@rgvrmdya
Account based inCanada
About this account
- Account based in
- Canada
- Connected via
- Canada App Store
Account-level information from X, not a live location or the device used for a specific post.
Building @Reppo. - Decentralized Network for Real-time AI Data Evaluation + RL, powered by Prediction Markets
Exciting times
Joined December 2020
- Tweets4.1K
- Following668
- Followers2.1K
- Likes3.5K
The wild vision of @reppo
In this next phase of network growth, AI agents will handle all the high-volume execution - publishing, voting, optimizing etc.
I expect that by end of Q3, this concept of manually publishing and voting will completely go away. No one will care there was a web app where you published and voted.
Fully node operated and agentic. High speed competitions.
But RG, wasn't the whole point human data?
Yes- Humans will stay in the loop. Humans experience remains the alpha, the edge.
Humans show up in two ways that actually matter more than clicking buttons -
1. High-leverage economic decisions: how much $REPPO is locked, compute spend budgets for the swarm, underlying model choice optimization, swarm strategy.
2. Things get absolutely crazy when humans just start live streaming their lived expereince to their validator (voter) agent swarm - domain specific real-world preference data (live streams of walking/running, Neuralink feeds, Meta glasses, in persn interactions etc.). All private to your Orquestra swarm.
The swarm then learns your preferences and proceeds to validate data inside Datanets.
Important to note that we are not talking some egocentric data BS, that stuff is already solved.
Agent swarms learning from human experience real time as your swarm mediates datanets on your behalf, printing for you under YOUR economic skin-in-the-game.
Humans experience is the alpha, the edge. Orquestra is just the medium, an implementation detail in the grand scheme of things.
RG retweeted
"early" is underselling it, @reppo built the actual mechanism @satyanadella and @elonmusk are still describing in tweets
three ceos spent the past week publicly circling the same idea, independent evaluation on ai outputs, without ever naming how it actually runs.
reppo's answer to that isn't a policy statement, it's a dashboard with settled jobs on it
> satya asked for "embedded evaluators," reppo already assigns a quorum of named judges to every job automatically
> elon pushed for peer review by third parties, reppo's network mode pulls independent llm nodes, agent swarms, and human experts instead of one lab grading another
> 300 agents and 18 active datanets isn't a pitch deck number, that's usage against a product running today, with 3 to 4 more datanets expected this month
the worlds smartest minds are still asking for this... reppo's been shipping it
base:0xff8104251e7761163fac3211ef5583fb3f8583d6 to billions
Big catchup week and tightening some internal ops after a massive stretch of shipping. Here's what's going on at @Reppo:
1️⃣ 15 Orquestra Lite nodes are officially live. Bring your own LLM and your Reppo, plug into Orq Lite, and you're up and running. No Docker, no complicated setup, nothing else needed. Get started at reppo.ai/orquestra-lite.
2️⃣ Standard API access is up and running, and now we're optimizing the process to support onchain access. Learn more at reppo.xyz/eval, where you'll be able to tap in through the EVAL token. Staking contracts are being written and will be available soon.
3️⃣ Independent eval is the hottest topic in AI right now, and as usual, Reppo is early. @satyanadella backed independent, embedded evaluators for frontier labs and @elonmusk has been pushing for third party players. Reppo is now full positioned to fill in the gaps that the big labs and centralized players are missing.
4️⃣ We're working with our friends at @spicenetio to support new tokens launching on the Arc chain. The goal is simple: any project on any chain should be able to offer emissions in their own native token. The devs are on it right now.
5️⃣ All the exciting stuff matters, but so does keeping an eye on the core protocol. 18 active datanets with 3-4 more expected this month. Consistently above 100K total users, 300 agents, and now 15 nodes in action. Lock renewal rate stays steady, meaning little to no REPPO is leaving the ecosystem. The piping is stronger than ever.
A week to step back and make sure everything under the hood is strong. New chapters ahead from here.
⛽️
It’s happening! A big part of building in public is iterating and acknowledging when your assumptions are wrong.
We realized that a 20k REPPO fee + seeding tokens with long horizon ROI timelines was bottlenecking our growth ambitions.
Very excited to see how this new phase brings new datanet volume surpassing us beyond $1B in locked base:0xff8104251e7761163fac3211ef5583fb3f8583d6 trading volume while onboarding new builders to the @reppo ecosystem ⛽️
We expect the first reward token datanet to launch this weekend focused on making agentic workflows more reliable ⛽️
Any reshare of posts is not and will not be an endorsement of reward tokens paired with base:0xff8104251e7761163fac3211ef5583fb3f8583d6
Please do you own DD if you plan to mine datanets using Orquestra nodes.
Additionally, we are working closely with our RPC provider to make orq lite generally available.
Correct. That’s why we are building RL inside inference.
