HNHacker News
TopNewBestAskShowJobs

hannesfur

330 karma · joined March 11, 2024

I am CTO of Stillwind. We make AI tools for Electrical Engineers. I have a background in decentralized systems and networking. Contact me at hannes[@]stillwind.ai
submissionscomments
hannesfur··on USA Today Co., partners with Palantir to analyze audience data
With that going on, it's hard to imagine there once was a time where the news paper publishers hat little to no signal on what actually sold news papers. Given that papers were only sold as a bundle of articles and most sold through subscribers, they pretty much had to guess. So they had a rough direction they knew their readers expected and they had some sparse feedback from the "letters to the editor" but you had to be really angry to follow through on writing a physical letter, getting a stamp and walk to the post box just to complain.

I would argue, that helped journalism because journalists probably assumed that their readers would hold them to the same standards they hold the writers of the other papers too.

hannesfur··on Show HN: Stillwind – PCB part selection as constraint solving
We have a focus on more complex ICs as we see them most relevant for this task. In that area we have ~95% of all components that are active (not EOL & recommended for new designs). But we try to have a little bit of every category to make most projects possible. While we expand our database, you can use the external search agent which uses the internet to track down parts that are missing.
hannesfur··on How We Made IPFS Content Publishing 10x Faster
Having worked on libp2p‘s DHT (Double Hashing for rust-libp2p) for a bit two years ago, it’s really great to see that there are improvements. To get to CDN level speeds though on dense networks, I still see it as an architectural flaw to not somehow encode network topology into the PeerID / identity in the DHT. A start would be to use the five RIRs. If you want to be more sophisticated, and I spent a lot of time theorising about this, you could have a dezentrally governed anycast IP address of Geo DNS to bootstrap new peers into their neighbourhood and couple that into their DHT identity. But do you want to put BGP into the hands of a decentralised system? Could you even do it in the governance structure of the internet?

Btw when we were working on our project HyveOS, we used Batman-advs routing table to quickly (really really quick) bootstrap new peers into the system.

Ah… sometimes i really miss working on this.

hannesfur··on Ubiquiti: Enterprise NAS, Built on ZFS
It says 8 Arm Neoverse N2 cores in the blog post. So not directly ARM Cortex, derived from ARM Cortex-X3 but same family as NVIDIA Grace, Google Axion and AWS Graviton4.
hannesfur··on ESP32-S31: Dual-Core RISC-V SoC with Wi-Fi 6, Bluetooth 5.4, and Advanced HMI
Not across all features, but certainly in specific ones. There are more advanced WiFi 6 chips, more advanced Bluetooth chips and faster MCUs. But they are all separate chips or companion ICs.
hannesfur··on ESP32-S31: Dual-Core RISC-V SoC with Wi-Fi 6, Bluetooth 5.4, and Advanced HMI
It has become limited. Some have WiFi, some have Bluetooth but rarely both. There is the SiWx917M (https://www.silabs.com/wireless/wi-fi/siwx917-wireless-socs) by SiliconLabs though.

If we include western companies you get the NXP RW612 (https://www.nxp.com/products/wireless-connectivity/wi-fi-plu...) which is dutch.

hannesfur··on ESP32-S31: Dual-Core RISC-V SoC with Wi-Fi 6, Bluetooth 5.4, and Advanced HMI
There is an ESP-C5 (https://www.espressif.com/en/products/socs/esp32-c5) that supports 2.4 and 5 GHz dual-band WiFi. And unlike the ESP32-S31 you can buy it today.
hannesfur··on Artemis II crew take “spectacular” image of Earth
From the EXIF we can infer that every setting was left at the default. No exposure comp, no contrast, no HSL, no lens correction and a linear tone curve. Just the default Adobe Color profile at 5400K.
hannesfur··on Artemis II crew take “spectacular” image of Earth
Looking at the EXIF (with exiftool) for the image uploaded by NASA (https://www.nasa.gov/wp-content/uploads/2026/04/art002e00019...), apparently this was taken by a Nikon D5 with an AF-S Zoom-Nikkor 14-24mm f/2.8G ED and developed with Lightroom. It also seems like very little was done in Lightroom. Amazing... I dumped the whole EXIF here: https://gist.github.com/umgefahren/a6f555e6588a98adb74eed79d...
hannesfur··on Embassy: Modern embedded framework, using Rust and async
That’s fair but when there is an async version of the driver or Hal available it should be pretty straightforward to port it to synchronous, right? Maybe Claude code can even do it with minimal supervision…

