Meetups/Infra/2026-01-12

busy busy night,

Introductions

Phillipe: (needs to go)
Alex: ML, no projects, trying to get back on feet from holiday
Doug: Likes computers 2026 is the year of self hosting (cont.) this year I'm committed to launching w/ claude an app every 2 to 3 weeks. See some buggy vibe coded, self hosted slop, he's got it (self deprecation :D)
Lulu: Come here to learn.
Alex: Gives out cool stickers, trying to help people use claude code better
Ash: works on infrastructure code for 10 years. works on opensource, came to SF 2018, 
Kevin/MrName: Today messing with ESP micro controllers (about 12 left). AI interests, GPT 5 ushers in a population of AI hackers.  
Peter: From Russia, travelled through turkey, costa rica, into dev ops. 
Jeanie: Sticks her head in the door, haven't had project for a while, lost job in 2023, when AI started, decided to train an AI assistant. (be part coworker, part therapist)
Robert: (newton os on discord) Operating systems, based on android, self hosting somethings, like NAT, currently self hosting AI. 
Zack:  cyber security engineer, working on home lab, needs to figure out.. Builds out pipelines for vendor products, but not so much on product side, source to delivery. dev-sec-op, jellyfin helps (what is jellyfin - https://jellyfin.org/)
Erik: Posts alot on discord working with graphics, interested in hardware, query chips, legit backdoors. 
Renaud : New to SF, couple month, just graduated CS. Hoping to build a JS engine for a subset of the language. Into self hosting, general CS, anything technical.
Derek: from New England, Programmer, Working on a game to learn mandarin chinese
Max: Recently graduated, programming things, going on a trip, be away for a month
Chris: aerospace engineer, satellite computers for fault tolerance. May be hiring,
Riley : Minneapolis maintainer of linux distro, remove 32bit libs from distro and virtualize them. Working 
Chris : quit job at AI firm. Learning spanish and language of the azteks(sp?)
Zacky: roped in friend co-acodeability, working with public domain mandarin course from the 1980;s
Daniel: Working on mesh drones, 
Dillen: Founded long range autonomous drones, looking for CTO, someone with computer vision experience.

Peter:

The 50 minute introduction

is android open source?

Swiftstack? 115 PB footprint -> 3 Exabyte, primarily used for machine learning training pipelines. Autonomous Vehicle Program for Mercedes generates 1.5 Exabytes / day. Kubernetes, but started with ansible in 2018. 250 people

Max: Building software? Need to store the software. As we update, many copies add redundnecy, only a little changes. How do we dedup?

Naively, store every version -> creates duplication, why store twice? Solution? Hash files, see if some are the same,

What happens when binaries are slightly different? Chunk. Hash chunks. technical difficult.

Content defined chunking, take a rolling window over the file, with a small hash of bytes that helps decide how to chunk. starting with 12 bytes. (A hash function is a one way function from domain to uniformly random output) If we contro how many bytes of hashing controls size of chunks with guarantees. (not going to hash entire file) every byte.

two approaches cheap hash functions, every byte is a coefficiant to a polynomial

gear hashing or precomputed random values, shift previous hash by 1 then add what new byte from the table, and will be randomized based on file position, better than source deduplication, offset in file? It will find it. Will get randomly defined chunks, but they are from the file, if at one split, all dups, then it will be detected.

Has investigated ZFS?

New ZFS has more efficient dedup. (Had to be memory) Now dedup is still not very useful because it's not block level. Need chunk level dedup to save more space.

think of it like "Hash function is a random number generator" based on the content, chunk the file and reset, may get some inefficiency, but not too much.

Uses hash collisions to detect chunks.

(WW2) When Polish were cracking codes, solving cipher, why does it have a specific fingerprint, was unique for specific position, radio transmisssion, find correct position, similar tecnique.

Chalk board experiment,

Not a product, running a caching server for nix package manager. Cachix is binary caching service. Replit has blog post about Cachix. (a rewrite of nix in Rust) uses content defined chunking, anyone running chaching service with similar properties, similar content by shifting. 90% in filesize.

nodejs will benefit. massive depend tree, use it to detect when libs have changed.

use content defined hashing to distribute that hashing.

Machine Learning algorithms are compression? Data is too small? Window size can adjust to? are smaller windows more efficient? large

Paper released to train on images?

With source code, unison. hashing the chunks, hash tree.

IPFS may also use hash trees, distributed file system nixos source based dedup.

All the binaries sit in a big cache, symlink, A B C A B C

Parkay files?

Content defined chunking. ZFS might consider adopting. BetterFS?

Guix was forked from Nix, the build daemon has not been converted. "Nix is beautiful, but where did it go wrong".

The android SDK is not totally open sourced, there are proprietary blobs? Which?

Aluminum OS, a desktop OS from google. Merging Chrome and Android. Featuring deep integration of Gemini AI.

Memes!

Antigravity? Google's answer to claude code. Agentic will destroy machine sometimes.

LLM messed up, was so apologetic it refused to help, then uninstalled itself. Virtualized end points.

Alex : Stop Hooks to save credits

Peter : ArgoCD