lot of money and couldn't achieve cloud gaming. Instead, gamers are buying 240Hz monitors and 1000Hz mice to reduce the input lag further.

@SurrealistShip youtube.com/watch?v=LqGqqz…

What films are the best to have on in the background of a party on mute? Samsara? Ghost in the Shell? Playtime?

@simonw @Voxyz_AI I wish @awesomekling did an in-depth review!

@Voxyz_AI It's ~3m lines so I've not even attempted to look

In case anyone's interested, that vibe-coded(ish) web browser project that Cursor released the other day does actually compile now, they fixed it up and added instructions to the README. I built it on my Mac and took these screenshots:

@josusanmartin Thanks for the call out :)

The game has changed: How I vibecoded my way into the #1 position on the highload.fun leaderboard josusanmartin.com/blog/2026/01/1…

I try hard to not be a pessimist but this is so demoralizing. WTF is happening