‘It's like coding with the curtains open’ SN13’s Amy & Vlad on DeAI's challenges and potential

Building and operating a subnet is complex, requiring a wide range of skills. Amy Chai and Volodymyr Truba, part of the SN13 team, reveal what it's like to work at the forefront of DeAI on Bittensor

How do you incentivise data-scraping on decentralized, open-source systems? That’s the core question that Amy Chai and Volodymyr Truba are devoted to solving on Subnet 13. To understand the challenges involved, we spoke with Amy and Vlad to explore their work and reveal an insider’s perspective from the frontlines of DeAI development.

How would you describe AI’s current level of maturity?

Amy: That’s a big question! I’d say the current state of AI maturity is at a teenage stage right now; there’s been a lot of promise and growth in recent years, but we still have a lot of room to improve.. However, the strides that AI has made in specialized domains like medical imaging and image generation are incredible.

In your opinion, what are some of the risks in developing AI behind closed doors?

Vlad: Unchecked biases, potential misuse… without public scrutiny, we can’t ensure that AI systems are being developed ethically and safely. Closed development can slow down overall progress in the field by limiting collaboration and knowledge sharing among researchers and developers.

Amy: To put it bluntly, the lack of transparency and accountability stemming from creating closed, centralized AI can lead to fear mongering and general mistrust around AI development as a whole. Yes, companies like OpenAI and Google make incredibly powerful large-scale models that are often people’s first encounter with AI, but cases such as OpenAI refusing to disclose details about their architecture and dataset construction for GPT-4 serve as harsh reminders that transparency is at the sole discretion of the company. And that’s not likely to lead to socially positive outcomes.

What has most surprised you about the emergence of LLMs?

Vlad: This is incredibly powerful technology. I am constantly surprised by its capabilities. But what's really blown me away is how LLMs are evolving into AI agents. It's like they've come alive! These agents can now tackle complex tasks step-by-step, make decisions, and even use digital tools.

What attracted you to the challenge of developing AI on the Bittensor network?

Amy: Having mined on Bittensor before the Revolution, I was super excited to see the emergence of so many new and interesting subnets. I’m all for visibility and responsibility in tech, and a decentralized open-source environment is the best example of that. 

Vlad: What attracted me to Bittensor is its unique approach to decentralized open-source. This open-source approach isn't just about being transparent – it's about sparking innovation. When smart people from all over the world can see how our AI works, they come up with ideas we'd never dream of. It's like we're all part of this global brain trust, pushing AI forward together.

How is developing on open source platforms different from engineering on proprietary technologies?

Vlad: With open source, it's like coding with the curtains wide open. Everyone can see what you're doing, pitch in, and learn from each other. It's way more collaborative and often moves faster because of all the brainpower involved. On proprietary platforms, you're kind of in a bubble. Sure, you might have more control, but you miss out the thoughts and ideas of the global dev community. With open source, like what we're doing on Bittensor, we're constantly getting fresh perspectives and ideas.

How do you see Bittensor’s long-term potential, and what role could SN13 play in it?

Amy: I see Bittensor as an incubator and environment for a wide variety of open source and decentralized AI products to thrive in. I think that Data Universe not only has the prospect to support other subnets through the more traditional route of providing training datasets, but in the future also as a tool to analyze emerging topics, track public opinion, and assist subnets in managing their market research. There’s so many things you can do with so much data!

How do you set incentives on SN13? What are the design decisions that have gone into those incentives?

Amy: Macrocosmos’s Data Universe is all about accumulating unique, fresh data. Our scoring is based on factors like data age: the newer the data, the more valuable it is, and data ‘labels’: certain hashtags for X posts and subreddits for Reddit posts are weighted higher than others, depending on what kind of data collection we’re trying to incentivize. Incentivizing newer data helps us keep up with the most up-to-date material, and prioritized data labels help us guide and reward scraping that we care about most, like posts about the crypto space. 

What excites you about the ecosystem Macrocosmos is developing?

Amy: The cohesiveness of all our subnets, and the potential that Data Universe has to support each one. From providing pretraining SN9 with open datasets for model evaluation, to fine-tuning LLMs in our emerging SN37, we’re really building a strong backbone for all our subnets.

Vlad: Every day, I see how we're pushing the boundaries in areas like pre-training, model throughput, inference performance, data scraping. In essence, being part of Macrocosmos means we’re directly involved in democratizing AI and opening up the future of intelligence. It's incredibly motivating work.