Steer Protocol | 🐂

SparkDEX just leveled up to V4 — powered by Algebra. ⚡️ @SparkDexAI is live on @FlareNetworks, built on Algebra Integral infra, bringing a modular CLAMM engine to Flare DeFi. What’s inside: ↳ Hooks / V4 plugins ↳ Dynamic Fee Markets ↳ Flash Accounting ↳ Strategic Liquidity Configurations $5M TVL already on the Integral-powered DEX — and growing. 👌🏼

x.com/i/article/1981…

Phase 1 of the LP rewards migration from SparkDEX V3.1 to V4 is happening soon. Rewards will be shifted to @steerprotocol V4 vaults (wide and narrow strategies). Learn more about how to prepare:

1/ SparkDEX V4 is now LIVE! 💥 Users can swap, LP, and farm on the V4 DEX, built on @FlareNetworks. Key features: • Hooks • Dynamic Fee Markets • Flash Accounting • Strategic Liquidity Configurations 30% of LP rewards from V3.1 have been migrated to V4 — more below.

Recently, I've become fascinated with bending Solidity in weird ways. Though this i've build some amazing things. Biggest achievement to date has been a custom virtual machine that runs entirely onchain at the transaction level called Matador (It's literally a breakthrough for the industry). I tried to do the same with the @paradigm / @meridiandotapp / @danrobinson AMM challenge. Today I share an onchain quantized neural network, or QNN for short. The idea was basically: could I build a tiny inference system onchain and, if so… could I train a set of weights that would be competitive within the challenge? Most competition on the AMM challenge is concentrated on hyper optimizations so they can rank #1. My goal was to rank high but also provide something extremely novel and also potentially something that has been done before. Training cycle for this system was short and one learning through the research was that QNN was limiting in regards to it's ability to achieve further past a localized basin (hint: gradient descent). To achieve onchian inference I had to work within a sadly small QNN are interesting because they can guide an already existing system/algorithm but whats more interesting is MLP, where you remove all the fee specific logic and allow the model to be the full controller. Think of it more of the full brain. It worked... and I trained it for much longer. Total edge gain... +4.41. The best result I got was through an MoE (mixture of experts) setup during the training proces with regime gating. This training is much slower but luckily for apple users out there, we can leverage MPS and run it on the GPU. The hiccup for this work though is that the amm challenge 'stack' is python and rust based which means you get bounded by CPU in certain areas of the process. Everyone talks about AI and crypto, but have you ever thought about running the actual model on chain?! Crazy. AMM design is fun and I learned a ton so shoutout to everyone who is submitting and I hope you've learned something through my small write up. I'll share a more technical deep dive later this week as well as the sample codebase. Now let's flip the script. A new challenge awaits: Uniquely, I work on @steerprotocol , a leading liquidity manager for CLAMMs. We currently support over 60 or so AMMs today and have a very robust web IDE. This allows for quants to refine their strategies, complete with built-in agentic AI support and Monte Carlo simulations with full Greeks analysis so you can fine tune and tweak to your hearts delight. It's a turn playground for those designing liquidity strategies. Next Friday, I'm going to release a competition under the same thesis: who can build the best strategy. Do you think you can do it? Start training now: ide.steer.finance p.s. go read about matador. it's time for a step function change in how we interact with blockchains. mark my words, steer will melt faces this year.