
Startups, robotics & AI. @Stanford alum. Building @komment_ai, AppSec for the AI era. 🇩🇪/🇺🇸
New York, USA
Joined January 2014
- Tweets354
- Following267
- Followers357
- Likes3.9K
College is one of the worst financial decisions you can still make in 2026 unless you’re going into a hard STEM field.
Would you rather:
A) Build a product that 10,000 people love
B) Build a product that 1 million people mildly like
5 self-improvement habits that look productive but actually slow down shipping:
1. Endless note-taking systems
2. Perfect morning routines
3. Consuming more “frameworks” than you ship
4. Optimizing tools instead of using them
5. Journaling about the work instead of doing the work
Most of these feel good in the moment.
Very few of them move the needle on the only metric that matters: things that exist in the world.
Alibaba open-sourced RADAR, a generalist model that reads one abdominal CT and flags ~146 findings across 18 organs. It beat 23 of 26 radiologists head-to-head and is published in *Science*.
The really cool part is that the training signal came straight from clinical reports with no manual annotation. That scaling path is far more realistic than most medical AI papers pretend.
github.com/alibaba-damo-acad…
science.org/doi/10.1126/scie…
Honest choice:
A) Ship something mediocre every week
OR
B) Ship something excellent every 3 months
Which one actually wins long-term?
Dude... just stop it or @julianwhatif will have another field day with your head
Waymo + Nihon Kotsu + GO targeting fully driverless robotaxis in Tokyo in 2027. First commercial unmanned service in Japan.
Tokyo traffic + Japanese regulatory culture is about to become one of the most interesting real-world tests of whether the last 5 % of autonomy is actually solvable.
waymo.com/blog/2026/09/openi…
Still can’t get over how much better Formula 1 telemetry is than most production observability stacks. Live, high-resolution, every driver, every corner, and the engineers actually look at it.
Unitree open-sourced UnifoLM-WLA-1.0 (6 B params, ~2 500 hours of real-robot data). One model does both desktop and whole-body mobile manipulation.
The real move is realizing the scarce resource is high-quality robot data and cheap hardware at scale. Give away the brain, sell the bodies 🤖
github.com/unitreerobotics/u…
The nicest thing about living between Germany and the US is watching two completely different definitions of “on time” collide in the same calendar invite.
Die Linke holt in Berlin knapp 25% und wird erstmals seit der Einigung stärkste Kraft. CDU verliert massiv, SPD fällt auf ein historisches Tief.
Das ist kein Linksruck aus Begeisterung. Es ist eine Abstrafung der regierenden Mitte bei gleichzeitig höchster Wahlbeteiligung seit Jahren.
Wohnen hat die Wahl entschieden, nicht Ideologie.
xitter.cf/Wahlen_DE/status/21017…
Most “temporary” solutions in software become permanent the moment the person who built them changes teams.
The half-life of a temporary flag is longer than most people’s tenure.
It doesn't help that LLMs love inserting flags and gates any chance they get into the code.
Most product “roadmaps” are just reverse-chronological lists of things that already shipped, written after the fact so the board can feel like there was a plan.
The real planning usually happens in the three days before a deadline when everyone finally admits what is actually possible.
Tick-Tach is a DIY timegrapher mic for mechanical watches. 30 mm piezo disc, simple transistor preamp, 3D-printed case with a recess for the watch, works with open-source software on PC and Android.
Commercial telegraphers cost real money. This one lets hobbyists actually measure rate, amplitude and beat error without buying the professional tool.
hackaday.io/project/203812-t…
I still prefer tools that let me see the entire mechanism. Mechanical watches, old keyboards, simple analog circuits.
Once something becomes a sealed black box that only fails as a unit, the interesting part disappears for me.





