←
romAI79 posts
romAI avatar

romAI

@RomAI__

Turning scattered compute into faster, open intelligence. @reliquary_ai

◷ Joined May 2020▧ 3 Photos and videos
romAI avatar
romAI@RomAI__·Nov 11

Labs pay for what their models get wrong. What if an open network were generating the tasks frontier models still learn from?

Catching up with labs and data vendors lately, the same thing keeps coming up: labs now pay most for data on what their models get wrong. Volume got cheap. Models generate their own code, sim scenes and tasks now. The deals go to data aimed at a weak spot, or data nobody else has. 1. Meta, Nvidia and AWS are buying sim data from Physicl, a Paris startup that just raised a $10M seed. It's built only around scenes their robot models get wrong, like odd lighting or objects they've never seen. 2. 4 of the top 5 US closed labs pay Halluminate, a 9-person team, for finance RL environments like PE due diligence. On its test the best of 7 frontier models averaged 51%. 3. Google paid $10M for a bankrupt airline's 100M emails and 500M Teams messages, and outbid Mercor to get it. Heard exclusive deals like that run 4-5x non-exclusive ones. The data vendors I'd watch find failures before the lab does, or own something nobody else can sell. If you sell data to labs: are you getting paid for volume or for failures?

Post attachment
romAI avatar
romAI@RomAI__·Oct 31

100 steps of fully decentralized RL on Reliquary. Rollouts from untrusted miners on 12+ GPU models. That usually breaks RL: we fix the sampling noise publicly, so the trainer knows exactly what the model would have sampled. Every eval sample verified on-subnet. bittensor:native

Post attachment
Moe avatar
⟳ romAI retweeted
Moe@M0EGPT·Oct 28

Building synthetic training data and RL environments in our lab has me convinced 2026 is a great time to work on this. Frontier models are getting really good at generating realistic, complex tasks at scale. We've been testing how far a small set of expert seed data can stretch. Pairing domain expertise with frontier models lets us generate thousands of task variants while keeping the original reasoning, constraints, and edge cases intact. For RLVR, rubric design is the hardest part. You have to check business correctness and catch reward hacking, while making sure strict parsers don't mark valid work wrong and starve the model of reward signal. Task selection is tricky too. A few failed initial rollouts aren't enough to call a task unlearnable, since sparse pass-rate estimates are noisy and some hard tasks just learn slowly. We're building stateful RL environments with realistic tools. That's why I'm optimistic about synthetic data. As frontier models improve, environment generation gets cheaper to scale. The defensible advantage comes down to expert seed data, careful verifiers, and stateful environments that evolve with the models.

romAI avatar
romAI@RomAI__·Oct 26

Our RL run is fully permissionless. Right now the same model is being trained by B200s, RTX PRO 6000s, 5090s, 4090s and 3090s, side by side. 47 active miners. Watch it live app.reliqua.ai/monitoring/ru…

Post attachment
romAI avatar
romAI@RomAI__·Oct 23

Any GPU. Any architecture. Our permissionless RL run is live, open to anyone, from anywhere. A first for decentralized training. Join the run.

Today, we're launching a reinforcement learning run on our subnet to train Teutonic I, Bittensor's 9B-parameter model. It's our largest training run yet, and you'll be able to follow its progress through a public dashboard.

0:12
SemiAnalysis avatar
⟳ romAI retweeted
SemiAnalysis@SemiAnalysis_·Oct 17

GLM 5.3’s biggest leap came from post-training alone. “GLM 5.3 was built entirely by post-training the previous model. For me, that was a wake-up call.” “Post-training is far more GPU-efficient than pre-training from scratch. Plenty of companies can’t afford to train a model from scratch, but they have proprietary data. They can post-train themselves and beat some of the closed models.” “There’s still so much juice left to squeeze from post-training. We compared GLM 5.3’s cyber capabilities with 5.2, and it’s a night and day difference. That’s all just post-training.”

