blintz

Are frontier models actually astronomically expensive to train? GLM 5.2 was trained on ~30T tokens, so ~10^25 FLOPs. Say you get B300's for $5/h (pretty high), and you get 50% MFU; that's ~$15M. Of course there's also a bunch of risk that the training itself goes badly, post-training, etc. But still, compared to the inference spend after, it's not that crazy.

Reminds me of 0 A.D. (https://en.wikipedia.org/wiki/0_A.D._(video_game)). The usual problem with an open-source game is the art/music, but this looks pretty decent.

This is cool. Something about dropping everything to go see a movie in an empty theater is sort of tempting.

I really want automated QA to work better! It's a great thing to work on.

Some feedback:

- I definitely don't want three long new messages on every PR. Max 1, ideally none? Codex does a great job just using emoji.

- The replay is cool. I don't make a website, so maybe I'm not the target market, but I'd like QA for our backend.

- Honestly, I'd rather just run a massive QA run every day, and then have any failures bisected, rather than per-PR.

- I am worried that there's not a lot of value beyond the intelligence of the foundation models here.

They are playing a bit fast and loose with the word "banned".

Your smartphone contains materials processed through semiconductor fabrication, chemical etching, metal anodizing, glass tempering, and electroplating — none of which you could start a new facility for in California without years of litigation.

I agree that we should make it easier to do things, specifically by decreasing the amount of litigation involved in doing stuff. But the risk of a bunch of litigation isn't a ban, right? I get that it's trying to be attention-grabbing, but calling it a ban when it's not just sort of confuses the issue.

I have tried many times to do this, but lack even the minor discipline required. I inevitably make changes to the commands I want to run at the command line, rather than in the script, and then later forget to edit them in the script.

Instead, I now swear by atuin.sh, which just remembers every command I've typed. It's sort of bad, since I never actually get nice scripts, just really long commands, but it gets you 50% of the way there with 0 effort. When leaving my last job, I even donated my (very long) atuin history to my successor, which I suspect was more useful than any document I wrote.

My only hot tip: atuin overrides the up-arrow by default, which is really annoying, so do `atuin init zsh --disable-up-arrow` to make it only run on Ctrl-R.

I see, yeah, I can see how if it's like 100% matching some parts of the style, but then failing completely on other parts, it's a huge pain to deal with. I wonder if a bigger model could loop here - like, have GPT 5.2 compare the fine-tune output and the Nano Banana output, notice that trees + water are bad, select more examples to fine-tune on, and the retry. Perhaps noticing that the trees and water are missing or bad is a more human judgement, though.

I was most surprised by the fact that it only took 40 examples for a Qwen finetune to match the style and quality of (interactively tuned) Nano Banana. Certainly the end result does not look like the stock output of open-source image generation models.

I wonder if for almost any bulk inference / generation task, it will generally be dramatically cheaper to (use fancy expensive model to generate examples, perhaps interactively with refinements) -> (fine tune smaller open-source model) -> (run bulk task).

I reset my network settings and updated before realizing this is just an outage. I kept searching “<MY LOCATION> Verizon outage” but did not even consider it could be nationwide. I guess it shows how rare nationwide outages are.

Google (and Vercel) are great for doing this! I would like to see Anthropic and OpenAI do something similar, since they too greatly benefit from Tailwind CSS.

PEP 658 went live on PyPI in May 2023. uv launched in February 2024. The timing isn’t coincidental. uv could be fast because the ecosystem finally had the infrastructure to support it. A tool like uv couldn’t have shipped in 2020. The standards weren’t there yet.

How/why did the package maintainers start using all these improvements? Some of them sound like a bunch of work, and getting a package ecosystem to move is hard. Was there motivation to speed up installs across the ecosystem? If setup.py was working okay for folks, what incentivized them to start using pyproject.toml?

I tend to avoid sci-fi that hits too close to home (don't love any of the AI/internet/crypto classics, same reason I can't bear to watch Silicon Valley), so I was a little bored by the top of the the list.

But, there's really good stuff that I've loved just a bit down the list: Foundation, The Left Hand Of Darkness, The Dispossessed, Stories of Your Life and Others, Exhalation, Children Of Time, Dune.

Was surprised the Mars trilogy was pretty low (might be the keyword indexing?) - highly recommend, as long as you don't get too bored by descriptions of rock.

Doesn’t rustc emit LLVM IR? Are there a lot of systems that LLVM doesn’t support?

There was lots of public scrutiny of Kyber (ML-KEM); DJB made his own submission to the NIST PQC standardization process. A purposely introduced backdoor in Kyber makes absolutely no sense; it was submitted by 11 respected cryptographers, and analyzed by hundreds of people over the course of standardization.

I disagree that ML-KEM is "obviously weaker". In some ways, lattice-based cryptography has stronger hardness foundations than RSA and EC (specifically, average -> worst case reductions).

ML-KEM and EC are definitely complementary, and I would probably only deploy hybrids in the near future, but I don't begrudge others who wish to do pure ML-KEM.

