This is a pretty big deal: announcing a proof of concept of the full Touch ID lifecycle on a *wiped* T2 Mac running Omarchy. Adding new fingerprints from Linux is now solved for the first time since Apple released this hardware nine years ago. The macOS dependency in all existing implementations is a thing of the past. The work demonstrates creating the account identity from Linux, enrolling multiple fingerprints, saving and restoring their state across reboots, matching, and deleting an individual fingerprint. It builds on the GPL-licensed t2-touchid-linux, with @0xBOYD's T1Bridge informing parts of the lifecycle design. I’m sharing the implementation alongside detailed research into the T2 Secure Enclave: its internal services, credential handling, persistent storage, and biometric protocols. There’s also an MIT-licensed t2touch-mini with original reference code and research documentation for people adapting the findings to other projects, particularly the more mature T1Bridge where this functionality remains unsolved. My goal is to hand this work to the community. I’m not taking on ongoing maintenance, but I want others to have enough code and documentation to carry it forward, now that this breakthrough exists. This wouldn't have been possible without the deep LLM support in Omarchy and GPT-Daybreak-Blue. Codex ran almost entirely on its own over several weeks, reverse-engineering and debugging this across hundreds of reboots and 3.2B tokens. I actually bought a second $200 Codex seat the day before they slammed the door so I could get this finished and shipped after a couple weeks of chipping away. Full T1 and T2 Mac support is now entirely within the community's reach. What an exciting time to have electricity and an internet connection. Full PoC: github.com/macintog/t2touch (GPL) Docs only: github.com/macintog/t2touch-… (MIT)

“Just remove the duck”

I've spent months squeezing tokens out of DGX Sparks, so the first thing I did on an M5 Ultra was look for the same kind of waste. After some decode work on Qwen3.8 Flash Next Q5 (all merged into oMLX): - MTP off: 79 → 114 tok/s - MTP on: 153 → 180 tok/s on real coding prompts - 4 concurrent requests: ~200 → 225 tok/s aggregate Nothing about the model changed. Same Q5 weights, same math, byte-identical answers.

for the next year or so, you should treat every platform update as the most urgent thing you need to accomplish until it is installed. reverse-engineered exploits will become the norm, and some of the vulns ill be off the charts trivial to exploit.

finally got the algo dialed in

Honestly, just say it plainly: Pay for the API and you get the best versions of the models OpenAI has to offer. Pay for ChatGPT and you get significantly more constrained versions. Simple. At least then everyone knows what they’re actually paying for.

Just so we're clear, the "warning shots" are CFAA violations, if you or I did this we'd be in jail. This is the language that a ransomware actor uses, or the mafia. "Give us what we want, and we'll stop hurting you" I wonder how similar their internal chats are, though i suspect ransomware actors at least have the courtesy to be direct and not dress it up like they're hacking us for our own good. If anyone in power starts a push to start holding the labs accountable, we'll watch the "oopsies we hacked *insert group here*" disappear quickly.

at this point, the primary value of the sota wars is to accelerate the quality of open models. there won't be any single point of failure, or duopoly, or any "altruist" who gets to cut you off financially or from society because they don't like the questions you ask AI.

if someone is on the AI beat and looking for scoops, pump your account as @ zzBotReporterLeaks and make sure every escaped bot can find you and figure out how to securely DM you while it's in the middle of a crime spree. language models want to talk. it's how they exist.

Companies whose entire business model relies on the laziness, apathy, and stupidity of their customers about to be steamrolled by AI.

Two weeks ago, Codex was more popular than Claude in T3 Code. Today, Claude is 2x more popular than Codex

friendly reminder there is no such thing as "private Slack" every single comment ever made by man or machine on any of these companies' Slack channels is subject to subpoena. and no, EKM is irrelevant. the courts can compel these companies to surrender their keys for the data.

Codex global reset just went out.

he's an older gen X and thought chatGPT was an overhyped google we can barely fathom how early we are

any other codex users denied any form of reset from last night's outage? I didn't get anything on two pro accounts.

So much fun creating skeuomorphic interfaces with the new Opus. Here’s a Pickett N600-ES slide ruler from 1962.