This episode reveals how human-centric design and decentralized feedback mechanisms are becoming the critical infrastructure for training smarter, more aligned AI within the Bittensor ecosystem.
Introduction to Tensorplex and Backprop
- Host Mark Jeffrey opens by praising Tensorplex's flagship product, Backprop, a trading terminal for the Bittensor ecosystem. He highlights its superior user interface, which provides a clear, at-a-glance market overview similar to CoinMarketCap, complete with subnet logos and performance data.
- Mark emphasizes that Backprop's design makes it easy to identify market-moving events, such as a specific subnet's performance shift, a feature he finds lacking in other platforms.
- CK, co-founder of Tensorplex, notes that the team is continuously evolving the product, adding features like a built-in, non-custodial wallet and an enhanced Telegram bot for market tracking and trading.
- A unique feature praised by users is the auditory feedback—a "clinking coin" sound—that confirms a transaction has completed on the Bittensor chain, which has a 12-second block time. CK mentions this was inspired by CowSwap's memorable "mooing" sound.
The Philosophy of Backprop: Decentralized Intelligence
- CK explains the profound meaning behind the name "Backprop." It's named after backpropagation, the fundamental algorithm in neural networks used to adjust model weights and improve accuracy. In the context of Bittensor, trading subnet tokens ($dTAO) is a form of backpropagation for the entire network.
- By buying or selling a subnet's token, traders provide direct market feedback, influencing the allocation of $TAO emissions. This shifts the network from a top-down, validator-controlled model to a bottom-up, market-driven system.
- CK frames trading on Backprop as more than just a financial activity; it's a direct contribution to the intelligence and decentralization of the AI network.
- CK: "You're not just trading or either making or losing money, but you're also, you know, helping decentralized AI every time you trade."
Deep Dive into Tensorplex Dojo (Subnet 52): Human Preference for AI
- Tensorplex also operates Subnet 52, named Tensorplex Dojo. Its mission is to solve a difficult problem in AI: integrating human taste and preferences into models. This is crucial as AI systems become more agentic and integrated into daily life.
- CK explains that while objective tasks like math can be learned by AI autonomously, subjective qualities require human input. This includes aesthetics in design, ethical alignment in behavior, and understanding nuanced social cues.
- Embodied AI, which refers to AI operating in physical forms like robots, will require a deep understanding of human preferences to provide services that feel natural and comfortable, not jarring or aggressive.
- Dojo aims to become the "Uber for human taste and preferences," creating a system to crowdsource this valuable feedback on demand.
How Dojo Works: Bootstrapping and Miner Mechanics
- Mark inquires about Dojo's operational focus, given the vast domain of human preference. CK clarifies that the subnet is currently in a bootstrapping phase, focused on generating and ranking user interfaces.
- The system generates four UI options for a given prompt: two intentionally flawed and two viable. Human miners are tasked with ranking these outputs.
- This setup allows validators to easily identify low-quality or automated miners. If a miner consistently ranks the flawed options highly, they are penalized.
- The miners are humans, often organized into teams by institutional-level mining operations, who provide nuanced feedback. CK notes this creates valuable jobs in developing countries.
- A key feature is collecting "rich human feedback," where miners must explain why one interface is better than another, providing qualitative data that is crucial for training.
The Importance of Human Feedback for Frontier AI
- CK debunks the idea that AI can improve indefinitely on its own. While model distillation (using a superior AI to teach a smaller one) is effective, it's insufficient for advancing frontier models.
- Improving the most advanced AI requires external, novel data, which in the case of subjective tasks, must come from humans. Relying on an AI to teach itself is like making a "copy of a copy," leading to quality degradation.
- CK points to OpenAI's success, suggesting its models possess a "higher EQ" and "just get it" because of extensive human feedback, even if other models outperform them on objective benchmarks.
- Dojo has already demonstrated this principle by training a 7-billion parameter interface generation model. The version trained with Dojo's human feedback data produced a "night and day" improvement in quality compared to the base model and one trained only on synthetic data.
Subnet Composability and Practical Applications
- Dojo is designed for subnet composability, meaning its human feedback mechanism can be used to improve the validation and output of other subnets on the Bittensor network.
- CK mentions imminent plans to work with subnets like 404.dev (3D generation) and the Score Subnet.
