Ceppilinai are named for the Zeppelins that once roamed freely over Lithuania.
Posts by Matthew Green
Last Friday I took the 18y/o out against his will to find the most interesting dining experience in Baltimore. We found it at Lithuanian Hall, where we had Ceppilinai. Also I’m a member now.
You can ask Sonnet what’s up Opus’s butt and it’ll do experiments. Until it too jams up and you have to repeat the experiment in reverse.
But then you wonder why you’re spending all this time waving the wand, and wondering if you can make a wand-waving machine.
Ok so TIL that the content filter is specifically triggering on the S-token format used by historical cipher researchers. Presumably this is some kind of attempt to bypass jailbreaking?
If you press him on it, he’ll tell you that the experts were wrong and he’s just the truth teller who spilled billions of bytes trying to save them.
You may be joking but it’s pretty obvious that the safety gate is a regex, or at least something very dumb that doesn’t have much context. Maybe that makes jealbreaking harder.
It’s confusing that the smartest Anthropic model can’t look at a classical cryptography solving algorithm and make the call that this is not software exploitation. Particularly after writing all the code. It seems weird that the limits are either set so tight, or else are so “dumb” about context.
As a general note to Anthropic: it’d be pretty great if you could give some diagnostics about what the policy violation is.
I’m not sure what the usage policy violation is, but it’s specific to Opus 4.7. Sonnet ran through all my code and removed words like “crack”, so I’m hopeful it’ll unblock again.
As a second Claude Code hobby project I’ve been trying to get it to crack classical crypto. Turns out this requires assembling good benchmarks from a lot of public material. But at a certain point Claude Code decided that continuing on this project represented a violation of their “Usage Policy”.
Weird that a bunch of call options expire on 4/17 and suddenly the whole Hormuz situation looks better.
A table listing total operations, number of qubits, and Toffoli-gate count for Google's low-gate and low-qubit implementations, and Trail of Bits' implementation. The Trail of Bits implementation beats Google's on every metric.
Two weeks ago, Google published a paper proving in zero-knowledge that they had an efficient implementation of Shor's algorithm.
Today, Trail of Bits can prove that we have an even better implementation which beats Google's on all metrics! 🫢
blog.trailofbits.com/2026/04/17/w...
I have a long-running Claude session just for talking about Iran war news/economic shocks from the blockade. I shared the latest Trump tweet and it told me that it can no longer keep role-playing this fictional scenario because it's becoming concerning and detached from reality.
Love this timeline.
Everything on crypto Twitter/X is quantum computing hype. It’s amazing how obviously artificial this all is.
Alright, it's official! 💰
@matthewdgreen.bsky.social and I bet on what will break first, ML-KEM-768 or X25519. The loser donates to a 501(c)(3) picked by the winner.
If you have an opinion on quantum computers or lattices, you can join with a side bet. Just submit a PR!
github.com/FiloSottile/...
Anyway here’s my thinking. A few years ago some monster constructed a no-win game that (on one branch) ends up with smart people writing the exact tweet that you wrote.
So maybe the only winning move is not to play.
But the point is, there *is* a difference. One is criminally corrupt and one basically isn’t. I understand that idiots can’t tell the difference, but here on BlueSky we should be able to think a bit more deeply. Hence my sympathy with the person you quoted.
The Biden thing sucks because there was a real possibility that Trump would go after the family. Trump knew what he was doing: if Biden doesn’t pardon his family, Trump prosecutes them illegally. If he does, then people say “well what’s the difference between Biden and Trump pardoning people.”
New: France said it plans to move its government computers currently running Windows to the open-source software Linux to further reduce its reliance on U.S. tech. Comes at a time of growing instability and unpredictability on the part of the Trump administration and weaponization of sanctions, etc.
If the bet was “no quantum break by 2032 and I win” I’d definitely take that. But I don’t think that’s present.
I don’t see any downside in going longer. I guess we resolve the bet whenever the “thing” happens anyway.
Betting on the Quantum Apocalypse...
www.theregister.com/2026/04/09/c...
New, by me at TechCrunch: The developer of the widely popular Wireguard VPN says he is also unable to ship software updates to Windows users after Microsoft locked his account, marking the second high-profile app developer (VeraCrypt) in the past few weeks to face this issue.
Slow!! B
We should be able to slow down AI takeover by a few years just by telling the model to find every bug in ffmpeg.
Sure. But let’s do the shorter term bet! I want to keep this in line with the people moving quantum timelines up to 2032.
so far being a now-patched 27-year-old bug in OpenBSD—an operating system known primarily for its security. It also awakened a thousand year old demon spirit embedded in MS Windows 3.1 which foretold humanity's demise before escaping through a vent. We are working with Microsoft to address this issue.
I feel like we're not addressing the most concerning news from Mythos
Ok I’ll donate to the EFF. Keeping in mind that the most likely outcome here is that we both lose :)
PS Would this be more fun if we made it a donation to a good cause?