There are so many data centers planned in Finland that the electricity demand of the whole country could go up 40%. www.hs.fi/visio/art-20...

Energia | Fingrid: Suomen sähkönkulutus voi kasvaa pian jopa 40 prosenttia

Sähkönkulutuksen suuri kasvu mullistaa Suomen sähkömarkkinaa monin tavoin.

hs.fi

I used to use AI for math, and then check its work. Now I use AI to supervise *me* when I do math.

Mapping all 19k genes in the human genome into the space of... anime characters? This project is nuts. www.lesswrong.com/posts/BJ7AqX...

Mnemonic portraits for 19,023 human genes — LessWrong

Back in 2013, Scott Alexander wrote in Extreme mnemonics: …

lesswrong.com

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/...

Critical Atlantic current significantly more likely to collapse than thought

Scientists say finding is ‘very concerning’ as collapse would be catastrophic for Europe, Africa and the Americas

theguardian.com

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

Computing k-mers in Graphs

We initiate the study of computational problems on $k$-mers (strings of length $k$) in labeled graphs. As a starting point, we consider the problem of counting the number of distinct $k$-mers found on...

arxiv.org

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!

bioRxiv Bioinfo@biorxiv-bioinfo.bsky.social · 6mo ago

10-minimizers: a promising class of constant-space minimizers https://www.biorxiv.org/content/10.64898/2026.03.16.712052v1

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).

GitHub - suomela/fin-dep: Finitely dependent distributions

Finitely dependent distributions. Contribute to suomela/fin-dep development by creating an account on GitHub.

github.com

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

Venice - Blog

Venice makes no sense. I’m talking about centro storico, the historic center. The place is walkable in theory. It is only a few kilometers end to end, but getting from point A to point B by foot might...

blog.jnalanko.net

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