In cryptography, a padding oracle attack works like this: you probe a system, and how it responds reveals something about your own input. You learn the game by playing it.
Working on River has challenged my thinking about human-computer interaction in a way nothing else has, and that analogy kept nagging at me.
Working with an AI agent in public surfaces something you don't get in a private window: you see how people think. The questions they ask, how they frame them, where they get stuck. River doesn't just flow through your work. It reflects it; and as an engineer learning from some of Shopify's best, that visibility is everything. It's how I've learned best & it's how I'm still learning.
That's the ripple I didn't anticipate, all of us figuring out how to play. So... if you could ask an oracle anything, what would you actually ask?
I'd ask if I'm onto something here, or losing it. Maybe both.
(It's been a total honour to be a part of this effort, and we're just getting started 😝)
@mitsuhiko DSV4f + this advisor.ts pi extension I've found to be a nice combo - a call out to a frontier model every once in a while is still relatively cheap, but can help the open model a lot https://t.co/Ng5GfHZVJT
It has been a privilege and a blast to contribute to this effort! It's really the best to be able to see all the incredible things people are already doing with River, and make those things better
@tnm I always wondered what ever happened to Manhattan DB. It was planned to be open sourced from what I remember, but then never was. https://t.co/iaNR3lc34K
@shazow Very good questions imo experienced devs have a real advantage but only if they rapidly progress through their grief cycle and adapt, now and onwards. Categorically rejecting or ignoring the new layer would be a mistake.
@DavidSHolz@nearcyan Random but seeing this interaction in the TL reminded me of your old SVVR talk that for me was one of the more inspirational of that era - hard to believe it's been 10 years, 2015. Things did get weird but differently than predicted https://t.co/iShAaYOeYB
Today @Shopify is open sourcing the tool we use for optimizing glTF 3D models.
🔧 Tweak compression settings per texture
⚡ See changes immediately
🧰 Mesh compression options
🌐 Hosted online & free
👇 Check the thread for details
so are we all just gonna sleep on the finding that entropy collapse doesn't happen on GRPO when you filter out positive advantage samples from the batch and, while it converges slower rewards wise, it does it in a way that doesn't make the model have less variability
That means that it's possible to send byte sequences directly to the steering wheel in the format it expects.
Logitech doesn't seem to publish this spec anywhere, but with Wireshark + USBCap I was able to reverse engineer the exact bytes to send to the wheel to rumble.
This is a GIANT sleeper feature.
Every Shopify storefront now exposes an MCP endpoint with tools to query merchant's catalog, look up store info, build a cart, and order status!
Docs: https://t.co/NOqqWUUj6a — this instantly turns any MCP-capable agent into authoritative commerce agent.
@kalomaze I run an indie Beat Saber alternative called Soundboxing where all the maps are user generated (it predates Beat Saber) and have something like 4TB+ of user generated beatmap data & replays. Have always wondered about doing something like this. Please keep sharing what you find!
@jazzychad App Clips sound like a great idea, and are convenient for the user, but will tank your app installs - nobody ever converts from app clip to full app - so not usually wise from a business or customer acquisition POV
@kalomaze I made a dataset like that, hooked it up to the normalized probability of the '1' token, but had OOMs with GRPO - it's a cool approach though https://t.co/mIIHpEJwFO