One incredible thing about AI coding is how the AI has to learn the codebase again at the start of every session. It starts from a blank slate. Then it reads documentation and greps some source code and gets to work. It does this bootstrapping every time! And it takes just a few minutes. Superhuman.

The human genome's most variable and clinically important regions (centromeres, telomeres, and acrocentric short arms) have been hardest to study at scale. Thrilled to share KaryoScope, our new preprint that brings them within reach. 🧵 www.biorxiv.org/content/10.6...

Today, I did some Rust coding. Claude could not implement a feature properly. It was complex. So I did it myself. I typed every character into the text editor myself! I had forgotten how laborious coding used to be. 2.5 hours later, it was done. 212 lines added and 71 lines deleted. Feels great!

What is this progress bar doing

www.war.gov/News/Release... "These agreements accelerate the transformation toward establishing the United States military as an AI-first fighting force". "AI-first fighting force".

I redesigned my personal website: jnalanko.net.

Feels important. www.theguardian.com/environment/...

www.erdosproblems.com/forum/thread... AI apparently just provided a new insight into number theory. Terence Tao comments: "In any case, I would indeed say that this is a situation in which the AI-generated paper inadvertently highlighted a tighter connection between two areas of mathematics".

Finding the right level of abstraction in software engineering is hard. It's tempting to generalize things, but as we generalize, we lose opportunities to optimize for specific cases. 1/4

The latest Trump post makes me think: Is there anything Europeans could do to influence US politics to a less crazy and less genocidal direction? Would economic sanctions work? Support for resistance movements? Anything?

I regret to inform you but, you are, in fact, the bad guys. So sad.

Did Mastodon already fail? It's very quiet. People tell me to follow hashtags. But tags like #bioinformatics have only 8 people posting in the last 2 days, and a lot of those are bots. Even #rustlang has just 10 people in the last 2 days. The posts that get on my timeline are not very interesting.

Accepted to CPM 2026! In this paper, we consider the problem of k-mer counting in *graphs*. We show that it's #P-hard even in deterministic DAGs. But on Wheeler graphs, it's tractable -- easy in O(nk) time. We can also do O(poly(n) log k) but it gets complicated. #CPM2026 arxiv.org/abs/2509.22885

Rust PSA: don't use serde + bincode for deserializing large things, like bitvec::Bitvectors. It seems to use a doubling buffer to load it instead of pre-allocating the space. So it can have up to 2x overhead. Loading a 130GB bit vector can take 256GB of RAM. #rust

There's some really nifty math in the analysis of this paper!

Dependency tree of a time-tracking app that I vibecoded in Rust.

I wanted to poke around the Gene Ontology database. Claude was immediately useful to get some basics statistics on the ontology graph. Then I asked it to make a visualization. In a few minutes, it gave me this really useful interactive browser. I keep getting blown away by these AI models.

New measure of software engineering ability: value-added-over-AI. This is what CS degrees will be about, and this is where the money is. Can you do a better code review than AI? If not, I could've just asked the AI myself -- no value added. AI is the new baseline.

Frustrations with Claude today. It generated some array indexing logic and added a comment that endpoints are inclusive. They weren't. The code worked though. But the comment wasn't helping! 1/2

ChatGPT5.4 (Extended Thinking) feels no better at math problem solving than 5.2 did. This is the first model that does not feel significantly better than the previous ones at this. Are we finally hitting a ceiling? My timeline to math research automation and AGI just got a bit longer.

64GB of DDR5 RAM costs... 1000 dollars now? Bullish for succinct data structures!

Donald Knuth is impressed by the creative problem solving capabilities of Claude.

The latest Claude seems to be able to synthesize small apps from a short description pretty much autonomously. I'm having hard time adjusting to this reality. I can now make small bespoke productivity apps just for myself. Computers just become a lot more useful for a lot of things.

Cool repository from Terence Tao: A list of various mathematical constants whose exact value is unknown. The page lists the best known lower and upper bounds for each constant. TCS is represented: e.g. undecidable busy beavers and random 3-SAT satisfiability threshold. github.com/teorth/optim...

Here is a small research project that I did recently with chatbots: github.com/suomela/fin-... — everything is vibe-formalized in Lean 4 (approx. 24500 lines of Lean code).

A quick review of Venice #dsb2026 blog.jnalanko.net/2026/02/20/v...

The architecture and art at the DSB 2026 workshop is out of this world. The talks are also great! #dsb2026

The feeling when you try to update some dependencies but get just a compiler error linking to the xkcd comic about all modern digital infrastructure relying on thanklessly maintained open source projects.