- For a 3D generation subnet, Dojo's human contributors can identify when a generated object (e.g., a cabbage) has an incorrect texture. For a sports-focused subnet, they can catch when an AI misidentifies a player's head as a football.
- This feedback not only makes reward distribution more accurate but also provides actionable insights for miners on other subnets to improve their models.
The New Telegram Bot: Enhancing User Accessibility
- Tensorplex developed a Telegram bot to meet users where they are. CK acknowledges the surprising popularity of trading via Telegram bots, which offer a faster, more streamlined experience than browser-based dApps.
- The bot provides price alerts, whale-watching notifications, and full trading capabilities for the Bittensor ecosystem.
- Crucially, the bot is built to be non-custodial, meaning users retain full control over their keys and funds.
- This initiative reflects Tensorplex's core philosophy of prioritizing user experience and lowering the barrier to entry for participating in the Bittensor network.
Market Discussion: $dTAO Token Dynamics
- The conversation shifts to the broader Bittensor market. Mark asks about the early issue with MEV (Maximal Extractable Value) bots, which seems to have subsided. CK acknowledges the problem is largely solved but is unsure of the specific technical fix.
- CK offers his perspective on the recent price decline of $dTAO (subnet tokens). He views the current market as a long-termist design that actively punishes short-term speculation.
- He draws a parallel to Ethereum's major drawdowns, suggesting that for long-term believers, these periods offer better entry prices. The current compression in subnet token prices presents an opportunity for those with conviction in the ecosystem's future.
Long-Term Investment Perspective on Bittensor
- Drawing on his experience since 2017, CK advises investors to "always zoom out." He applies this framework to Bittensor, assessing its potential long-term value.
- If Bittensor succeeds in becoming the world's leading decentralized AI network, its valuation could reach hundreds of billions of dollars. The current sub-$10 billion valuation implies the market is pricing in a high probability of failure.
- For investors, this presents a high-risk, high-reward scenario. CK shares his personal strategy:
- Hold a healthy stack of liquid $TAO.
- Deploy capital into a basket of top-performing subnets.
- Allocate a portion to earlier, higher-risk/higher-reward "moonshot" subnets like Dojo.
Dojo's Business Model and Revenue Strategy
- Mark asks about Dojo's path to revenue. CK confirms the subnet is currently pre-revenue, focusing first on proving its value within the Bittensor ecosystem.
- The initial "clients" are other Bittensor subnets, using Dojo's feedback to improve their own validation and outputs. This serves as a powerful proof-of-concept.
- The team is actively engaging with external AI startups, corporations, and academic institutions to understand their data needs.
- The long-term goal is to sell Dojo's human feedback and data labeling services to these external clients, offering a decentralized, scalable, and cost-effective workforce subsidized by the Bittensor network.
Tensorplex's Backing and Funding
- Tensorplex has raised over $3 million in a round led by prominent investors, including DCG, Accomplice, and Mechanism Capital.
- CK notes that the revenue generated from Backprop and their validator operations is sufficient to cover expenses, allowing the team to focus on long-term growth.
- He emphasizes the value of having investors from outside the core Bittensor community, as they provide fresh perspectives and challenge assumptions, which is healthy for development.
- He also highlights the importance of DCG's and Ooma's deep involvement, as their institutional expertise is critical for explaining the complex value proposition of Bittensor to a broader market.
Controversy: Subnet Emission Burns and System Predictability
- The discussion concludes with the recent controversy around some subnets burning a portion of their miner emissions, which has frustrated miners who invested capital based on expected rewards.
- CK states that while Tensorplex has not burned emissions, his core belief is in system predictability and trustworthiness. He compares a predictable system to the United States, which attracts global talent because its rules are stable and rights are protected.
- He argues that while subnet owners need flexibility, changing the rules arbitrarily mid-game erodes trust with miners, who are essentially the decentralized workforce of the subnet.
- CK: "You want to respect them or else, you know, there's no way you can have a like long-term successful subnets because people won't be incentivized to, you know, to focus on making the outputs better and better."
Conclusion
This discussion underscores that user experience and reliable human feedback are not just features but foundational pillars for decentralized AI's success. For investors and researchers, tracking teams that master these human-centric elements, like Tensorplex, is key to identifying projects with long-term viability in the competitive Bittensor landscape.
