Jan Ole Suhr (@janole) header
Jan Ole Suhr (@janole) profile picture

Hands-on, open-minded Product Engineer since 1986. Entrepreneur at heart. Former InsurTech CTO/CPO. Creator of Gravity, the social net app for Nokia w/ 1M users

Berlin, Germany
Joined July 2007
Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)Photo by Jan Ole Suhr (@janole)
GLM 5.2 has taken control of a tmux session with Claude Code in it. It's watching and waiting for approval prompts. Then it's trying to approve or reject them, but that tool call itself is running through another GLM 5.2 approval agent 🤯 The fact that I can tweet this means neither of those agents accidentally wiped my MacBook yet ... 😂🤣
3
369
Noooooooooo! cache expired · rewrote 122k (idle 1h 1m)
154
Jan Ole Suhr retweeted
building systems is like assembling lego now it’s so awesome 🤩
9
5
1
46
1,856
Very interesting analysis from Opus 5.5 on my session traces: With gpt-6-astra, one minute took about 5.6% of the 5h limit, while gpt-6.2-astra only takes 1.3%.
First tests with gpt-6.2-astra are promising. On my cheap $20 sub, the 5 hour usage limit lasts for about 60-80 minutes now ... that's a considerable improvement. "time ~58 min / tokens ~17.0M / uncached ~168k (1%) / output ~25k (1.3k)"
2
175
First tests with gpt-6.2-astra are promising. On my cheap $20 sub, the 5 hour usage limit lasts for about 60-80 minutes now ... that's a considerable improvement. "time ~58 min / tokens ~17.0M / uncached ~168k (1%) / output ~25k (1.3k)"
1
1
1
296
The previous gpt-6-astra took about 15-20 minutes to reach the 5h usage limit of a $20 sub and rarely to never finished a task in one stint.
52
Jan Ole Suhr retweeted
Both NeXT and 3DFX tried to “do the right thing” with color spaces, and it wound up generally making them look worse than competing options. I didn’t really understand it properly until years later. It is natural to think that a pixel with a value of 1.0 is emitting twice the photons as a pixel with a value of 0.5, so you can do the obvious linear math to calculate partial pixel coverage, blending a foreground over a background, or interpolating between colors in a texture. If displays are “linear”, that is correct. NeXT and 3DFX were much closer to linear than other contemporary displays. NeXT had crossover from the deep graphics understanding at Pixar, and 3DFX had the SGI heritage. They did this intentionally. If you have 10+ bits per channel like the professional rendering systems, linear displays are The Right Thing. Unfortunately, NeXTStation Colors were 4 bits per channel and 3DFX was 5 or 6 bits per channel, and there… it was not the right thing. Human eyes aren’t linear intensity sensors. We mostly register relative changes in intensity, so we can tell the difference between 0.01 and 0.02 as a huge doubling, while 0.99 to 1.00 is almost imperceptible. This distribution of intensities is often referred to as the “gamma” of the display system, but the actual curves like sRGB are a bit more complicated. If you only have a small number of discrete pixel values, you definitely don’t want to spread them evenly from 0 to 1! If you do, you will have good precision in the brights, but very stark “intensity bands” in the dark colors. If everything was authored specifically for 3DFX, you would still have the fundamental problem with the dark bands, but more accurate blending and filtering. The problem was that artists made all of their textures in paint programs on conventional non-linear displays, so when they were drawn by 3DFX cards, they got the characteristic “washed out color” look. You can mathematically correct textures for differences in display gamma, but I didn’t really understand the issue back in the glQuake days, and we were so enthralled by the absolute performance of 3DFX that the color fidelity didn’t get as much attention as it should have. Even today, tons of games don’t properly handle gamma across their textures and render targets. High resolutions and 24 bit color masks most of it, but you still see it in some slight shimmerings in high contrast textures and edges.
Does anybody know why 3DfX 1 cards had such a washed out color output? In my research, circa '97 I found many people complaining about the colors. Is it gamma? 16-bit color? What can it be?
154
497
45
6,621
205,211
I don’t want to live in a world where my graphics shaders (using modern tensor cores) must be approved by the government. We don’t want to live in a dystopia, where goverments dictate what code our computers are allowed to run.
