Databases and distributed systems. Working at Datadog on Metrics.
New York, NY
Joined April 2009
- Tweets498
- Following250
- Followers632
- Likes33
What a time, M-VAVE FM1 - a pocket FM synthesizer, was reverse engineered with LLMs. Now with custom firmware it can do a lot more, it can be a groovebox with drum machines, a sequencer and different sync engines. It can also play Doom and emulate GameBoy.
I of course couldn't resist buying the device. It's not much larger than a phone and feels like a toy, but I paid only $52.99.
There are 140 firmwares now, with 30 or so released just this week, most obviously vibecoded fm1-editor.com/firmware/. I think this is the kind of malleable software that we've been hearing about since LLMs became good at writing software that compiles, runs and kind of works.
Wtf is wrong with Silksong, 7 minute runback from a bench and 5 waves of enemies before a boss fight. I don't remember anything like this in Hollow Knight.
Intuition developed over the years - if a new LSM key-value store claims it's 10x faster than RocksDB, it's very likely using mmap and the whole benchmark dataset fits in memory.
I keep hearing how Sol is better than Opus, but it has not been my experience. I've been giving Claude + Opus 5 and Codex + GPT-5.6 Sol the same tasks for the past two weeks. Opus is clearly better at research and produces more simple and better solutions.
Sol randomly stops and forgets things (might be Codex issue?). The only area where Sol is better is code reviews. Both are bad at writing English, Opus is worse, but you shouldn't be using LLMs for writing intended for humans.
Claude Code's auto memory gets stale quickly. It kept making wrong assumptions based on something it decided to remember from two architecture iterations ago, made me disable the feature entirely.
Polonius is on by default in nightly Rust blog.rust-lang.org/2026/08/0… looking forward to getting rid of a few copies and unsafe in performance sensitive code paths once it's in stable.
Artem Krylysov retweeted
Software infrastructure at a glance (July 23, 2026)
NYC Apache Iceberg Community Meetup in New York next month. @akrylysov shares how MVCC and transactions work in RocksDB. Atoms raises a $1.7B round. And @DatabentoHQ is hiring a C++/Rust developer.
New post on how MVCC is implemented in LSM-trees and how RocksDB builds transactions on top, with a benchmark comparing optimistic and pessimistic transactions artem.krylysov.com/blog/2026…
Solved the permission fatigue from Claude by running it with dangerously-skip-permissions inside a VM that only has access to the host code directory.
claudevm() {
limactl shell --workdir "$PWD" default -- claude --dangerously-skip-permissions "$@"
}
Turns out my local wiki setup, which is just a pile of markdown files tracked in Git that I've been using for a decade was so ahead of its time. With AI it's a Second Brain™ now.
After a decade of wanting to remove a CSS framework from my personal website, it's finally done thanks to Claude! The first attempt, where I just asked Claude to get rid of the framework, didn't go well, the website looked very janky.
This is similar to my overall experience with LLM driven coding. The LLM output gets better (at solving the task, not necessarily better code) as you give an LLM better ways to verify its output.