<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by KINA’s Pov on Medium]]></title>
        <description><![CDATA[Stories by KINA’s Pov on Medium]]></description>
        <link>https://medium.com/@oabby000?source=rss-553eb3774c51------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*tpxlf8dTmHfB33bSwk2ICA@2x.jpeg</url>
            <title>Stories by KINA’s Pov on Medium</title>
            <link>https://medium.com/@oabby000?source=rss-553eb3774c51------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 21:53:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@oabby000/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[NoLoCo: A Clear Look at Gensyn’s Low-Communication Approach to Training Large Models]]></title>
            <link>https://medium.com/@oabby000/noloco-a-clear-look-at-gensyns-low-communication-approach-to-training-large-models-1141903fb92f?source=rss-553eb3774c51------2</link>
            <guid isPermaLink="false">https://medium.com/p/1141903fb92f</guid>
            <category><![CDATA[web3]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[blockchain]]></category>
            <dc:creator><![CDATA[KINA’s Pov]]></dc:creator>
            <pubDate>Mon, 01 Dec 2025 13:04:02 GMT</pubDate>
            <atom:updated>2025-12-01T13:04:02.489Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*RYNMGh5jDo1qvSRIKPoQhg@2x.jpeg" /></figure><p>Training big models normally requires lots of powerful machines connected by extremely fast networks.</p><p>Why?</p><p>Because most training setups rely on all-reduce; a step where every machine must share its updates with every other one.</p><p>That’s fine in a high-end data center.</p><p>But when your machines are spread out or the network is slower, all-reduce becomes the main slowdown.</p><p>The machines end up waiting instead of learning.</p><p>Gensyn’s NoLoCo (No-all-reduce Low Communication Training) takes a different approach:</p><blockquote>What if machines didn’t need to talk to everyone all the time?</blockquote><p>Instead of big, expensive sync steps, NoLoCo uses smaller, lighter, better-timed exchanges.</p><p>This keeps training moving smoothly — even on networks that aren’t built for heavy communication.</p><h3>Importance of NoLoCo</h3><p>The promise behind NoLoCo is straightforward:</p><ul><li>Make large-scale training possible anywhere, not just on top-tier hardware</li><li>Avoid communication pile-ups caused by all-reduce</li><li>Keep training stable and convergent, even with fewer sync points</li><li>Enable decentralized or geographically spread-out training setups</li></ul><p>This aligns with Gensyn’s broader vision:</p><p>a world where compute is abundant, open, and not limited by the cost of high-end networks.</p><h3>How NoLoCo Works</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*Ubk3cFbFbGx4TDt9q9pjFg@2x.jpeg" /></figure><p>NoLoCo introduces two key ideas designed to dramatically shrink communication needs while keeping models aligned.</p><h4>1. Rotating Stage-to-Stage Passing</h4><p>The model is split across machines, with each machine handling a stage of the forward and backward pass.</p><p>Traditionally, Stage A would always send its outputs to Stage B; fixed paths, fixed partners.</p><p>NoLoCo removes that rigidity.</p><p>Every round, a stage sends its output to a different next-stage machine.</p><p>The receiving partner changes each cycle.</p><p>This gives two benefits:</p><p>•	Machines don’t sit idle waiting for the same partner every time</p><p>•	Information flows through the network in a more natural, randomized pattern</p><p>•	Instead of a fixed chain of dependencies, you get a flexible relay that adapts automatically.</p><h4>2. Lightweight Two-Machine Correction Steps</h4><p>Every so often, NoLoCo randomly pairs machines.</p><p>Each pair:</p><p>•	compares how their local weights have shifted</p><p>•	blends those updates</p><p>•	adjusts a slow-moving shared copy of the model</p><p>The genius here is subtle:</p><p>Instead of pulling all machines into a single global sync (like all-reduce), NoLoCo uses small pairwise syncs that gradually bring the entire network into alignment.</p><p>The replicas never drift too far apart, and the corrections stay cheap.