Hacker News Reader

Closed model unfortunately, continues the theme of TTS / STT being really not great for local deployments.

At least the one I’ve seen in person appeared to be literally that, an android phone without the phone bit; so an intentionally awkward user experience for no purpose.

drum55on OpenPrinter

To amuse myself I bought a modern Epson dot matrix printer with all the convince features like USB and tractor feed. Yes it works no matter what but it’s a good reminder that we do actually care about the quality of print jobs.

drum55on I never authorized an integration that scans my secrets

"You can just opt out" is the worst possible response for something the user doesn't want, and didn't know about in the first place.

drum55on Registration without a phone number on Signal will use zero-knowledge proofs

I've literally never seen anybody mention it, much less use it since it was announced.

drum55on Apple Unveils iPhone Duo

Right, because it scratches some nostalgia thing that people have obtained osmotically. Nobody actually wants a replaceable battery that makes their phone vastly thicker and harder to waterproof, nobody wants phones that randomly catch fire because their OOPLMXC brand battery from amazon wasn’t up to snuff, nobody wants to be carrying a pocket full of extra batteries to swap all day. Devices with field replaceable batteries just sucked a lot more than people don’t remember. As much as people clamor online about how they want these things, given the option they will shy away from it because it doesn’t match the fantasy that’s been created about “what has been taken from us”.

drum55on It took a year to ship WebAssembly in Anubis

That’s because its weird and mostly irrelevant, it’s nothing to do with the proof of work being any sort of economic impediment. It takes Claude a few minutes to write a near instant solver, that’s the maximum proof of work cost this adds.

drum55on It took a year to ship WebAssembly in Anubis

It’s not seconds though, it’s in the order of tens of milliseconds of work, it’s faster to complete the challenge than most http requests will take.

last difficulty 5 nonce 645376 in 9 ms (101.6MH/s, avx512-x16)

This takes many minutes to solve on a cellphone, 9 milliseconds on a cheap VPS. It’s not stopping anyone.

drum55on Georgi Gerganov on llama.cpp/ggml future after Nvidia acquisition of HuggingFace

They got a C&D.

drum55on Memory prices climb 500% in 12 months

Most devices never test their memory and don’t have ecc, so you generally never know. Random instability can be fixed by restarting and your program not being in an area with bad bits anymore.

drum55on A year of fighting scrapers on my 1.5 million-page website

https://blog.mozilla.org/en/firefox/block-cryptominers-with-...

    To help you combat these unwelcome scripts, we’ve partnered with Disconnect and compiled lists of domains that serve cryptomining scripts so you can block them directly through Firefox. When you download Firefox,  tracking protection is enabled by default, blocking known cryptominers.

drum55on A year of fighting scrapers on my 1.5 million-page website

People tried this in 2013 or so, there's no point to it. Doing proof of work in javascript in a browser is so crushingly, pointlessly slow that there's no value at all. Some browsers also intentionally detect attempts to do proof of work and attempt to block it entirely.

drum55on 99% of My Website Traffic Is Bots

The prompt used for Opus 4.8 was:

    write a implementation of the anubis proof of work in native c code, optimized for speed above all else. use every trick available to make the proof of work as efficient and fast as possible, including modern processor tricks on the x86 platform. your code should avoid using external libraries where possible, include tests, and be readable and concise. a reference for what needs to be met is in this repository. https://github.com/TecharoHQ/anubis

Then

    let’s develop this more. turn this solver into a local HTTP server that can be given work in the request, and it returns solved work. make an end to end tester that sends test work to the solver and waits for a valid response. add support for solving with a GPU using cuda.

Then it was done more or less, it happily made a local server that supports solving the challenges given to it in bulk with priority based queue and can tolerate potentially tens of thousands of requests a second with no issue. The CPU time spent solving the challenges is less than the SSL setup for the connections. The GPU version does in excess of 20GH/s (but with high latency) though I didn't really test it, I'm not using this for anything but proving a point that the LLM itself can write the bypass tools and run them happily.

drum55on A year of fighting scrapers on my 1.5 million-page website

It's more or less designed for it, it's SHA256 with a break in the middle for midstate compression to be effective, and the difficulty system is based on a misunderstanding of how bitcoin PoW works ("number of zeros" is never, ever a consideration in bitcoin, it's a match to a floating point target).

sha256(challenge + ascii(nonce)) means that the first compression round of the function can be cached and the second compression round is just the nonce plus the cache. This is the same trick used in Bitcoin mining and would have been avoidable by putting the nonce first, so immediately any non-naive code has to do half the proof of work as the vanilla solver.

drum55on A year of fighting scrapers on my 1.5 million-page website

Which is trivially bypassed by an actual implementation of the proof of work in non-javascript, rendering it absolutely useless. The website is approximately 3800x times slower than native code, and hundreds of thousands of times slower than the CUDA kernel claude wrote. The "proof of work" is just non existent at that point, they're solved in milliseconds for what would take the browser version 10 minutes or more, it's security by obscurity being dressed up as something more.

