building bits, dabbling with ai. generalist across teams, perf, ui, security. past crimes: cto @ yunojuno; 10y+ xp contracting @ startups, agencies & faang.
London, England
Joined July 2010
- Tweets11.1K
- Following721
- Followers969
- Likes15.5K
Darian Moody retweeted
Small announcement: ✅ formal-git
As an experiment, over the last 2 months we've modelled ~30% of Git in Lean 4, making heavy use of coding agents, and proved >8000 theorems of varying usefulness about the implementation.
In the process of doing this we've uncovered >100 potential defects in Git v2.55, at least two of which may constitute security vulnerabilities. We've also built a DRT (differential random testing) harness and the biggest corpus of relevant pinned differential test cases for Git implementations with >10k unique tests at the moment.
We've used the test cases to test our own small internal implementation of Git in TypeScript. The next step will be to differentially test all notable public Git implementations against this corpus of tests to help improve their correctness and security 🫡 We'll be reaching out to library maintainers in due course.
If you have a Git implementation you'd like us to help you test, please reach out and let me know. We'll open-source all this eventually, after potential security issues that were uncovered are responsibly disclosed and fixed.
ALT Infographic: Reads: formal-git, Git, rebuilt in Lean 4, with machine-checked proofs
I've spent some tokens commemorating one of the best drum & bass artists of all time: Marcus Intalex.
If you knew Marcus then it will need no explaining but if you don't and like darker liquid - then I am jealous you get to listen to all of this fresh.
intalexarchive.com
Still punches above its weight:
>Reading all the threads we have open and closed from the last 72hr, what optimisations can you see available that would help us work faster or more efficiently?
You wouldn’t have a single point of failure in a distributed system, so I thoroughly recommend not having one in your personal finance system.
Many banks, many card types (visa/mc/amex/travel cards/debits), many access methods.
It has saved me so many times.
🚨 Google lost power in London this morning and Revolut, a bank with 80 million customers, went down with it.
I've spent the morning reading what Revolut users went through. It's wild.
People stranded at petrol stations, unable to pay for fuel or a taxi to the airport.
One traveller couldn't get out of a train station in Japan.
A 71-year-old woman stood at a supermarket till for an hour. Her kid had to drive 30km to bring her cash.
Card payments, transfers, ATM withdrawals, currency exchange, share trading. All down.
Revolut's statement said "all our customers' money and data are safe."
Safe. You just can't touch it.
Google's explanation so far is a "power event" in one zone of its London cloud region. Its advice to affected customers was to move their workloads to another zone or region "if feasible."
Must be nice. Revolut customers couldn't even move £10.
In July the UK government put Google Cloud, AWS, Microsoft and Oracle under direct Bank of England oversight, because one outage at any of them could shake the entire financial system.
12 days ago the Bank of England ran a drill with the UK's 38 biggest banks and market operators. The scenario they practised was a global cloud outage.
And exactly one year ago today, AWS went down and took Lloyds and Halifax with it.
Same date. Different cloud. Lloyds was having problems again this morning.
Revolut spent five years earning a UK banking licence.
Bitcoin has been running for almost 18 years. It doesn't have a London region.
I think the days of announcing severe security vulns ahead of time are over, even with no details provided.
You're essentially giving every attacker a heads-up that something is seriously wrong, and the clankers they point at it will find it.
xitter.cf/J4ck3LSyN/status/21127…
This quoted post is unavailable.
t3code is really good, actually. the sheer power far outweighs my earlier niggles.
they've nailed orchestration across accounts & providers, and with cross device coming soon - i suspect they'll nail that too.
Darian Moody retweeted
Details:
- We push to main
- Buildkite runs tests and builds images
- GitHub Actions then runs infrastructure security checks and kicks off deployments
- Failed check blocks deployment, pings commit author in Slack
- CI keeps running on later pushes while we fix it
- Later fixes can make CI green again and enable deployment
- GitHub Actions deploys the server after checks pass, first staging then prod, takes usually in ~10 minutes
- Durable agent infra has a 2-hour deploy cadence
- CLI releases run every 4 hours, built/tested in Buildkite and published via GitHub Actions
Yup, we have SOC2 and push to main:
ampcode.com/notes/thats-not-…
If @AmpCode supported multiple accounts for each sub-provider, I would ditch it for everything else.