Don't make the mistake of thinking this is about his infidelity. These articles will not stop until he agrees to follow Anthropic's proposal to block inference for unsigned and registered open-models in the firmware of modern GPUs. This is a pretty transparent concerted effort at centralizing control over compute ressources and the ability to execute ever more powerful local models. Jensen is the tip of the spear pushing back at this (in part for his own economic reasons). I'm team Jensen, his affairs are none of my business, not even the alleged support of circumvention of export controls are my business, his position on open-weight models is.
22
92
3
1,003
19,127
Jan Ole Suhr retweeted
Why are old stories that were not even covered up all that well become sudden international and geopolitical news? The timing is suspicious, so worth asking: who benefits and why now? If Jensen, the premier champion of open models, is taken out then anyone who wants AI closed, centralized, and controlled might benefit. Why now? The midterms seem relevant. But also perhaps the upcoming IPOs.
Breaking: Nvidia’s board is preparing to oust CEO Jensen Huang after he had two extramarital affairs, including one with a suspected Chinese spy. on.wsj.com/4jcpr01
291
762
114
6,351
477,998
Jan Ole Suhr retweeted
In 10 years as CEO of two startups, I can't count how many times engineers pitched me a complete rewrite of our codebase. It always reminded me of Steve Blank's Startup Suicide: Rewriting the Code and Joel Spolsky's Things You Should Never Do. Joel compared rewriting software from scratch to demolishing a building instead of renovating it. I always agreed and pushed back hard on rewrites. But AI has completely changed my mind. When you can coordinate a swarm of thousands of coding agents, the economics are completely different. Suddenly there's an enormous premium on having an agent-native codebase: super fast CI, minimal dependencies, clean architecture, every decision documented, and everything optimized for agents rather than humans. For the first time in my career, I'm starting to think a complete rewrite might actually be the right call.
The thing with AI: You REALLY can replace existing software with something that is much faster in performance, much cheaper to operate, as reliable (or more!) as before, with shockingly little effort when you know how to use AI Expertise has never been this valuable!!
59
17
8
411
39,868
Jan Ole Suhr retweeted
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
3,205
9,250
6,970
96,503
21,095,569
Always good to analyze all your sessions. Apparently, Opus 5.5 was switching away from using Bash for reads/writes/edits from 11-05 to 11-07. Back to normal now. I didn't notice any difference in quality, and usage was feeling normal, too ... 🤷
3
215
Argh! I seem to have been targeted by another Anthropic A/B test. Probably worth monitoring regularly. It's when you see lots of Bash usage ("sed ...", "cat ...", ...) instead of Read/Edit/Write. For me, "bashFirst" works better.
1
1
200
It's nice that you can just use Claude to analyze itself: "Since about 09:13 UTC today the server has Bash-first switched on for us anyway, so nothing urgent is wrong right now. ..."
66
Jan Ole Suhr retweeted
Cybersecurity expert #1: AI is becoming superhuman at finding vulnerabilities and soon no computer will be safe Cybersecurity expert #2: also our private key is '4' and if you mash the keyboard long enough it logs you in as root
*sshole is out if you run sshd on alpine, openwrt or any musl box with password auth on: update now official writeup: sshhole.com
5
37
3
244
5,940
Wild bug …
Background subagents in Claude Code now use less of your usage limit. A running subagent updates its status line every 30 seconds, and each update used to re-read its whole conversation. Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less.
172
When the agent starts to talk to sub-agents although it shouldn't and I'm reading ... "Great work — the operator loves the Plan tab." ... I'm thinking to myself: are they alive!? 🥰
3
224
Jan Ole Suhr retweeted
Sorry bro, weekend is for the lazy, I let my robo clean up my desktop (sound on)
120
174
58
3,045
270,721
Jan Ole Suhr retweeted
the moat is to be a solo founder
there is no moat
4
1
44
2,000
🧐 „Claude Code burns ~15–20% of your tokens on a 3-word status line. Here's how to turn it off“ gist.github.com/flamerix-ai/… Not sure if that’s really true. Can’t reproduce it (I am using Claude in my own harness with my own remote-/sub-agents only.)
159