Founding Engineer @MacrocosmosAI | @bitstarterAI 👾 Previous: Machine Learning (@opentensor, @parallel_domain) 🇨🇦
Joined January 2020
- Tweets1.1K
- Following506
- Followers913
- Likes3.9K
Pinned Tweet
I cannot overstate how excited I am to share this with the world. 16B trained over the internet by 4090/5090s with help from @green_compute_ and @TargonCompute. MFUs upwards of 40%, with compute all the way from the USA to Hong Kong. @IOTA_SN9
iota.macrocosmos.ai/dashboar…
Brian McCrindle 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.
Brian McCrindle retweeted
How can you tell if an incentive mechanism will survive mainnet?
By launching it on @Apex_SN1 first.
Join @macrozack and @macrocrux at 5pm to discover how our upcoming teams will be accelerated on SN1.
📅 Today
🕐 5pm UK | Noon EST
📍 Live on X @MacrocosmosAI
Brian McCrindle retweeted
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.
This is the exact subnet that @bitstarterAI needs. Apex is a playground for incentive, and what better way to get into Bittensor than to play in the fire. So excited to see where this goes!
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.
Brian McCrindle retweeted
I don't think people fully appreciate what @IOTA_SN9 is trying to build on #Bittensor..
They're essentially figuring out how to turn scattered machines all across the world into one coordinated training system.. even when that hardware is different and individual machines come and go
Which in itself.. is extremely powerful and difficult to do
But now they're also moving beyond pretraining.. into fine tuning and reinforcement learning
Genuinely think people are massively underestimating the scale of this
$TAO
Brian McCrindle retweeted
.@sfcompute fixes this
GPU financing is getting even more creative. The pattern i keep hearing about:
- Raise a seed/Series A round from a top 5 VC on a hypemaxxed idea.
- Other VCs pile in and now you have $800M
- Use $700M to buy GPUs
- When the hypemaxxed idea inevitably doesnt work, you have a small neocloud to fall back on.
- No VC pushes back or sues for fear of the tier 1 VC cutting them off of deal flow.
🍿
Brian McCrindle retweeted
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
Brian McCrindle retweeted
Pretraining is only one job. The iota SDK also runs RL and post-training, fine-tuning, and inference on the same disaggregated compute.
Brian McCrindle retweeted
The iota SDK, under the hood: three components form the engine, and a few lines of code give you the steering wheel.
Brian McCrindle retweeted
Dudes are gonna blow up $TAO
Great to talk to @TheStreet about our work at @MacrocosmosAI on aggregating compute to train frontier-scale training workloads.
We are steadfast in our work towards fusing the long tail of idle, heterogenous, interruptible compute into a global supercomputer.
finance.yahoo.com/technology…
Brian McCrindle retweeted
Great to talk to @TheStreet about our work at @MacrocosmosAI on aggregating compute to train frontier-scale training workloads.
We are steadfast in our work towards fusing the long tail of idle, heterogenous, interruptible compute into a global supercomputer.
finance.yahoo.com/technology…
Brian McCrindle retweeted
The iota SDK makes fragmented, interruptible compute train like one centralised cluster. Here's how, in two parts: the engine and the steering wheel.
Brian McCrindle retweeted
Very soon distributed training will just be called training.
Data center boundaries will disappear.
Clusters will be liquid and ephemeral. Workloads will be elastic and cheap.
Reserving compute will become unnecessary for most.
On-demand AI training.
Elon has Colossus. This is your own mini-Colossus.
On tap. Liquid.
Powered by Bittensor $TAO.
Brian McCrindle retweeted
Proud to announce the next stage of our work in @IOTA_SN9, the iota SDK and liquid compute, live at @ExploitSummit today!
We’ve distilled 2 years of R&D into a powerful set of communication primitives so that that anyone can train models using globally distributed, heterogeneous and unreliable compute with just a few lines changed from pure PyTorch.
We believe this fundamentally disrupts the economics of AI training.
Brian McCrindle retweeted
‘Train anything you want, decentralized’
@WSquires + @macrocrux just dropped the mother of all Bittensor SDKs 🚀
Pretraining, inference, fine-tuning: everything the @MacrocosmosAI team has developed is being opened to all.
Brian McCrindle retweeted
so dope, trained a 16B model with on what many would call old gpu's but it worked!
I cannot overstate how excited I am to share this with the world. 16B trained over the internet by 4090/5090s with help from @green_compute_ and @TargonCompute. MFUs upwards of 40%, with compute all the way from the USA to Hong Kong. @IOTA_SN9
iota.macrocosmos.ai/dashboar…
Brian McCrindle retweeted
Two machine learning teams. One #Bittensor subnet slot. You decide who gets it.
At pitchtensor, hosted by @bitstarterAI, two aspiring teams pitch to @macrocrux, @KeithSingery, @maxsebti, @EvanMalanga + @willobrien - then you choose the winner in a live crowdfund, online + in the room.
Live from the Exploit stage: luma.com/exploitsummit26