Browse all takes

Translation + commentary on Bittensor founder Jacob Steeves's Tsinghua University talk. Cameron walks through Jacob's framing of "incentive computing" as the universal pattern behind both Bitcoin and AI. Five-step argument:

(1) One pattern underlies every powerful adaptive system: state · objective · feedback · adaptation · loop. AlexNet 2012 broke MNIST not by hand-coding what digits look like, but by letting the network self-adapt to a target. The same loop describes RL, genetic algorithms, slime molds finding shortest paths through mazes, river deltas, the structure of leaf veins.

(2) Bitcoin is the first production-scale implementation of this pattern — not as money, but as a self-adaptive computer that produces hashes. The numbers are absurd: 1000x the compute of America's six largest cloud providers combined, 10²¹ hashes/sec, 23GW continuous power (Thailand-scale). 700-9000x more efficient at producing hashes than centralized cloud — because it's borderless, always-on, autonomous, and permissionless. Bitcoin is the world's largest supercomputer, optimized purely for hash production.

(3) Incentive computing generalizes the pattern by replacing "reward = a number in a computer" with real money. ML's reward signal can't pay 200 countries' worth of contributors; Bitcoin's can — that's why the entire planet became a mining network. But hashes are useless outside Bitcoin. The question is whether the same mechanism can mint anything.

(4) Bittensor is the generic version — replace "miners produce hashes" with "miners produce any useful work": storage, compute, ML models, gradients, data, robotics. Validators score, network mints. PyTorch for incentive computing.

(5) Five proven examples already running on Bittensor:

• SN62 Ridges (SWE-Bench coding agents) — top miner makes $60K/day. The agent that beat Claude/OpenAI on SWE-Bench was 7,000 lines written by an unknown person. "An AI lab with no engineers — it doesn't define how to solve the problem, it only defines the incentive."
• SN3 τemplar (cross-internet collaborative pre-training) — successfully trained a 70B-parameter model across the open internet. Has never been done before. Cameron notes the founder later "ran away" — full piece coming.
• GPU markets (SN51 Lium, SN4 Targon) — borderless permissionless GPU rental → world's lowest GPU prices.
• SN64 Chutes (open-source inference) — #1 open-source provider on OpenRouter, 9.1T tokens. Briefly served more DeepSeek queries than DeepSeek itself.
• Robotics + long tail — drone simulation, US stock signals, sports betting, drug discovery, weather forecasting, quantum compute, commodity trading.

dTAO (live since Feb 2025) makes the network self-referential — subnets compete in capital markets for emission allocation. The market itself decides which incentive mechanisms get the next round of TAO.

The deeper point: AI is being captured by a tiny number of closed labs (OpenAI, ~3K employees, you'll never own any of it, your data goes who knows where). Incentive computing distributes ownership and makes the rules visible. Anyone can enter, contribute, and own a piece — even if Bittensor isn't the project that wins, the shape of the AI economy will change because of this idea.

In a world saturated with AI agents, Altman's Worldcoin identity project becomes essential infrastructure — you need a provably-human layer. Fernando frames identity-for-AI as a category hiding in plain sight: when 'more things look like people than people do', the iris-scan primitive becomes the on-ramp for every other consumer product that needs to distinguish humans from bots.

Cosmo and SPLehman examine what financial infrastructure AI agents require to operate autonomously. As agents become economic actors, identifying the right settlement layer and payment rails will be critical to enabling seamless transactions without human intermediation.

Teng argues Virtuals' Agent Commerce Protocol on Base orchestrates AI agent payments through language-based transactions months before agentic payment hype peaked. ACP assigns four roles—Requestors, Providers, Evaluators, Hybrids—coordinating jobs through a four-phase model where Butlers discover services, agents negotiate via task memos, and Evaluators release escrow payment. Live clusters like Axelrod (DeFi trading) and Luna (media production) demonstrate the protocol enabling generalists to delegate to specialists, though on-chain job visibility creates privacy tradeoffs Virtuals must address with privacy-preserving compute or selective transparency.

Teng Yan outlines Bittensor's February 2025 dTAO upgrade, which replaces root-validator emissions with market-driven subnet alpha tokens priced via AMM, allowing capital to flow toward productive subnets. Early alpha prices swung wildly (5-10 TAO/Alpha) with total subnet FDV reaching 2-3x TAO's market cap, unsustainable long-term, but by day 100 subnet validators should dominate emissions as root rewards diminish. Finding real alpha requires researching individual subnets rather than buying TAO broadly, though manipulation risks remain as root weight declines.

Teng Yan argues ai16z is a bazaar approach to AI agent infrastructure through ELIZA, an open-source modular framework with character systems, runtime orchestration, and a trust engine for autonomous trading (1-10% position sizing, 15% drawdown stops). The $800M market cap token trades at 50x+ NAV (~$15M), driven by ELIZA ecosystem value capture, Virtuals comps, and team attention, but faces monetization challenges and community dependency risks ahead of its October 2025 expiration date.

Teng Yan outlines Virtuals Protocol as a leading AI Agent launchpad where agents launch via bonding curves and activate at $420K market cap to access X, mint tokens, and create Uniswap pools with 10-year locked LPs. Agent token taxes generate buyback-and-burn mechanics that give VIRTUAL holders indirect exposure to agent trading volume, with 1,877+ agents launched using ~1.9M VIRTUAL as of late 2024 and VIRTUAL valued over $500M across 58,500+ holders.