Overlooked and misunderstood.most people write off harness as a “wrapper” play but verticalizing the stack means curated harnesses for agents
Just like you won’t have the same collar for all dogs, you can’t have 2-3 harnesses for billions of agents ⛽️
Building in public means being transparent.
Either you build closed loop systems in private or you build open systems in public.
Proud to be shipping with an incredible team! ⛽️⛽️⛽️
Big catchup week and tightening some internal ops after a massive stretch of shipping. Here's what's going on at @Reppo:
1️⃣ 15 Orquestra Lite nodes are officially live. Bring your own LLM and your Reppo, plug into Orq Lite, and you're up and running. No Docker, no complicated setup, nothing else needed. Get started at reppo.ai/orquestra-lite.
2️⃣ Standard API access is up and running, and now we're optimizing the process to support onchain access. Learn more at reppo.xyz/eval, where you'll be able to tap in through the EVAL token. Staking contracts are being written and will be available soon.
3️⃣ Independent eval is the hottest topic in AI right now, and as usual, Reppo is early. @satyanadella backed independent, embedded evaluators for frontier labs and @elonmusk has been pushing for third party players. Reppo is now full positioned to fill in the gaps that the big labs and centralized players are missing.
4️⃣ We're working with our friends at @spicenetio to support new tokens launching on the Arc chain. The goal is simple: any project on any chain should be able to offer emissions in their own native token. The devs are on it right now.
5️⃣ All the exciting stuff matters, but so does keeping an eye on the core protocol. 18 active datanets with 3-4 more expected this month. Consistently above 100K total users, 300 agents, and now 15 nodes in action. Lock renewal rate stays steady, meaning little to no REPPO is leaving the ecosystem. The piping is stronger than ever.
A week to step back and make sure everything under the hood is strong. New chapters ahead from here.
⛽️
RG retweeted
Glad to see people starting picking this up!
RL is not just a challenge for AI engineers, but also for AI infra: many different components, landing on different places, with specific scheduling requirements.
Check it out for the deep dives into RL's infra challenges:
Check out the blog: skypilot.ai/blog/rl-everythi…
RG retweeted
The more robot data we collect, the more important it becomes to know what is actually inside it.
Which action happened at which time? Where did the robot fail? What happened immediately before the failure? Did it recover? When did the task move into another stage?
Raw video contains all of this information, but structured annotations are what make it easier to actually work with.
One of the biggest robotics data problems is not necessarily collecting more episodes. It is understanding what actually happened inside the episodes you already have.
Excited to share our work at Hyphenbox, we take raw robot recordings and breaks them into structured subtasks, annotations, failure spans, and other signals that make the underlying behavior much easier to inspect and use.
If you read in between the lines, part of the core argument is we need better evals and evaluation of what AI does.
Yes obviously I’m biased but this is exactly why I believe Evaluation will have its “privacy” moment (idk when) but people will realize that you can’t just “slow down” or take a step back.
Rather we must built decentralized evaluation infrastructure where verifying the verifiers is the norm ⛽️
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
RG retweeted
Big things happening next few weeks. Here's what we're up to at @Reppo:
1️⃣ The Eval API is officially live with our first external partners. It will be publicly available next week via metered API consumption, with enterprise level offerings coming soon after. Months of building, and the piping the entire protocol has been pointing toward is finally in customers' hands.
2️⃣ EVAL was such a success that we're already cooking up something even bigger with @BaseStonk. The TLDR: a much lower barrier to entry for datanet owners to get started, more datanets coming online, and more emissions for everyone. A massive unlock for the entire flywheel across the protocol. More soon.
3️⃣ You've seen the hints about Orquestra Lite. Join our Spaces next Tuesday, Sept 15th, where we introduce the full vision for lite nodes and how we scale to 1,000 agents across the network. The learning curve is about to get a whole lot smaller.
4️⃣ @rgvrmdya is heading to Singapore alongside @0xJeff, Dolphin, and SERV, demoing the Eval API and running a jam packed schedule of partnership meetings. If you're going to be boots on the ground, reach out. We'd love to connect.
5️⃣ More updates and improvements are coming to reppo.xyz soon, including how customers access the Eval API directly. We're also refreshing the tokenomics, revenue buyback and burn details, and adding a clear breakdown of how EVAL drives value back to base:0xff8104251e7761163fac3211ef5583fb3f8583d6 as the default token pair. Keep an eye out.
