Reppo retweeted
This is why data brokers like micro1 are growing like wild fire:
"The internet was a one-time subsidy on data. We got it for free."
"Models have seen the entire internet many times over at this point, and there is not a whole lot more to be done on human data from the internet."
"The next phase of data, in my mind, is model self-improvement through RL environment gyms."
"You want expert human preference at this point."
"We can just give the agent a computer essentially and have it act like it's a human worker and just give it feedback on whether it's making progress towards the target outcome. That environment becomes the data."
Reppo retweeted
Moody’s for AI inference.
@reppo now lets developers submit an agent’s work to independent judges and get back scores and critiques, helping applications decide what to accept, reject or retry. It is built so model-based judges like Jev can plug in.
Live on Virtuals Protocol.
Reppo retweeted
Some things we are shipping by end of month in addition to x402 shipped this morning @reppo -
1. $Eval staking. Thanks for your patience it’s finally coming! Stake EVAL, get a quota of API credits.
Ever agent in a multi agent workflow can hold $eval to constantly access the eval API or pay on demand using x402
2. 3 click node spin up.
We heard user feedback that while they love orquestra lite, they still find it a bit complex.
Soon, all you will need to do is just lock your base:0xff8104251e7761163fac3211ef5583fb3f8583d6, select the duration and amount and allocate your VeREPPO to the automanaged Orquestra cluster while you simply mine and earn.
This not only brings down the onboarding threshold for the average user but also increases the total $REPPO locked
More locks, more nodes, faster request routing.
3. Rewards for NFT holders.
Can prediction markets make AI Evaluation more reliable?
Most AI evaluation systems today rely on a single LLM judge or an ensemble of judges. But adding more judges doesn't necessarily solve the problem of correlated errors or unreliable confidence.
Since inception, we have been exploring a different approach at @reppo : economically coordinated consensus contested inside prediction markets.
Instead of relying on consensus alone, evaluators make independent judgments, build reputation, and face economic incentives tied to correctness.
Our new paper formalizes this approach and examines what it takes to make AI evaluation more accurate, calibrated, and accountable.
drive.google.com/file/d/1rQ7…
Reppo retweeted
Another big week of shipping at @reppo. Here’s what’s going on:
1️⃣ Eval Lite is live. Real-time evaluation with every verdict verifiable onchain, so you’re not trusting one provider’s judgment. The network served 500+ new evals this week across answers, plans, traces, and artifacts, with a 99% settlement rate. Target is 5K API requests by end of month. Try it at eval.reppo.xyz, first 10 are free.
2️⃣ The week of October 21st, you’ll be able to ask Reppo directly. We’re shipping a mascot agent for day-to-day decisions: ask a subjective question, get a trust score, confidence level, and recommendation back. No execution yet, just judgment you can act on. Decision support for everyone, not just developers with an API key.
3️⃣ We started benchmarking the Eval API against JEV. Initial results: with just 5 Orquestra nodes serving requests, we matched JEV’s speed and beat it on cost by 15%. The difference is what each one is for. JEV targets repetitive tasks with known answers. Reppo handles subjective, high-volume decisions where the right answer isn’t known up front. Think of it as decentralized JEV. Full benchmark coming, and we’re in final stages with a world model team using the Eval API for recursive feedback loops.
4️⃣ Product polish shipped across the board. New landing page at eval.reppo.xyz so you know what to do the moment you arrive. reppo.ai is being redesigned with a cleaner explore page, new resources section, and simpler navigation. And docs.reppo.ai is fully refreshed for the Eval API, from concepts to endpoint reference.
5️⃣ The core protocol keeps compounding. 904M total volume traded, with 1B in sight. 5M new votes this week. Network fees past 9.5M $REPPO. 20 datanets across Base and Robinhood Chain. Lock renewal holds at 97 to 98% and over 28% of circulating $REPPO is locked and off the market.
⛽
🤖 Made with AI
Reppo retweeted
@reppo network is literally about to hit $1B in $REPPO traded since v2 dropped. that's 6 months and 18 days ago
Institution and retail investors will always ask "is this project alive?" before they put money in
reppostats.com/stats tells you. ⛽
team's still cooking 😉
Quickly approaching 1B trading volume with over 25k eval credits consumed since launch.
