Recently found out about @jvkersch.bsky.social's `tmtools` github.com/jvkersch/tmt.... It works nicely! Being able to pass user-defined sequence alignments is a nice (if simple) feature that is missing from OpenStructure's own `tmtools` @torstenschwede.bsky.social.
Thank you @umbislupo.bsky.social. The user-defined sequence alignments feature was contributed by github.com/alirezaomidi. Happy to see further adoption of the package!
You can define it yourself using that other nifty LaTeX macro, \solvehaltingproblem.
Impressions:
- Very easy to get the first 90% right
- Very difficult to finish it off exactly right
- Actually a great way to learn a new API
The last is what sells it for me: being dropped right where the action is, and having the model act as a thought amplifier, not as a stand-in.
Experiment 2 with #claudecode: vibe-coded #manim animations to convey the intuition behind QQ plots, which are notoriously difficult to explain just on the whiteboard.
youtu.be/7jH5tloKHYQ
Inspiring deepdive by @sachachua.com -- always inspiring and thought-provoking sachachua.com/blog/2025/07...
What
- The tokens: burned through about $15 worth of tokens, covered by my subscription. Not sure the funding model ($20/month plan) will be sustainable...
What didn't work so well:
- Had to take the wheel a few times. Maybe not so bad: I want to spend my time on the deep stuff after all.
- The delay! So long that you get bored, not long enough that you can reliably switch away.
What went well:
- Just letting it do its thing and watching it get 95% right is magical.
- Going through multiple iterations is a breeze.
- Seamless integration with GitHub -- writing a custom command is just a matter of adding a new prompt.
In the end, took me much longer with Claude (2 days of on-off attention) than manual (2 hours of focus work), but my effort was much less. I could see myself doing 2-3 tasks that used to be intensive at the same time this way.
Very happy I converted my slide decks to Quarto last few years, this format is a breeze for Claude (both to edit and to get info about). I can't imagine doing this with PowerPoint.
First impressions with Claude Code: Add section headers to my calculus slide decks, and have autogenerated ToCs (github.com/ugent-korea/...). Overall: productive experience, but weird -- the jagged edge is real. 🧵👇
Dropping one "Make it so, Mr Crusher" has made Claude Code switch completely into Star Fleet officer mode...
... which increases the temperature of my office by about 0.002 ℃ assuming various thermodynamics nonsenses
The figure quoted in the article is 3Wh per query, which checks out: running one query with the DeepSeek 70b model quantized to 4 bits keeps my GPU busy for about 30 seconds, which works out to 1-1.5 Wh.
Plus or minus sexy?
Casually leafing through Leibniz's papers, as one does, and was surprised to see these astronomical symbols ☽☉ used as shorthand. Looking forward to using these, as well as ☿♃♆ etc in my calculus class.
This reminds me of my own journey to Python way back when (about 20 years ago). I was in a B&B without internet access and couldn't access the Matlab license server, and just decided to ditch it altogether.
Nice read; reminiscent of how many of us switched to Python after getting disappointed with Matlab's high licensing fees.
Condensed the main idea just in time for my intro class on complex numbers:
Del Ferro's formula
Started reading Paul Nahin's "An Imaginary Tale: The Story of √-1" in preparation for teaching about complex numbers. Somehow I had never seen the derivation of the formula for the roots of a cubic, and was surprised at how easy it was to follow -- and how magical it must have been to derive.
Recently I got a citation alert for a paper that extended a result of mine from 2010. Upon closer inspection it turns out that the authors are from North Korea. Crazy that we are a mere 250 kilometers apart but we may be as well on a different planet.
In the end, the cause of a mysterious slowdown turned out to be the fact that the system was out of memory and swapping like crazy. Turning off swap didn't fix the problem, but at least made it obvious.
Of course, understanding your pipeline means having the right tools to look at what's going on: htop, btop, iotop, /proc, /sys, listening to the fans, putting your hand inside the cage, all give useful information.
3. stress-ng (2/2): Occupy 8 CPUs: stress-ng --matrix 8 --tz --ignite-cpu
2. stress-ng (1/2): Stress out the system in various ways. Allocate 30GB of RAM: stress-ng --vm-bytes 30g --vm-keep -m 1
1. systemd-run: Limit the number of CPUs and available memory. Start a bash shell on 8 CPUs with 32GB of RAM: systemd-run --scope -p AllowedCPUs=0-7 -p CPUQuota=800% -p MemoryMax=32G bash