  pow_server  http://127.0.0.1:8080   backend avx512-x16
  ──────────────────────────────────────────────────────────
  uptime   00:03:12
  solver   ● BUSY  difficulty 9, 0.3s
  queue    [####################............] 5/8   peak 12
  ──────────────────────────────────────────────────────────
  accepted 1240        solved 1180
  503 shed 48      504 timeout 2      4xx/5xx 10
  ──────────────────────────────────────────────────────────
  last     difficulty 5  nonce 645376  in 9 ms  (101.6MH/s, avx512-x16)
  hashes   3.90GH total   avg 65.3MH/s
  Ctrl-C to stop

Claude even made a nice little API server for it after implementing midstate compression, AVX multi way hashing, and a CUDA kernel. This doesn't stop the literal LLM it's trying to block from solving the challenges, it's really annoying that everybody is using it and claiming that it's something that's usable in the real world as a result of it using proof of work. It's obscure, and obscure is fine so long as nobody is pretending that it is secure.

drum55on Who does Anubis actually stop?

The barrier is a two sentence prompt to Claude to add a patch to curl that does 500MH/s to the 50kh/s the javascript version does on the same hardware, it only works because it’s obscure and largely irrelevant.

drum55on Who does Anubis actually stop?

Then the only people who can use it are the ones with a GPU implementation.

drum55on Who does Anubis actually stop?

Claude Opus can make an optimized version of the proof of work solver that’s more than 10000x faster than the javascript one, in about 5 minutes time. Who is this stopping exactly? It’s the wrong tool in the wrong place.

drum55on Apple to increase spend with Broadcom to produce billions more U.S. chips

Very little uses crystal oscillators, they’re gigantic compared with electronics today and have very wonky performance over temperature and shock.

drum55on OpenPrinter

That's why this is just using off the shelf cartridges with commercial heads.

drum55on Midjourney Medical

1 petabyte per 60 second scans implies a kind of comical data rate to storage, even at RAM speeds that’s implausible. Imagine we need to write these to hard drives, they happily sustain 150Mb/s on the high end, which would imply you’d need 115,000 hard drives to absorb that amount of writes. Even with top end NVMe drives you’d need a thousand of them writing simultaneously.

drum55on Making glass-to-metal seals for homemade vacuum tubes

xray, photomultiplier and laser tubes are still SOTA. PM tubes in particular have a huge number of glass feed through for the intermediate plate voltages.

drum55on Ultra-processed foods in the global food system: The role of tobacco companies

The idea that you could buy any food that doesn’t fit that definition is silly, all foods have additives that’s why you can buy them and they last for more than 60 seconds on a shelf, all foods are processed because we don’t eat raw seeds as the majority of our staple diet. You have to come up with a definition of what “process” is good and bad, and what about them is “bad” before making statements like that.

drum55on Ultra-processed foods in the global food system: The role of tobacco companies

It’s basically a wildcard, “ultra processed food” is a classification of nothing and everything. There’s nothing inherently bad about processing food, lots of food is terrible for you but that’s unrelated entirely.

drum55on Chuwi Minibook X

There's actually a known prototype MacBook Pro from 2006 with a cellphone radio, and the release MacBook Pros from the time all have a weird looking area near the battery and RAM where the SIM slot was supposed to be, and some leftover parts for the goofy little extendable antenna on the screen. Hopefully they end up doing it.

https://www.macrumors.com/2011/08/14/photos-of-a-prototype-m...

drum55on Chuwi Minibook X

I had a thinkpad at one point that had a slot, but because it wasn't optioned for it you had to patch the BIOS or it wouldn't boot with anything in the slot, it seemed so hostile as to be worthless.

drum55on Chuwi Minibook X

I somehow managed to get it working in 2016 with a lot of hackery, I'd still have it as a usable device if the weird little pouch cells it had didn't die, repacking those batteries seemed like enough of a fire hazard I just didn't bother.

drum55on Chuwi Minibook X

The cause is just that the panel is mounted rotated on the device. It's supposed to be used in a tablet where the top is the short end and the side is the long end, opposite to a laptop.

drum55on Chuwi Minibook X

The GDP devices are amazing except for the keyboard, which is some fever dream layout I've never been able to understand. https://img.website.xin/contents/sitefiles3601/18006016/imag...

drum55on Chuwi Minibook X

I miss my Sony Vaio P series which fitted in a similar sort of niche, the cellphone radio made it just by far the best laptop I've ever used. Modern laptops don't seem to have provision for a LTE/5G radio which always confuses me a bit, in this form factor it would be ideal. I'm surprised nobody has cloned this actually, with phone screens being the right aspect ratio it seems obvious.

https://www.zdnet.com/a/img/2014/10/03/9f923860-4b47-11e4-b6...