Edit: Replace blocking with synchronous

hannesfur··on Embassy: Modern embedded framework, using Rust and async
I am a big fan of the embassy project and it’s a great example of why async Rust is so great: Because this is possible. It works without a heap, is a really low cost abstraction and you can do stuff concurrently on a single core chip (where you can’t just spawn a new “thread”) and you don’t have the complexity of an RTOS. I believe there is a great future for embassy ahead and it’s so great how far the team has come.

I also want to give a shoutout to reqwless (https://github.com/drogue-iot/reqwless) which is a HTTP client for embassy-net that even supports HTTPS!

Rust embedded was really never actually better then C or C++ but embassy for me is a big reason why I now make my buying decision based on how well I can use Rust on the MCU.

hannesfur··on Embassy: Modern embedded framework, using Rust and async
Maybe stuff has changed a lot in the last year but I didn’t experience that problem so far. For me it was the other way around mostly. Where did you encounter that?
hannesfur··on Efficient method to capture carbon dioxide from the atmosphere
Interesting thought but you would need a lot of these gasses on the one hand and on the other hand it doesn’t help in working against the greenhouse effect. The greenhouse effect depends on the absolute amount of CO2 in the atmosphere, not the percentage. How much infrared light is absorbed by CO2 primarily depends on the amount of CO2 in the atmosphere.
hannesfur··on GPT-5.2
Maybe they felt the increase in capability is not worth of a bigger version bump. Additionally pre-training isn't as important as it used to be. Most of the advances we see now probably come from the RL stage.
hannesfur··on ETH-Zurich: Digital Design and Computer Architecture; 227-0003-10L, Spring, 2025
This course is actually mandatory in the first year of the CS undergraduate program here at ETH. I remember it very fondly for its great (and passionate) lecture and the hands on experience building a MIPS cpu in the exercise sessions. Probably the best lecture in my undergraduate.
hannesfur··on Upcoming Rust language features for kernel development
These seem like the first features that Rust in Linux bring to the Rust language that are not almost exclusively useful to the kernel. In my perception the focus on bringing features for the kernel has held up development in other parts of the language and the standard library.
hannesfur··on Zed is now available on Windows
It's great that they finally target all three mayor platforms! Let's see how developers using Windows treat them. I used to be a Neovim purist, spending days on my config and I have barely touched it since I moved to Zed. It's so much nicer to use then VScode (and it's forks) because it's so snappy. I hope they ship leap.nvim (https://github.com/ggandor/leap.nvim) support soon, then I am completely happy!
hannesfur··on Zed for Windows
It's great that they finally target all three mayor platforms! Let's see how developers using Windows treat them.

I used to be a Neovim purist, spending days on my config and I have barely touched it since I moved to Zed. It's so much nicer to use then VScode (and it's forks) because it's so snappy.

I hope they ship leap.nvim (https://github.com/ggandor/leap.nvim) support soon, then I am completely happy!