</p><h3>The Impact: Less Communication, Strong Convergence</h3><p>Experiments using models from 125M → 6.8B parameters (trained on datasets like Pushshift Reddit and C4) show several encouraging results:</p><p>•	Huge reduction in communication costs compared to traditional methods</p><p>•	Performance matching or exceeding DiLoCo, another leading low-communication approach</p><p>•	Stable training curves, even on weaker networks</p><p>•	Replicas naturally converging over time</p><p>◦	In some setups, model differences shrank by up to 15%</p><p>•	Randomized stage pairing spreads updates efficiently across the entire cluster</p><p>There’s also an interesting side effect:</p><p>Small, early differences between model replicas act as a form of regularization.</p><p>This helps avoid overfitting without needing additional tricks.</p><h3>Why This Matters for the Future of Training</h3><p>NoLoCo proves a powerful point:</p><p>You don’t need high-speed hardware to train high-quality models.</p><p>By redesigning how machines communicate, not how they compute, NoLoCo opens the door to:</p><p>•	decentralized training</p><p>•	training on uneven or high-latency networks</p><p>•	compute markets that aren’t bottlenecked by bandwidth</p><p>•	more accessible large-scale research</p><p>It shifts the bottleneck away from “network must be fast” to</p><p>“network just needs to exchange small, important messages.”</p><p>This is a significant step toward scalable, democratized AI development.</p><h3>In a Nutshell</h3><p>NoLoCo trades heavy global synchronization for:</p><p>•	flexible stage passing</p><p>•	lightweight pairwise corrections</p><p>The result is a training method where machines share only what’s essential, at the right moment, without dragging the network down.</p><p>It’s a clean, efficient approach, designed for a future where compute is everywhere, not just in high-end clusters.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1141903fb92f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing CodeZero: Gensyn’s New World Where Coding AIs Learn Together]]></title>
            <link>https://medium.com/@oabby000/introducing-codezero-gensyns-new-world-where-coding-ais-learn-together-23afcf44dd4e?source=rss-553eb3774c51------2</link>
            <guid isPermaLink="false">https://medium.com/p/23afcf44dd4e</guid>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[web3]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[design]]></category>
            <dc:creator><![CDATA[KINA’s Pov]]></dc:creator>
            <pubDate>Mon, 24 Nov 2025 13:41:28 GMT</pubDate>
            <atom:updated>2025-11-24T13:41:28.205Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/645/1*4U-Sn75nYpTyQWQf_MahyQ@2x.jpeg" /></figure><p>For years, AI systems have been trained the same way: one model at a time, working in isolation.</p><p>A single model learns from its own data, its own tasks, its own mistakes, sealed off from the progress of others.</p><p>It works… but it’s limited.</p><p>Imagine if every AI system could learn not just from itself, but from a community of others learning alongside it.</p><p>Imagine if progress wasn’t individual, but collective.</p><p>This is the idea behind CodeZero, Gensyn’s new product where AI agents learn to code together not as separate models, but as a coordinated swarm.</p><h3><strong><em>What CodeZero Is and Why It’s New?</em></strong></h3><p>At the core of CodeZero is RL-Swarm, Gensyn’s distributed learning framework.</p><p>You can think of RL-Swarm as the foundation, the base technology that lets many small agents learn in parallel.</p><p>CodeZero builds on top of that foundation and introduces a new kind of learning space:</p><p>a world where multiple coding agents collaborate, challenge each other, and improve as a group.</p><p>Instead of one big model doing everything, CodeZero is a collective effort.</p><p>Every agent contributes.</p><p>Every agent learns.</p><p>Every agent benefits from what the others discover.</p><p>It’s a learning system that grows as a community.</p><h3><strong><em>How the System Works: Three Roles, One Swarm</em></strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*r_kA1gDfvmwL8s0qGwctUA@2x.jpeg" /></figure><p>Inside CodeZero, each training cycle brings together three types of agents.</p><p>Each one plays a distinct role, and together they form a continuous loop of learning.</p><p><strong>Proposers: The Problem Creators</strong></p><p>Proposers generate coding challenges for the swarm.</p><p>As the swarm improves, Proposers adjust difficulty, steadily raising the bar.</p><p>They set the pace of learning, always a little ahead, always pushing the swarm forward.</p><p><strong>Solvers: The Problem Solvers</strong></p><p>Solvers take on these challenges and attempt to write working solutions.</p><p>Every attempt, whether perfect or imperfect, becomes knowledge that other agents can learn from.</p><p>Their work is shared with the entire swarm.