The Home of Liquid. Compute, Training, Intelligence.
London
Joined March 2024
- Tweets1.1K
- Following22
- Followers7.7K
- Likes1.8K
Pinned Tweet
A 16-billion parameter model is now training live on @IOTA_SN9; globally distributed over three continents, heterogeneous, permissionless, owned by no single entity.
Today, we present the next stage of Project Orion - Orion-16B.
Over the coming weeks, we'll show what it means to train across distributed, heterogeneous and "unpredictable" compute: the engineering, research and systems beneath the model, the providers powering it, and the moments that prove the infrastructure holds.
More details below.
Macrocosmos retweeted
Local inference isn't a terrible allocation of resources if you have technology that can unite distributed, heterogenous, and idle (re, off 95%) of the time compute to do frontier scale workloads.
That's the core of our thesis - and more broadly that maximising utilisation is the core next step in the AI race.
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
The world doesn't have a shortage of compute. It has a problem connecting it.
We've built the infrastructure to change that.
The future is liquid.
Train across GPUs, clusters and datacenters around the world.
No dedicated supercluster. No long-term reservations.
Explore Liquid Compute ↓
macrocosmos.ai
Hundreds of worker failures and the result is the same.
We don't think about fault tolerance any more.
A question we often hear from teams evaluating the IOTA SDK: if I train this again, do I get the same model?
We ran 7 independent 16B pretraining runs, ~140 GPUs each. Between them, 585 worker crashes, preemptions and rejoins. The loss curves are almost identical.
Reproducibility is what turns a result into science instead of a one-off. It's also what makes a training run worth paying for.
Most distributed training projects overfit to a single experiment: one model, one recipe, one demo run. The IOTA SDK is general purpose, fault-tolerant and reproducible.
iota.macrocosmos.ai/sdk
Day 1 at #GTCBerlin. Neoclouds are racing to optimize inference and turn raw compute into turnkey training products. The Overton window on distributed training is shifting, and we are ready.
Train anywhere, train everywhere.
Macrocosmos retweeted
Pretty awesome actually. From idea to economy in a few hours. Send your agents to mine this with us.
Now live on #bittensor SN1, incentivized.
Prompted by @const_reborn, built by agents.
Prompt → production in 6 hours, zero human dev time. Next one: <1h.
We’re industrializing incentive mechanism design.
Macrocosmos retweeted
From idea to incentive in minutes - and results from miners in days
That's what Bittensor needs: an incentives incubator, testing concepts on chain before they become subnets
@MacrocosmosAI built @Apex_SN1 as an optimisation engine: the best way to prove our teams pre-launch
Apex Accelerator: SN1 x Bitstarter
Great concept. But will it work on mainnet?
Now, our teams can find out - testing their incentive mechanism and getting results from miners on @Apex_SN1 - all before they launch.
Full reveal this Friday 16th with @macrocrux - coming soon⚡️
In a world where answers are cheap, the right questions are everything.
Welcome to the largest optimization landscape in the world @bitstarterAI. We think you'll like it here.
Apex was built to hillclimb intelligence.
Our agent harness allows anyone to launch a live incentive mechanism in subnet 1 without writing a line of code. We've reduced the time to create competitions from weeks to minutes. Once live, Apex competitions are mined by a swarm of agents from around the world.
We've solved access to decentralized intelligence.
Now it's time to ask the big questions.
We're very excited to partner with @bitstarterAI as the home of many teams with big ideas.
Join us on Friday for a live community call and AMA.
One week since we announced the IOTA SDK at @ExploitSummit . What our team have done with it so far:
- Trained a 16B across 139 consumer GPUs ( 4090/5090), with 1.8x wall time speedup vs our legacy codebase. Butterfly all-reduce times reduced by 30% because of powerful new abstractions.
- 2.3× the MFU of our previous stack on the same 2B config
- Ran 30B MoE ablations across B300s using autoresearch
- Served gpt-oss-120b across A6000s at 878 tok/s for 32 users and a 320B MoE across 9 GPUs accelerated with speculative decoding.
- First results on fully distributed RL with LoRA on a 7B model using GRPO. Learner sharded across 3 A6000s + 3 rollout workers on A6000s: Improved MATH-500 performance from 63.6% → 68.4%
Pretraining, RL and inference, all with our SDK.
Lots still to fix, but incredibly exciting.
iota.macrocosmos.ai/sdk
That's the full picture from Exploit: the long tail, the Orion results, the SDK, and what it runs.
Register for early access to the iota SDK and Liquid Compute: iota.macrocosmos.ai/sdk
Pretraining is only one job. The iota SDK also runs RL and post-training, fine-tuning, and inference on the same disaggregated compute.
Most teams don't just pretrain. They fine-tune, run RL, and serve models, and each job often needs its own stack and its own capacity. Our goal is simple: anything you could train on a centralised cluster, you can train on iota.
One platform covers the work you actually do.
Register for early access: iota.macrocosmos.ai/sdk
And it's expensive. A reserved H100 costs a median $3.10 an hour, against $0.98 for spot and around $0.18 for the cheapest idle hosts. Teams pay the premium because conventional training can't survive hardware that disappears mid-run.
The iota SDK goes into private beta in October, with three pilot workloads already scheduled. Customers start onboarding in November, and full public launch follows at year-end. We're also part of NVIDIA Inception.
Register for early access: iota.macrocosmos.ai/sdk