rckoepke

Please link me to anyone pushing back against the precedent of US v. Heppner, S.D.N.Y. (2026). Would love to learn more! I'd rather not spout off on things that I'm not an expert on if I can help it, and I may have commented a bit fast above.

There's no digital attorney-client privilege. If a lawyer uses LLM's on your behalf, the communications are protected. If you use the same LLM's on your own behalf, the chats are not protected.

DOJ could cut Z.ai off from US-based payment systems (Mastercard/Visa) if they don't comply with subpoenas.

Indeed. Would be nice to have a rock-solid “correct” example to critique and learn from. And a write-up from an expert which goes into detail about all the traps for new players, what failure states / eFuse settings could be truly unrecoverable even if you have remote control of the MCU’s PWR / COM / RST pins, etc.

Years ago I had to design some robust OTA systems for both ATmega’s and ESP32’s. I was very confident that my ATmega solution could recover from absolutely any failure state that wasn’t a true “not my fault” hardware failure. At the time though, I was never quite sure if I properly covered every last edge case of the ESP32, which has significantly higher complexity of things that can go wrong from a bad OTA firmware update.

Just FYI, LoRa / LoRaWAN (Part 15) explicitly allow even unlicensed operators to utilize encryption. WiFi and Bluetooth, similarly, of course (also under Part 15). And surprisingly long range wireless network bridging like the Ubiquiti Building Bridge XG over 5GHz and 60GHz to bridge two building's LANs up to 500meters away. Part 15 also covers the unlicensed usage of the 900MHz band popular for citizen mesh networks at 2-15Mbps.

MURS (Part 95J) allows unlicensed operators to utilize encryption for digital data transmission as long as it's not encrypting voice communications, though it disallows store-and-forward packet operation and repeaters which makes doing anything similar to Reticulum/Meshtastic rather difficult.

For $400-800 nearly anyone can get a FCC Part 90 Nationwide Itinerant Data License which would allow them to do most of the "fun stuff" they might like to do for just $45-75/year, which is pretty darn affordable. Part 90 explicitly allows encryption, both of voice and data.

FCC Part 96 (CBRS) allows "unlicensed" (licensed-by-rule) operators to run their own private LTE/5G cell sites, if you have a WAN connection like Starlink or whatever. Obviously encryption is allowed.

Encryption is really only hard-banned for "amateur" radio (Part 97), and voice rules of Part 95 (i.e. CB / FRS / GMRS).

*piqued

> During the winter time the GPU also doubles as a 300W in-house heater.

Just a reminder that heat pumps can consume 300W of electricity to provide 1200W of heat.

Chrome has a browser extension API which allows plugins to access all cookies, but its use is considered suspicious and a red flag; an extension which uses it would generally get caught during initial review. However, Chrome extensions are also allowed to “hotload” portions of their own code/scripts from external 3rd party servers.

So an extension will seem benign when it initially gets checked by Google as part of becoming part of its submission to the Chrome Store. Then, later, the external “3rd party” script that is hosted remotely will get replaced with a different, malicious script. The malicious extension carries on stealing cookies, credentials, and fingerprints until someone reverse engineers it and reports it to Google.

Google will not always recognize the issue immediately because the 3rd-party malicious code is not strictly “part of” the extension so there’s a bit of a song and dance while the person who reversed it convinces Googles reviewers that “yes, this really is actually malicious, you need to analyze the third party code that loads later” and then Google eventually takes it down after a semi-involved back-and-forth where extensive documentation and video walk-throughs are provided by the exasperated white-hat Good Samaritan.

Revisiting this. Isn't it a bit of a red herring to enquire about the number of 2-faces that an n-simplex has? It still only has n+1 vertices. A 768-simplex may have 75.5 million faces but it will still only have 769 vertices which completely define the shape. So why would I expect a large number of the other >90% of the 10,000 samples I have to lie on the surface, rather than inside the interior volume?

To be more direct, what's the specific relevance of bringing up the number of 2-faces that an n-simplex has?

I believe the answer is:

(n+1)!/((k+1)!*(n-k)!) where n=768 and k=2

Or about 75.5 million triangular faces. Which explains a lot. Thanks for that.

> The real problem is that in dimensions that high, the point set probably already is the hull and all this is a zero signal gain operation.

Well, if I have 10,000 samples of a 768-dimension volume, most of those points will probably be inside the volume, and not per se a vertex of the hull.

I’m very comfortable rolling my own solution, so thank you for pointing me to Jarvis’ algorithm!