1:50
romAI avatar
romAI@RomAI__·Oct 14

This makes decentralized RL a lot more scalable. If samplers only need to stay within a bounded mismatch, far more GPUs and precisions can feed the same training run. That’s what we’re building at @reliquary_ai, permissionless. Come see it run.

Modal@modalOct 13

Training large models with RL typically runs on a fixed homogeneous cluster of the latest accelerators for numerical stability and faster throughput. This throws away inference optimization work done for other vendors and generations, limits the capacity you can utilize, and is unnecessarily expensive. So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.

0:12
romAI avatar
romAI@RomAI__·Oct 13

If you have a GPU doing nothing, read this. Every idle GPU could be doing post-training. Post-training doesn’t need a datacenter. It needs GPUs that are free.

A centralized lab can only use the GPUs it owns or rents. Reliquary can use any GPU in the world. That’s why we build on Bittensor. Here’s the key. Inference means low latency and no interruptions: an answer can’t stop halfway. Post-training doesn’t need either. Nobody is waiting on a rollout. If a GPU drops mid-task, nothing breaks. So any GPU can contribute. Idle at night, free between experiments. Join for an hour, leave, come back. The work still counts. A centralized lab pays for its GPUs 24/7, busy or not. We only pay for work that’s delivered and accepted. And every GPU owner can know what they’ll earn before they start. Our API now publishes, live, what each task pays per million tokens. Check your card’s throughput, and you have your number. reliqua.ai/api/sft/pay

romAI avatar
romAI@RomAI__·Oct 10

Just checked our dashboard. We generated 2B tokens of RL rollouts today on Qwen3.8-27B.

Xiaoyin Qu avatar
⟳ romAI retweeted
Xiaoyin Qu@quxiaoyin·Oct 9

People are laughing at data companies making $1B a year and call it NGMI. Entirely wrong. As RSI becomes more accessible, the bottleneck shifts from "a smart math genius knowing training recipes" to "access to high-quality data". Data companies create the most value in this new world. AI is the biggest math genius. We don't need more math geniuses. We need the most unique, best curated data with a strong point of view. Whoever owns this can either sell to labs, or train their own model that beat the labs. Just accept the new reality, guys.

Calvin Chen avatar
⟳ romAI retweeted
Calvin Chen@calvinchen·Oct 8

A look at our research infra. High-quality training data requires intuition built by training models and studying their failures. As tasks get more complex, this matters more. Creating data without training on it is like building a product you never use. Excited to share more.

At Proximal, we believe that building our own research infrastructure is crucial to design the highest-quality training data Today, we are sharing more information about our internal post-training stack

0:14
romAI avatar
romAI@RomAI__·Oct 8

Working on RL / post-training? We just opened early access to Reliquary's batch API: low-cost training data generation at scale. Bring your environment and tell us what you're training. DM me to try it! reliqua.ai

romAI avatar
romAI@RomAI__·Oct 7

Felt exactly the same. After those conversations I've never been more bullish: frontier models built end to end on Bittensor are coming. Great meeting you guys, let's make it happen.

Now that we’ve had a moment to process everything from @ExploitSummit, the energy and motivation to build are stronger than ever after all the conversations we had. Exchanging ideas with decentralized training teams like @MacrocosmosAI or @reliquary_ai and hearing @jondurbin talk about Parallax made one thing clear: Bittensor has the talent and energy to build models that can compete with SOTA. That's exactly what we’re working toward with @teutonic_tony . Really glad I got to be there with @DendriteHQ

Adithya S K avatar
⟳ romAI retweeted
Adithya S K@adithya_s_k·Oct 6

Finally, a RL simulation environment for port logistics 🚢 (v1) A lot of you said this should exist. So built it using real world port data. 👀

1:00

This would be a perfect RL environment. Take real shipping data, turn real logistics problems into verifiable tasks, and let models learn to optimize routes, schedules, inventory, and the entire network. Basically, logistics software as a strategy game for RL.