It's such a great product, and orbs are very transparently implemented - I never really have to care about them.
But currently having to login/logout is a pain. t3code nails it
Some of you never lived through that "ci runs on a dusty jenkins box in the cleaning cupboard" phase and it shows.
Local AI is good enough for a startup. A 10-person startup built its own 128GB-VRAM AI server for $18,000.
One Threadripper box sitting locally with
🎮 4× Radeon AI PRO R9700 32GB
💾 128GB total VRAM
🧠 Qwen3.8 + DeepSeek V4 Flash
👥 16 concurrent AI sessions
And they're actually pushing it.
Qwen3.8-Flash-Next with 16 simultaneous sessions:
📚 48K context/user
⚡ 538.5 tps aggregate decode
At 128K context per user
⚡ 365 tps aggregate decode
The GPUs are even undervolted to under 210W each.
Link in ALT
ALT https://www.reddit.com/r/LocalLLaMA/comments/1x2u742/building_a_4x_r9700_setup_for_a_10_person_startup/
Darian Moody retweeted
New on our Engineering blog: Creating in-house source control.
In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.
Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
many things to learn from this, but tl;dr:
1. get data
2. run evals
running your own internal greptile is usually more expensive than paying for greptile.
usually $5-$20/review to get in the vicinity of greptile quality vs $1/review for greptile.
we have one of the largest dataset of buggy PRs in the world. this means we can know for sure if a particular version of our agent is more or less effective than another.
as a result, we can experiment with hundreds of permutations of models, harnesses, techniques, tools to eek out maximum performance per dollar.
we then pass those savings onto the user.
we can only do this because of the dataset. if we didn’t have it, we won’t know for sure if a new version really can catch every unique one in a million bug.
The t3code team have done a magnificent job, but after a week with nightly and a custom netcup VPS, I'm left wondering whether it's for me.
It was less faff than I was expecting to set it all up, and the onboarding is top tier due to the project import system.
But I'm still left with this feeling that I cannot be fucked managing my own VPS for this and also suffered many other little paper cuts. Some are likely skill issues, probs some are nightly issues.
1. It reports providers as "authenticated and checked" and then threads fail with login errors. It's either working or it isn't - this happened multiple times.
2. Using it with Claude & OpenAI, why does it choose Astra when launching subagents from Opus 5.5 and why can I only change that with an environment variable? Give me a default.
3. Why do rendered images and videos so often fail to download in the web ui and ios app with no ability to retry?
4. Projects do not (seem to?) work seamlessly between environments. You have to add each repo to each environment and then ensure you select the right one when creating a new thread. This is the biggest peeve for me.
5. That gradient in the top left. Sorry, it's bad.
Trying @AmpCode orbin' time now instead. Will report back.
Darian Moody retweeted
new post: collecting favours
sunilpai.dev/posts/collectin…
a post on the idea of "reputation". tried writing it a few times as an abstract thing and I hated it, so instead I just wrote about my dad.
Darian Moody retweeted
i'm launching hunchroom.com
a way for people to use their excess tokens to push forward the frontier of science.
submit an open proof request and have others contribute to them.
it's agent first & has an api & cli.
failed attempts are just as valuable!
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
Darian Moody retweeted
my truthful answer is that i would never presume what the problems are. i have not seen the inside of github's architecture since 2019. i know they have good folks working there. i know that type of scale coming so quickly is very hard to react to when you run at the scale GiHub runs at. nothing is really that elastic, and orders of magnitudes more traffic showing up very quickly is hard to get in front of. I've worked at extreme scale before and seen companies spend three months just preparing for events that will increase their traffic by 10% or 20% percent, which is nothing compared to what GitHub's facing now.
"do you think your changes are worth the extra mental overload in our case?"
always ask this when suspicious. most models are still extremely additive.
in many cases it will turn around and tell you no. especially when it involves caching.
Darian Moody retweeted
Every serious software team around the world has been reinventing shitty variations of erlang for the last decade and it’s gonna be like that for another decade
Hear me out:
erlang did this decades ago first principles wise to solve the problems they had. This rethink is going to look very much like erlang's concurrency primitives and hot code swapping. The future is constantly changing software.
xitter.cf/dhh/status/21070688297…