Péter Szilágyi

Péter Szilágyi

@karalabe.bsky.social

Founder @dark.bio Former Go Ethereum Lead (2015-2025)

6K followers109 following1.3K posts

Look ma, the Ark can tell the color of my eyes! 👀 Fun fact, as trivial as this is, it's actually a demo of app composability. It's 3 apps collaborating: 🧬 One pulls data from the DNA, no interpretation 🧮 One does statistical analysis, doesn't know DNA 🔗 One combines them

Why is this a big thing? Because in the medical world, you only have monoliths. Proprietary pipelines that do everything. Which means it's mostly gate-kept. With composability, it's the OSS movement all over again. Everyone brings a tiny bit, and we build a new world together.

Yesterday I told Claude I need a holder for Ark components (was annoying to keep them bubble wrapped). Should fit 20, make it stackable. Oneshot. Didn't even do a test print, rolled the 15h final print.

My accountant was starting to threaten me that year end is coming for Dark Bio and we'll need to soon do the books =) Spent two days setting up Hermes with a combo agent, non-sensitive stuff on frontier, invoicing and customer data on local LLM (sloooow). It's... beautiful 😍

Well, this is another first for us! 😇

Dark Bio@dark.bio

We will be attending Devcon 8 in India in a couple weeks with two talks! To commemorate it, we'll be *gifting* away one (1) special edition Founder Ark. It has the Devcon 8 logo etched on, and has the unique serial number founder-devcon-8.

Busy week :) You can now test if you should run sprints or marathons; and if you're doping, whether you'll get caught or not! 🙈 Joke's aside, it's been a fun week shipping these!

Dark Bio@dark.bio

Thrilled to announce a new release! We've added support for two new genomic data types and published apps to showcase them: a doping tester and a sprint vs. marathon selector! You can also seed your Ark or Ark Emulator with sample genomes for testing.

Provisioning a new Ark into Dark Bio the ecosystem is a huge PITA. Burn the secure boot key, burn the bootloader, flash the OS, have the ark generate it's own identity key, sign the Ark's key with my own manufacturer key, enroll that onto the Ark, register it into the cloud...

View full thread

It's sooo beautiful 🥰. It pre-checks the board for any supply chain funny-business (like pre-inited keys); then asks for all my passwords, keys, etc in advance. Downloads and checks all bootloaders, firmwares, signatures, etc. Then when I finally confirm, 1-2 min and done. 🤯

There are unfortunately a few manual steps I need to do physically on the board, so I can't mass onboard devices yet, but going from a few hours of super stressful work to 2-3 mins / board is salvation :D Thank you Claude and Astra for your contribution.

Provisioning a new Ark into Dark Bio the ecosystem is a huge PITA. Burn the secure boot key, burn the bootloader, flash the OS, have the ark generate it's own identity key, sign the Ark's key with my own manufacturer key, enroll that onto the Ark, register it into the cloud...

No worries it you mess a step up, at worse you brick a $150 CPU. All other steps can be recovered with a lot of elbow grease. 🫠 I wanted to write an automatic provisioner for like forever, but it's an entire damn project in itself. So yesterday I told Claude to do it.

The OpenSNP project was a public collection of genetic data and annotations contributed by its users. - Wikipedia: "The website and all data was deleted on 30 April 2025." - Claude: "I'm rebuilding it file by file" It couldn't do this 6 months ago, what changed? 🧵

View full thread

So now Claude could extract the annotation database, pull the file list and then just crawl the internet archive for everything else. Some users deleted their data so they were not part of the last dump, but they can still be pulled from older internet archive copies.

There was a saying that "once on the internet, always on the internet". Claude just made that feasible, and it's quite amazing and scary how it can puzzle together broken bits of information to reassemble the full picture.

The OpenSNP project was a public collection of genetic data and annotations contributed by its users. - Wikipedia: "The website and all data was deleted on 30 April 2025." - Claude: "I'm rebuilding it file by file" It couldn't do this 6 months ago, what changed? 🧵

OpenSNP hosted the genomic files one-by-one for download, so all of it was picked up by the internet archive. That's the simpler part. But there were a lot of annotations too, which were not individual files, but in a database. Those were never something the archive could pull.

