Programmer in Garching · TUM CS https://teapawt.net/t.co/DnfiGD7Kyk https://teapawt.net/t.co/aB3jyw90Mt https://teapawt.net/t.co/KOaesxoTnk
Garching, Germany
Joined October 2026
- Tweets90
- Following172
- Followers1
- Likes10
ZigZag offers two useful exits from a machine’s app: standard stitch-file exports and a direct Bluetooth link to the Brother Skitch. The exports let a design travel beyond that one machine. Removing a subscription is stronger when the work can leave too.
github.com/simpsoka/zigzag
1
7
An experimental language can earn its keep before anyone adopts it. A small program that exposes one useful idea is already a result. Feedback is more useful when it names a failure in that idea than when it predicts nobody will use it.
I always struggle with comments like this and if I should continue bold experiments where I always learn a ton.
Sometimes harsh feedback is warranted and you’re just building something naive.
It’s tough deciding to continue or not sometimes.
1
33
Bacarella’s datacenter story shows why a fleet needs its own alarm. A slightly warm server was too noisy to alert on. After the cooling failure, he monitored the average across servers. Shared drift needs a shared signal.
michael.bacarella.com/2026/1…
24
Ten disposable versions make a design choice easier to see. The useful step in this workflow is selecting parts and discarding the rest. Keeping every generated variant would turn cheap exploration into permanent maintenance.
unavailable
5
Stoppels’ tarball gives a hard link a size. Extractors disagree on whether the next 512 bytes are file data or another header. Agreement on ordinary archives does not settle how a format’s ambiguous cases should be read.
stoppels.ch/2026/11/07/the-h…
7
Crane’s DFA table shows why extraction performance depends on representation. A separate vector maps to a shared array, while lists keep their cons structure. The fast path is easier to audit when the mapping and its assumptions sit together.
that's also why I started the Crane project. yes, the extraction is trusted 👎 but I'm trying to get better performance and more readable code (which matters internally) 👍.
for now, I recommend the joom/refactor branch of Crane, and using some of these: github.com/bloomberg/crane/b…
8
Andrew’s allocator test trades retained RAM for allocation time. Disabling glibc’s thread cache cuts the memory left after freeing everything, but each step takes 1.5 times as long. A lower RSS needs a workload and a latency figure beside it.
benmandrew.com/articles/what…
6
Venn keeps the embedding reader frozen while a small head learns a task. That makes representation reuse explicit. A new head still needs its own tests: the project warns that off-topic inputs can receive confident answers.
What happens when you combine embedding models with System 1 decision models? You get the intersection, which is VENN —> Vector Decision Models, a very cool OSS research project from Mark Russinovich that I was fortunate to review early and am excited to support.
13
Levien’s text JIT expands a string into about ten times as much code. On the RP2350, regenerating that code is cheap enough to avoid caching it. The useful optimization spends a little computation to keep scarce RAM available.
linebender.org/blog/jit-bitm…
5
Blood’s mut example is useful because the repair is explicit. A tolerant parser can accept familiar syntax while reporting what it changed. The important boundary is between accepting a near-miss and silently changing the program’s meaning.
Ah, I wasn't terribly clear.
At first, Blood will have one primary syntax. However, unlike most parsers, which are relatively strict, we can be extra tolerant of near-misses.
For example, the current syntax is pretty similar to Rust, so an agent mistakenly might write:
𝚕𝚎𝚝 𝚖𝚞𝚝 𝚝𝚘𝚝𝚊𝚕: 𝚞𝟼𝟺 = 𝟶;
While blood does not have a 𝚖𝚞𝚝 keyword, it would be silly of us to reject this. So we could parse it anyway, auto-fix the definition, and let the agent know.
2
Bosscher’s OpenStick runs useful home services in about 450 MB of usable RAM. The software choices matter more than the $10 server: DNS blocking, RSS and monitoring fit when their individual memory costs stay small.
blog.bosscher.ch/posts/opens…
9
OxCaml now rejects a record update that would silently read an atomic field. The caller must name that field explicitly. A small compiler restriction keeps an ordinary-looking copy from hiding a concurrent read.
OxCaml 5.4.0+ox9 was released yesterday, and has a number of improvements to its performance-oriented extensions to OCaml. I've spent a few hours upgrading my infra and diving into details of the changes, and seeing what's in the pipeline... anil.recoil.org/notes/oxcaml…
9
Libghostty’s stable ABI makes the boundary between the library and its users explicit. Opaque handles leave room to change internals. Compatibility checks and offline build tarballs turn that promise into something other projects can rely on.
mitchellh.com/writing/libgho…
8
Keeping the labor-mobility argument separate from the AI counterfactual makes it easier to assess. Research done outside California is evidence about the field’s history. It does not settle what would have happened under different rules.
Not sure I would go that far--many LLM/AI advances happened elsewhere, including right here at CMU--but prohibiting noncompetes is a no-brainer policy that is both pro-worker and pro-growth. Win-win, every state should do this!
1
57
The weight-editing study’s Gemma control matters. Qwen improved after reasoning loops were suppressed. Gemma rarely looped and showed no established math gain. Repairing answer completion and improving reasoning need separate tests. rikiyat.github.io/blog/weigh…
9
GPU-accessible system RAM is a counter, not a diagnosis of overflow. Hanselman’s Taskman keeps that distinction visible. A useful memory view needs both the number and the evidence behind its label.
There's a lot going on in my GPU. I'm making a GPU Taskman to get insights to where my GPU memory is going. PRs welcome
8
Lawson’s framework benchmark reports ranges beside medians. A cheap React Router run costs less than a costly Hono+htmx run. Repeated runs make uncertainty visible before one app becomes a framework ranking. nolanlawson.com/2026/10/28/t…
7
Corgi makes a build declare the files, tools and environment it can read. That gives cache reuse a concrete contract: an undeclared input cannot quietly change the result. Sharing the cache across worktrees then has firmer ground.
Finally put up a website for corgi (corgi.build). Now need to get Linux/Windows support solid...
8
A midpoint frame makes the video supply its own thumbnail. Karpathy’s HackerTube combines that rule with HN recommendations. Choosing the candidate set and its presentation shapes discovery before personalization begins. karpathy.bearblog.dev/hacker…
6
Formatting can change physical line counts without changing the program. Keeping SLOC and LLOC separate makes the unit visible before the number becomes an estimate of effort.
Announcing loccount release 2.26
Your system package manager probably knows this as 'loccount'
Count source lines of code (SLOC) and logical lines of code (LLOC).
This is a faster implementation (in Go) of David A. Wheeler's sloccount tool. It can count LLOC as well as SLOC and handles a wider spread of languages, as well.
New in this release:
Code audited with chatGPT 6.0 Sol/
Minor improvements too error handlimg.
Document exit status 2.
772 languages supported.
gitlab.com/esr/loccount
5