An attempt at explaining why I'm finding it hard to get really high productivity boosts from AI-assisted programming. #AI blog.jnalanko.net/2026/02/02/t...

DSB Program is out ! Seems incredible (as ever) dsb-meeting.github.io/DSB2026/

I pointed ChatGPT5.2 at the greedy superstring conjecture, open since 1988. It could not prove it for all n, but it found simple arguments for n = 3 and n = 4 strings, which haven't been mentioned in the literature to my (limited) knowledge. n = 5 still open? blog.jnalanko.net/2026/01/20/a... #math

AI fails the Turing test again. It's too smart and knowledgeable to be believable.

I think the time has come to modernize math/CS research. Now that AI is starting to be capable of formalizing proofs, it might soon be inexcusable not to provide formally verified proofs with papers, like trying to ship software without a test suite. This would also make peer review so much easier!

Thoughts on the FrontierMath benchmark on AI. #AI #math blog.jnalanko.net/2025/12/26/t...

Why are there no European frontier AI labs?

Korkeakoulujen aloituspaikkoja ei voi mitenkään tolkullisesti lisätä kertaluontoisella rahoituksella, vaan siihen tarvitaan nimenomaan perusrahoituksen kasvattamista. Sisäänotettua opiskelijaa pitää palvella 5+ vuotta, tarvitaan pysyvä lisäys opetushenkilöstöön. www.hs.fi/politiikka/a...

"Valtiovarainministeriön kansliapäällikkö Juha Majanen sanoi [..] että ylimpien marginaaliveroasteiden keventämisen hyödyistä on vankkaa tutkimustietoa ja toimi rahoittaisi pitkällä aikavälillä itse itsensä." Mikä ihmeen vankka näyttö? www.uusijuttu.fi/juttu/sZRTks... #politiikka #puoliväliriihi

More on latest AI models. The AI deep research feature from OpenAI is impressive, but critically, it has no BS detector. I asked for a literature review of a field I know well. It repeats all the hype authors put into introductions of papers and all the cherry-picked highlights in the conclusions.

OpenAI o3 is a curious beast. It's vastly superhuman in the amount of encoded knowledge. It seems to have an encyclopedic PhD-level knowledge of every existing topic. It's able to recombine what it knows to create something new. It should be able to sketch some pretty good research projects! 1/3

I'm considering skipping all US conferences until the political nightmare is over. Thoughts? I wonder if an academic boycott would have any effect on the administration.. most likely not. Getting turned back on the border for personal messages on my phone would really suck, even if unlikely.

Today I learned that you can be denied entry to the US if your phone has personal messages that are critical of Trump. www.theguardian.com/us-news/2025...

Here is one of the problems in this year's Finnish high school math final exam (like SAT). It asks to prove this inequality with the hint below. This gave me some pause, but I found the proof eventually. Then, I tried giving it to ChatGPT. It found the same proof as I did! AI is getting good.

Vihreiden ja Vasemmistoliiton erot Helsingin kuntavaaleissa, osa 2! blog.jnalanko.net?p=779. Tein hierarkisen klusteroinnin Ylen vaalikoneen perusteella, tässä tulokset. Mukana myös vertailu Kokoomukseen. #kuntavaalit #vaalit #vihreät #vasemmistoliitto #kokoomus #politiikka.

Mitä eroa on Vihreillä ja Vasemmistoliitolla kuntavaaleissa Helsingissä? Latasin kaikki Vihreiden ja Vasemmistoliiton ehdokkaiden vastaukset Ylen vaalikoneesta ja piirsin tällaisen kuvan. Blogissa myös nostoja ehdokkaiden perusteluista pariin kysymykseen! Täällä: blog.jnalanko.net?p=687

I understand the US doesn't want to be involved in wars that do not threaten them. They have the luxury and privilege of having peaceful neighbors and two oceans around them. But how about, standing up against a bully that is beating up your friend? I know it's not *obligatory*, but... come on. 1/2

Does NATO need the US? It doesn't feel like we're allies anymore with the US. Would the US come to help Finland in a time of need? I think it's time for Europe to take the lead.

People of the US, why are you not already on the streets about this? x.com/FluteMagicia...

marcan.st/2025/02/resi... If it was possible to work with the Linux Kernel in Rust, I might be interested in contributing something someday (when I have the time, hah). Having to write C at this scale is just a no-go for me.

So, it seems that so far Bluesky has won over Mastodon. This place feels more alive. Why is that? Personally, I find the UI here nicer, and the feed more relevant to me.

This is nuts. If you told me this five years ago, I would have thought you are crazy. Solving IOI problems requires some serious problem solving skills and reasoning. AI is now a far better competitive programmer than I will ever be.