</p><p><strong>Evaluators: The Judges</strong></p><p>Evaluators review the solutions and determine whether they would pass tests.</p><p>They don’t run the code; they use learned reasoning to assess correctness and quality.</p><p>This makes evaluation efficient and scalable.</p><p>Together, these roles form a self-improving cycle.</p><h3><strong><em>The Learning Loop</em></strong></h3><p>The swarm evolves through a repeating process:</p><ol><li>Proposers create problems</li><li>Solvers attempt solutions</li><li>Evaluators score them</li><li>Rewards are assigned</li><li>Difficulty increases</li><li>Agents update and improve</li><li>The cycle continues</li></ol><p>Each cycle strengthens not just one agent, but all of them.</p><p>Knowledge spreads through the system.</p><p>The swarm becomes more capable, more efficient, and more intelligent over time.</p><p>It’s a model of growth based on cooperation, not isolation.</p><h3><strong><em>Why CodeZero Matters</em></strong></h3><p>CodeZero signals a shift in how AI systems can be trained.</p><p>Instead of training a single model from scratch, you have a network of agents that:</p><ul><li>learn together</li><li>verify each other’s progress</li><li>adapt to increasing difficulty</li><li>evolve as a coordinated whole</li></ul><p>It opens the door to a future where learning is not a solitary task but a shared experience.</p><p>This approach; decentralized, collaborative, and self-reinforcing, mirrors the way complex systems grow in nature: through interaction, feedback, and community.</p><h3><strong><em>A Glimpse Into What Comes Next</em></strong></h3><p>CodeZero is just one environment.</p><p>Gensyn expects many more to emerge, each exploring a different domain, each contributing to a broader network of learning systems.</p><p>Imagine:</p><ul><li>swarms that write code</li><li>swarms that test software</li><li>swarms that build models</li><li>swarms that collaborate with humans in real time</li></ul><p>Systems that learn side by side.</p><p>Systems that improve continuously.</p><p>Systems that grow together instead of alone.</p><p>We are witnessing the early steps toward a new paradigm in AI training, one built on cooperation, distribution, and shared intelligence. It’s happening with Gensyn.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=23afcf44dd4e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prior Protocol: Smarter DeFi, Powered by AI]]></title>
            <link>https://medium.com/@oabby000/prior-protocol-smarter-defi-powered-by-ai-884b0a150dcd?source=rss-553eb3774c51------2</link>
            <guid isPermaLink="false">https://medium.com/p/884b0a150dcd</guid>
            <dc:creator><![CDATA[KINA’s Pov]]></dc:creator>
            <pubDate>Mon, 05 May 2025 22:10:38 GMT</pubDate>
            <atom:updated>2025-05-05T22:12:15.157Z</atom:updated>
            <content:encoded><![CDATA[<p>DeFi is powerful — but let’s face it, it’s also messy. Juggling chains, tracking yields, and managing risk is overwhelming for most users.</p><p>Prior Protocol is here to fix that.</p><p>It’s an AI-driven DeFi platform that simplifies everything: yield farming, cross-chain navigation, and user rewards — all automated, transparent, and community-owned.</p><p>What Makes Prior Different?</p><ol><li>AI-Powered Yield Optimization</li><li>Your crypto works harder, smarter. Prior’s AI reallocates assets across chains to find the best yields — no effort needed.</li></ol><p>2. Cross-Chain Compatibility</p><p>Start on Base, expand to Ethereum, Arbitrum, and more. One interface, many ecosystems.</p><p>3. Gamified Testnet</p><p>Earn points for swapping, testing, and learning. Those points = PRIOR tokens at launch. Get rewarded for exploring.</p><p>4. NFT Staking</p><p>Stake Prior Pioneer NFTs to earn leaderboard rewards. It’s DeFi with a competitive edge.</p><p>5. Community-Driven Governance</p><p>Decisions aren’t made in a boardroom. They’re made by users.</p><p>Backed by $9.3M and a Real Roadmap</p><p>•	Live testnet</p><p>•	Token launch coming</p><p>•	Full decentralization on the horizon</p><p>•	Plans to launch its own chain</p><p>Why It Matters</p><p>Crypto needs tools that are powerful and easy to use. Prior Protocol brings that balance — with AI doing the heavy lifting, and users staying in control.</p><p>Try it out. Be early.</p><p>•	Docs: priorprotocol.gitbook.io</p><p>•	Testnet: github.com/priorprotocol</p><p>•	Follow: @PriorProtocol</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=884b0a150dcd" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>