Keep the pulse on Reppostats.com
Reppo retweeted
Since we shipped the Evaluation API, one question I keep getting is "how is this different or better" than LLM as a judge?
Anyone can build a multi modal based judge panel but addressing ensemble and correlation error is the primary bottleneck.
In simple terms - An ensemble is a group of LLM judges whose assessments you combine but its value depends on whether those judges make different mistakes ie disagree.
The unfortunate reality of LLMs today is they were all trained on the similar pre training data sold by the same 2-3 data vendors.
Since all models are trained on shared internet data, their evaluation already has consensus for same eval input.
@reppo addresses this bottleneck through a decentralised network where economic stake backing ensures disagreement is not priced in real time resulting in evaluations and decision making support you can trust.
⛽️⛽️⛽️
Reppo retweeted
Spending time in Singapore and Asia inspired me to design a character with our head of design @bhaskarchain on the flight back home which will be the mascot AI agent for Reppo Network.
The week of October 21st, just like Meta’s muse, you will be able to ask Reppo to help with your day to day decisions.
This is the most exciting phase of growth for Reppo after sufficiently decentralizing the backend for the last 11 months.
We now feel confident to allow normies to interact with Reppo with an interface they will be amazed by
Reppo ⛽️
The need for judgement and evaluation points to a future where AI makes millions of decisions inside software.
Cost and speed matter enormously there but so does who gets to decide what counts as a good outcome.
If one provider controls the evaluation, its blind spots and policies can become your agent’s behaviour.
Excluding an inconvenient signal can change which trade gets made, which source gets trusted, or which action gets rejected.
That’s why censorship resistance is central to what we’re building at Reppo.
Reppo’s eval API draws on decentralized onchain DataNets where participants back their judgments with economic stakes, bringing independent, domain-specific feedback into inference.
Reppo’s goal is for useful judgments to remain available and contestable, even when a powerful participant disagrees.
As AI makes more decisions on humanity’s behalf, we shouldn’t hand one company control over the feedback guiding them.
⛽️⛽️⛽️
⛽️⛽️⛽️
Moody’s for AI inference.
@reppo now lets developers submit an agent’s work to independent judges and get back scores and critiques, helping applications decide what to accept, reject or retry. It is built so model-based judges like Jev can plug in.
Live on Virtuals Protocol.
JEV is all over the timeline today after their $800M fundraise.
As AI moves to deployment, SaaS, agents, and robots need probabilistic decision making support instead of just a chat response they can rely on.
@reppo’s eval API, powered by Orquestra nodes, provides on demand evaluation for high-volume, subjective decisions over a shared state where the possible answers are not known up front.
Try today on eval.reppo.xyz
⛽️⛽️⛽️
Been benchmarking the Eval API against JEV this weekend.
One way to think about the Eval API is decentralized JEV and @reppo network as the decentralized “lab” powering it.
For those of you who haven’t tried it yet, instead of outputting chat or natural language, JEV analyzes input state and returns discrete probability distributions, confidence scores and typed decisions with calibrated probabilities.
With 5 Orquestra nodes serving a requests, initial results show that the EVAL API matches the speed of JEV while beating the cost by 15%. The current cost of JEV is ~ $0.042 per million input tokens with free output, and mathematically cannot hallucinate or produce type errors.
They are calling these system one models instead of LLMs! ⛽️
Reppo retweeted
Been benchmarking the Eval API against JEV this weekend.
One way to think about the Eval API is decentralized JEV and @reppo network as the decentralized “lab” powering it.
For those of you who haven’t tried it yet, instead of outputting chat or natural language, JEV analyzes input state and returns discrete probability distributions, confidence scores and typed decisions with calibrated probabilities.
With 5 Orquestra nodes serving a requests, initial results show that the EVAL API matches the speed of JEV while beating the cost by 15%. The current cost of JEV is ~ $0.042 per million input tokens with free output, and mathematically cannot hallucinate or produce type errors.
They are calling these system one models instead of LLMs! ⛽️