Bittensor creator Const on Affine, dTAO, "mining reasoning," and more | E2326

Episode Summary

Executive Summary: The episode positions BitTensor/Tao as a permissionless, crypto-native marketplace for AI compute, inference, training, and eventually intelligence itself. Jake Steves (Const) explains how BitTensor adapts Bitcoin’s proof-of-work and Ethereum’s smart-contract abstraction into subnet-based digital commodities, where validators enforce performance and price becomes a market signal for useful work. The hosts frame it as a potentially huge, adversarial, world-scale experiment.

Main Topics: Bitcoin as the template for BitTensor (Priority: 5/5): The conversation starts by framing Bitcoin as the first digital commodity and permissionless peer-to-peer monetary network, then extends that idea to BitTensor as a computational network for AI workloads. Permissionless markets and adversarial design (Priority: 5/5): A major theme is that open networks attract both top talent and bad actors, so BitTensor’s core challenge is to create mechanisms that stay valuable even under active attack. Subnets as AI service marketplaces (Priority: 5/5): Const explains that BitTensor subnets are open networks where contributors provide inference, training, storage, and other services, with validators checking quality and speed. Tao, alpha tokens, and incentive engineering (Priority: 5/5): The episode details BitTensor’s token model: Tao has a fixed supply, subnet alpha tokens represent individual networks, and market price is used to rank and reward productive subnets. Affine and the race toward intelligence (Priority: 4/5): Affine is presented as a subnet focused on mining reasoning and ultimately measuring intelligence, which the speakers frame as the next frontier beyond inference and compute. Templar, decentralized training, and the rug-pull lesson (Priority: 4/5): The hosts discuss a failed decentralized training subnet and what it revealed about the difficulty of building robust incentives in crypto and the need for better vesting/lockup norms. Foundation, governance, and ecosystem maturity (Priority: 3/5): Const explains the split between foundation roles, internal governance, and the project’s evolution from experimentation into a more structured ecosystem with multiple teams and evolving mechanisms.

Key Arguments: Bitcoin’s core innovation was not just money, but a permissionless system for producing a digitally defined commodity through computation and electricity. BitTensor generalizes that model from hashing to AI-related work, letting open participants compete to provide inference, training, storage, or other useful services. The key problem is adversarial: if a network is open, it must remain valuable even when malicious or low-quality actors join. Validators and cryptographic proofs create programmatic SLAs, so networks can verify that a miner is actually performing the advertised AI task. Market price is used as a ranking signal for subnet quality; useful subnets attract capital and users, while poor ones lose support. BitTensor’s open design intentionally lets anyone participate, which increases innovation and scale but also introduces scams, speculation, and attack surfaces. The ultimate ambition is not just cheaper inference or training, but a decentralized way to compete with frontier AI labs and measure intelligence as a commodity. The Templar failure showed that decentralized training is far harder than inference, because training requires synchronization, bandwidth, and extremely fragile correctness. The ecosystem’s incentive model works because it converts excess compute into productive revenue through permissionless participation. Speculation is inevitable and not the goal, but it can still help bootstrap adoption and awareness if the underlying product has real value.

Data Points: Bitcoin cap: 21 million - Const says Tao is capped similarly to Bitcoin. Bitcoin block timing: ~12 seconds - Const references Tao issuance after the first halving and says a hash is produced every 12 seconds. BitTensor subnet count: 128 - The hosts discuss the current number of subnets and examples like Affine at subnet 120. Typical subnet league size: 256 - Const says a normal subnet is usually 256 participants. Subnet registration cost: 688 Tao - Current Dutch-auction price to register a subnet at the time of the conversation. Approximate USD cost to register: $121,000 - Roughly equivalent cost for one of the 128 subnet slots. Revenue scaling example: 3x - Const says some mechanisms tripled miner payments and tripled compute in two months. Model comparison: 35 billion parameters - Const says a BitTensor-produced model beat Qwen’s best model in the 35B range before being overtaken. Temporal scope: 2 years - People have been building mechanisms on BitTensor for two years. Project timeline: Launched in 2021 - Const says the network launched in 2021.

Pivotal Quotes: "It's not an understatement to say that we are up against nation states because this is truly like China versus the United States." — Const: Describing the competitive landscape for decentralized AI and why BitTensor must be built to withstand extreme adversaries. "For some people, mining BitTensor is like the most fun they've ever played ever." — Const: Explaining why participation is not just financial speculation but also an engaging, game-like incentive system. "I would say that is the entirety of the project." — Const: Answering whether adversarial thinking is central to subnet design; he says defending against bad actors is the whole point.

Implications: BitTensor is trying to turn AI infrastructure into an open, market-driven game where usefulness is rewarded and bad actors are filtered out. If it works, it could create a decentralized alternative to centralized AI labs and a new model for building internet-native companies.