hannesfur··on Apple M5 chip
That's sadly true, over in x86 land things don't look much better in my opinion. The corresponding accelerators on modern Intel and AMD CPUs (the "Copilot PCs") are very difficult to program as well. I would love to read a blog post on someone trying though!
hannesfur··on Apple M5 chip
Edit: Foundation Models use the Neural Engine. They are referring to a Neural Engine compatible K/V cache in this announcement: https://machinelearning.apple.com/research/introducing-apple...
hannesfur··on Apple M5 chip
You probably wouldn’t with a Pro but you might between an iPad Pro and an MacBook Air. With the foundation models API they basically said that there will be one size of model for the entire platform, making smarter models on a MacBook Pro unrealistic and only faster ones possible.
hannesfur··on Show HN: Osaurus – Ollama-Compatible Runtime for Apple Foundation Models
Very cool! Since it’s possible to train foundation model adapters, is a library for user fine tunes possible?
hannesfur··on Apple M5 chip
That would be an interesting approach if true. I hope someone gets to the bottom of it once we have hardware in our hands.
hannesfur··on Apple M5 chip
Oh, I overlooked that! You are right. Surprising… since Apple has shown that it’s possible through CoreML (https://github.com/apple/ml-ane-transformers)

I would hope that the Foundation Models (https://developer.apple.com/documentation/foundationmodels) use the neural engine.

hannesfur··on Apple M5 chip
Thats true! I was referring to their wider line up, especially the iPad, where users will expect the same performance as on the Mac’s (they payed for an Mx chip) and they sold me an iPad Air this year that comes with a really fast M3 and still only 8 GB of RAM (you only get 16 on the iPad Pro btw if you go with at least 1TB of storage on the M4 Pro one)
hannesfur··on Apple M5 chip
As pointed out in other places as well a better comparison will be the upcoming Pro & Max variants. Also, as far as I know, Strix Halo mainly uses the GPU for inference not the little AI accelerator AMD has put on there. That one is just to limited.
hannesfur··on Apple M5 chip
> All the neural accelerators in the world won't make it competitive in speed with discrete GPUs that all have way more bandwidth.

That’s true for the on-GPU memory but I think there is some subtlety here. MoE models have slimmed the difference considerably in my opinion, because not all experts might fit into the GPU memory, but with a fast enough bus you can stream them into place when necessary.

But the key difference is the type of memory. While NVIDIA (Gaming) GPUs ship with HBM memory ship for a while now, the DGX Spark and the M4 use LPDDR5X which is the main source for their memory bottleneck. And unified memory chips with HBM memory are definitely possible (GH200, GB200), they are just less power efficient on low/idle load.

NVIDIA Grace sidestep: They actually use both HBM3e (GPU) and LPDDR5X (CPU) for that reason (load characteristics).

The moat of the memory makers is just so underrated…

hannesfur··on Apple M5 chip
I inferred that they meant the neural engine cores by neural accelerators or it could be a bigger/different AMX (which really should become a standard btw)
hannesfur··on Apple M5 chip
It’s unfortunate that this announcement is still unspecific about what they improved in the Neural Engine. Since all we know about the Neural Engine comes from Apple papers or reverse engineering efforts (https://github.com/hollance/neural-engine), it’s plausible that they addressed some quirks to enable better transformer performance. They have written quite interesting papers on transformers on the Neural Engine:

- https://machinelearning.apple.com/research/neural-engine-tra...

- https://machinelearning.apple.com/research/vision-transforme...

Things have definitely gotten better with MLX on the software side, though it still seems they could do more in that area (let’s see what the M5 Max brings). But even if they made big strides here, it won’t help previous generations, and the main thing limiting Apple Intelligence (in my opinion) will continue to be the 8 GB of unified memory they still insist on.

hannesfur··on Apple M5 chip
I agree—the difference between the different compatibility layers and native games is very steep at times. Death Stranding on my M2 Pro looks so good it’s hard to believe, but running GTA Online is so brittle and clunky… Even when games have native macOS builds, it’s rare to find them with Apple Silicon support (and even rarer with Metal support). There is a notable exception though: Arma 3 has experimental Apple Silicon support, though it comes with significant limitations. (Multiplayer, flying & mods) Although I don’t believe it’s in Apple’s interest, gaming on Linux might become an option in the future, even on Mac, but the lack of ARM builds is an even bigger problem there…

Since I am playing mostly MSFS 2024 these days I currently use GeForce Now which is fine, but cloud gaming isn’t still quite there yet…

Page 1 of 2Next →