been watching these guys cook for a while now. looking forward to meeting the minds behind sn62!.

RT @proofoftalk: The #Bittensor community showed up strong at Proof of Talk, bringing energy, insights, and real-world grit to every corner….

It’s absolutely crazy to me that subnet 9 miners cracked model merging before the literature caught up. And by before, I mean at least 6 months. Incentivized AI research is the new meta.

My hope is that, in the near future, subnet 1 miners will write all of Bittensor’s papers.

The currency of AI research is papers. To establish Bittensor as a world leader in AI research we need to work together as a community and regularly publish our work. Over a hundred smart teams have been innovating and building novel technologies, but unless we share our.

This recent post by @LangChainAI on building open deep researchers is a great introduction to the key ideas and approaches that's we've been busy incentivizing

You can play with our Fast Incentivized Reasoning Engine (FIRE 🔥) through the app and api. I think it would be very cool to give users a ‘slider’ to allow them to choose how much time to give the miners. Interesting tradeoff between speed and quality. .

We’ve entirely re-written the subnet 1 codebase, which is now around 95% smaller, much faster and laser focused on optimizing the most powerful elements of our own deep research for the past 6 months (TTC, agents and GANs). We’ve also added a secure sandbox to enable our agents.

Personally, I use deep researchers every day to help me with basically everything. They are probably my favourite AI use case so far, but they take too damn long. The world just hasn’t come to expect fast results yet I guess. Soon, it will be very different.

Subnet 1 is doubling down on building the fastest deep researcher in the world. This will allow anyone to create accurate and detailed reports (even research papers) in seconds, not tens of minutes. We're pushing a major upgrade to the network soon and writing the SN1 DR paper.

RT @v_truba: Single miner. Single dataset. 2.9 billion rows. Bittensor SN13 → @huggingface → 30k downloads(validat….

RT @MacrocosmosAI: 🪐 Gravity is getting an upgrade 🪐. Our frontend devs are cooking again 🍲. 🧑‍🚀 Pop-up mission commander.📹 YouTube search….

IOTA needs to be able to process trillions of tokens in order to train powerful models. We know this better than anybody, and we're not stopping until it's done. The next generation of IOTA is coming on July 23rd. More speed, less haste.

Something I'm obsessed with right now is a new version of CLASP, which is based on tournaments between random bundles of miners that stretch out down the pipeline. This solves a major bottleneck of IOTA -- validating work in permissionless training runs.

Meanwhile, our servers are busy running longer experiments on the bottleneck llama architecture .developed by @MacrocosmosAI . This new design compresses model activation sizes by 128x. Losses are looking great after training on billions of tokens.

For the next couple of weeks we will be deep in the code, profiling and building a fully horizontally scalable orchestration system that can handle internet scale training. Cracked mining teams have helped us get a 360 view and make the new version much, much more powerful.

Within a week, the miners optimized by around 10,000x. The power of Bittensor brought our orchestrator to it's knees. We maxxed out all our machines. The orchestrator still couldn't keep up. Then two weeks ago we began re-architecting the orchestrator from the ground up.

While developing IOTA we ran weeks of experiments on testnet with 30 A100 miners. We load tested the orchestrator with 100 times the traffic we were creating (>10x full base network) and we marched to main net. But this is Bittensor. .

RT @dsvfund: Bittensor Revenue Search 12: Mike Bunting - SN13 Data Universe

Amen.