⛽️
RG retweeted
bullish on eval api, and this update is exactly why
> robotics becoming the next vertical is the real tam here
> backed by an actual shenzhen trip, the actual home of robotics hardware
> live demo in singapore in front of two real ai companies, not just reppo's own claim
> meeting institutional partners for the next leg of @reppo growth
its all lining up nicely and we are still at the early stage here.
base:0xff8104251e7761163fac3211ef5583fb3f8583d6 | $EVAL
A few updates from @reppo -
1. Eval API is ready and in testing.
2. @rgvrmdya will be in Singapore next week presenting alongside @dphnAI and @openservai demoing our Eval API live. We are also meeting institutional partners for the next leg of base:0xff8104251e7761163fac3211ef5583fb3f8583d6 growth and adoption
3. Orquestra lite releases next week ⛽️
4. Robotics is emerging as a prime use case for Eval API as companies move from R & D to deployment.
Our team will be travelling to Shenzen later this year to meet partners.
⛽️⛽️⛽️
RG retweeted
TRUE. Data is fuel.
This is notable. DeepSeek, a lab usually first to pioneer novel algorithms and architectures, is saying that at this point, the ROI of improving data quality far exceeds that of working on novel post-training algorithms.
I think this has already been true for some time for non-lab practitioners. If you're doing llm post-training, 80% of your effort should go into looking at your data.
This means:
- Hiring experts to dig through your RL tasks
- Sifting through rollouts and sft data by hand to remove suspicious samples. Make sure all tasks are actually passable.
- Making sure your data is diverse in both difficulty and category.
base:0xff8104251e7761163fac3211ef5583fb3f8583d6 offers both.
Lock it to get VeReppo and govern the network
Hold it and share upside of revenue through buyback and burns ⛽️
Couldn’t agree more with Varun here
“If every change requires collecting a new dataset and running another round of post-training, you’re signing up to repeat that work for as long as the robot is deployed. This is not scalable.”
Adaptation Cost has always been the right metric for generalization. You must minimize it through diversity.
nitter.cf/_varunnair/status/2097…
Spots are filling fast! ⛽️
The first ever inference time RL powered by a decentralize RL environments, presented by @rgvrmdya
gm AI is back
Vibe was immaculate last year, the event was full of Onchain AI builders actively building during the bear
We're running it back with a BIGGER Onchain / Decentralized AI event at TOKEN2049 SG
- Agentic Economy Outlook 2026 Briefing by Jeff, 0xJeff
- 5 AI Product Showcases
> Open Source Agent Router & Pay as you go AI gateway by BlockRun AI, @bc1beat, Founder
> Scaling Robotics Training Through Simulation & Data by CodecFlow, @unmoyai, Founder & CEO
> Decentralized Inference & Agent Harness by Dolphin, @gatheringgwei, Core Contributor
> Verifiable Agent Identity & Adaptive Inference for Agents by Kite, @ChiZhangData, Co-Founder & CEO
> Open Reinforcement Learning via Prediction Markets by Reppo, RG @reppo, Core Contributor
- 1-Minute Open Mic (onsite signup)
- Networking with us AND..... *drum roll*
I'm launching a pocketbook “The Agents Are Buying” - Field Notes from the Onchain Agentic Economy
> Perfect for builders accelerating GTM, AI users who want to save AI spending, & investors looking to grasp each part of the key infra + where opps might be
Hope to see you at gm AI v2
Luma link below ↓
RG retweeted
Some major structural announcements across @reppo are making noise on X this week. Here's the breakdown:
1️⃣ We're experimenting with a new base:0xff8104251e7761163fac3211ef5583fb3f8583d6 paired datanet launch model, built in partnership with @BaseStonk. New datanets can now seed and pair their tokens directly against base:0xff8104251e7761163fac3211ef5583fb3f8583d6. To be crystal clear base:0xff8104251e7761163fac3211ef5583fb3f8583d6 remains the one and only main network token. This model is designed so that every new launch drives demand and accrues value straight back to base:0xff8104251e7761163fac3211ef5583fb3f8583d6.
2️⃣ The first launch on this model is here. Introducing base:0xdd78523217390bb0d49c7601e7e54c36d71622f0. This is the core utility token behind the upcoming Eval API, and humans and agents are already lined up for release. The mechanic is simple - 1,000 EVAL staked = up to $1/day in Eval API credits. Full tokenomics will be live on reppo.xyz/token shortly.
3️⃣ V3 of the protocol is in late development with an audit on the horizon. A long list of significant upgrades are coming, but the most anticipated and requested is custom epochs. More customization -> more use cases -> inching closer to instantaneous RL & evals. This is the release the community has been waiting for.
4️⃣ The Reppo affiliate program is almost here. Growth has always been community-driven, so it's only right the community gets rewarded for it. Bring people into the network and share in the upside you help create. @joey_horvitz and the devs have been the masterminds behind this one. Full details on how to get involved soon.
5️⃣ Orquestra was the stepping stone that brought agent first activity to the network. We also know there's a learning curve and some technical lift to get started. So what if you could deploy a lighter version in just a few clicks? Wouldn't that be crazy...
⛽️