I think the whole point is that some people would be forced to use it due to other standards picking previously-standardized ciphers. He explains and cites examples of this in the past.

If an organization wants to force its clients or servers to use pure ML-KEM, they can already do this using any means they like. The standardization of a TLS ciphersuite is besides the point.

He comes with historical and procedural evidence of bad faith. Why is this ridiculous?

Yes, the NSA has nefariously influenced standards processes. That does not mean that in each and every standards process (especially the ones that don't go your way) you can accuse everyone who disagrees with you, on the merits, of having some ulterior motive or secret relationship with the NSA. That is exactly what he has done repeatedly, both on his blog and on the list.

why wouldn't you equate that to working for the NSA (or something equally bad)?

For the simple reason that you should not accuse another person of working for the NSA without real proof of that! The standard of proof for an accusation like that cannot be "you disagree with me".

Since ML-KEM is supported by the NSA, it should be assumed to have a NSA-known backdoor that they want to be used as much as possible

AES and RSA are also supported by the NSA, but that doesn’t mean they were backdoored.

TLS 1.3 did do that, but it also fixed the ciphersuite negotiation mechanism (and got formally verified). So downgrade attacks are a moot point now.

Standardizing a codepoint for a pure ML-KEM version of TLS is fine. TLS clients always get to choose what ciphersuites they support, and nothing forces you to use it.

He has essentially accused anyone who shares this view of secretly working for the NSA. This is ridiculous.

You can see him do this on the mailing list: https://mailarchive.ietf.org/arch/browse/tls/?q=djb

Love atuin - has saved my ass many more times than I can count. The more you guys can monetize the better; will help keep the base product good. Even pretty senior devs (who don’t always love changing their workflows) can find a lot of value in it.

I would pay you guys for E2EE syncing, but I think it’s free at the moment. Charge me!

It’s symmetric keys, so quantum doesn’t matter.

This page measures the concentration of the Fediverse and the Atmosphere according to the Herfindahl–Hirschman Index (HHI), an indicator from economics used to measure competition between firms in an industry. Mathematically, HHI is the sum of the squares of market shares of all servers.

I had not heard of this metric before - it’s neat and simple to understand. If you scaled it down to 0-100 (by dividing by 100), I think it would make the numbers more immediately understandable. I’d even consider inverting it (so 0 = centralized and 100 = decentralized), since the website title implies measuring progress ‘towards’ decentralization.

I'm surprised at how normal some of the unseen words are. I expected them to all be archaic or niche, but many are pretty reasonable: 'congregant', 'definer', 'stereoscope'.

This is a good one: https://eprint.iacr.org/2019/720.pdf

There has been a theoretical breakthrough that makes search a O(log n) problem, actually, (https://eprint.iacr.org/2022/1703) but it is pretty impractical (and not getting much faster).

Technically, there are rate-1 homomorphic encryption schemes, where ‘rate’ refers to the size ratio between the plaintext and the ciphertext. They’re not super practical, so your general point stands.

I say this as a lover of FHE and the wonderful cryptography around it:

While it’s true that FHE schemes continue to get faster, they don’t really have hope of being comparable to plaintext speeds as long as they rely on bootstrapping. For deep, fundamental reasons, bootstrapping isn’t likely to ever be less than ~1000x overhead.

When folks realized they couldn’t speed up bootstrapping much more, they started talking about hardware acceleration, but it’s a tough sell at time when every last drop of compute is going into LLMs. What $/token cost increase would folks pay for computation under FHE? Unless it’s >1000x, it’s really pretty grim.

For anything like private LLM inference, confidential computing approaches are really the only feasible option. I don’t like trusting hardware, but it’s the best we’ve got!

This isn’t some conspiracy, it’s just CYA. They know your general location from your IP and device APIs, they don’t encrypt business messaging, and they comply with subpoenas.

Excited to see someone finally doing this! I can imagine folks with sensitive model weights being especially interested.

Do you run into rate limits or other issues with TLS cert issuance? One problem we had when doing this before is that each spinup of the enclave must generate a fresh public key, so it needs a fresh, publicly trusted TLS cert. Do you have a workaround for that, or do you just have the enclaves run for long enough that it doesn’t matter?

Call it the iron law of LLMs:

"No worthy use of an LLM involves other human beings reading its output."

If you use a model to generate code, let it be code nobody has to read: one-off scripts, demos, etc. If you want an LLM to prove a theorem, have it generate some Coq and then verify the proof mechanically. If you ask a model to write you a poem, enjoy the poem, and then graciously erase it.

I know this would be less fun, but given that the key space was only 36^4, why not just run the actual decryption functionality in QText? Like, even if it takes 1 second to decrypt, spin up 32 cores and wait a day. They allude to the idea that checking the key derivation is faster, but I wonder by how much.

(of course, it’s still interesting to read about 90s encryption, so I appreciate that they did it the fun way)