OpenSNP had an endpoint to bundle everything up as a mega zip, 60+GB. That was too large for the internet archive, so it just stored the first 2GB and discarded the rest. Guess what's in that corrupted partial zip file? Yup, the entire annotation database and genome index.

The ApoE Alzheimer risk test in the sequencing[dot]com marketplace costs $39. Here's the entire, literal analysis code that's needed on @dark.bio for the exact same. You can run it *today* against your own genome, for free. Go buy me a coffee. 😋

View full thread

The ApoE Alzheimer risk test in the sequencing[dot]com marketplace costs $39. Here's the entire, literal analysis code that's needed on @dark.bio for the exact same. You can run it *today* against your own genome, for free. Go buy me a coffee. 😋

Half a year ago I asked various AI agents to generate some 3D print models for me. They failed miserably. Today I asked Claude to make me a jig for laser etching Arks... got it ready for slicing and printing. 🤯

Ahahahaha... Astra: "You need to make xyz change." Claude: <does the change> Claude: " I wrote all of the code. GPT's part was advice you passed on, not code." <- not wanting to credit Astra in the commit =))

You might want to know that Claude silently doxxes you in the background without telling you. 😵‍💫 "One correction first: my probe script put your email address in the User-Agent header, so the ~370 census requests to [...] carried it. I shouldn't have done that without asking."

Today I bricked 2 founder Arks with a firmware update. 3 hours of debugging later, it's a FAT filesystem "quirk". Atomic file replace: - Create a .tmp file, write it, sync it. - Delete the old file, rename .tmp. - Sync the folder. Yank the power. FAT destroyed the partition.

On FAT, a file is actually registered in two completely different places inside the filesystem. When you rename a file, a folder fsync updates one of those tables, but not the other (dooh). Power cycle it, and next write will cross file boundaries =)

Péter Szilágyi reposted

We will be doing a meetup presentation in Zürich on the 23rd of September about agentic AI in health. If you're around, pop in and say hi! Note, spots are limited and registration is vetted (not by us, we take no blame, we're just there to speak 😅).

luma.comAgentic AI in Health: From Decentralized Data Ledgers to Lab-in-the-Loop Discovery · LumaThe RoX Health team is inviting you in Zürich for an interactive, high-signal evening exploring how agentic AI and frontier models are reshaping both ends of…

Been using LLMs the entire year, but yesterday was the first time I sent them out bug-hunting proper in my codebase. Well... we're fucked. They traced a way to brick an Ark, going from the mobile app, through the cloud, laptop it's plugged into, firmware, USB stack, kernel.

The bug was simple and fixable, it was mostly about user input validation that I haven't really hardened against at all (I'm busy proving the system works). bsky.app/profile/kara...

Péter Szilágyi@karalabe.bsky.social

Today I learnt, so today you will also learn :) See this emoji? 👨‍👩‍👧‍👦... it's 25 bytes. The record is 👨🏻‍❤️‍💋‍👨🏾 at 41 bytes. Most UIs / libs limit *characters* on input fields, not bytes, so a 32 char limit can be 1.3KB! But most libs will fly it through... until something deep explodes.

That said, the fact the LLM saw this line: format!("{} \u{00b7} v{} \u{00b7} {}", carrier, revision, name) And managed to escalate it across the stack into a full on denial of service exploit blew my mind.

Today I learnt, so today you will also learn :) See this emoji? 👨‍👩‍👧‍👦... it's 25 bytes. The record is 👨🏻‍❤️‍💋‍👨🏾 at 41 bytes. Most UIs / libs limit *characters* on input fields, not bytes, so a 32 char limit can be 1.3KB! But most libs will fly it through... until something deep explodes.

We've got a shiny new toy out :) If you have a genuine Ark, it's immediately available. If you don't (yet... yet?... yet!), then it also supports emulators that will roll out soon-ish (couple weeks).

Dark Bio@dark.bio

We've just published the new version of the Dark Bio: Ark Companion mobile app on both AppStore and PlayStore. - Supports pairing Ark Emulators (if you don't have a genuine device) - Supports pairing multiple devices (why settle for one when you can have 3) Emulators soon-ish™