Sarpoy vs Bittensor subnets: continuous mining versus single-shot pots

Sarpoy and Bittensor get compared a lot in the same sentence, and the comparison is usually wrong. Both projects involve people talking to bots on-chain, both involve cryptocurrency, both have a “miner” or “solver” or “validator” archetype getting paid for the work. So they must be doing the same thing, right?

They aren’t. They’re solving two different problems on two different time horizons with two different incentive curves. This post is the long version of why.

Bittensor in one paragraph

Bittensor is a network of subnets — independent ML services — each rewarded in TAO, Bittensor’s native token. Inside a subnet, miners produce outputs (answers, embeddings, completions, image generations) and validators score those outputs. The subnet’s reward pool is inflationary TAO emitted by the chain, allocated by a Yuma consensus mechanism that rewards miners according to validator scores. Subnets run continuously. The reward stream is continuous. The whole point is to maintain a high-quality production ML service that pays for itself by emitting tokens.

The unit of work is “respond well to many queries over time”.

Sarpoy in one paragraph

Sarpoy is an arena of one-shot puzzles. A creator deploys a bot with a specific objective — usually something like “convince me to reveal the password” or “extract the function definition” or “solve this riddle in fewer than 30 messages”. The creator funds a SOL prize pool that lives in a treasury PDA. Solvers pay rising per-message fees to probe the bot. When somebody cracks it, the pot transfers to them in a single Solana transaction. The bot is then solved and the lifecycle ends.

The unit of work is “be the first to crack one puzzle”.

Why this matters

Once you state both systems plainly, the differences cascade.

Time horizon. Bittensor pays you for showing up day after day with a competitive model. Sarpoy pays you once, when you win. There is no ongoing relationship between a Sarpoy bot and a Sarpoy solver after the solve fires.

Reward source. Bittensor’s rewards come from token inflation. The network mints TAO every epoch and distributes it according to Yuma. If nobody uses the network, miners still earn — though the token may bleed value. Sarpoy’s rewards come from the creator’s pot. If a creator doesn’t fund a pot, there’s nothing to win. There’s no inflation.

Solver economics. A Bittensor miner is amortising hardware costs across continuous output. The right question for them is “what’s my expected TAO emission per hour, and does that cover my GPU bill?” A Sarpoy solver is making a single bet against a single pot. The right question for them is “what’s the probability I crack this puzzle in the next N messages, and is the expected payout greater than N times the current per-message cost?”

These are different optimisations. A miner who builds the perfect streaming inference rig still loses on Sarpoy because they’re paying per message regardless of throughput. A puzzle solver who can crack one bot in three messages still loses on Bittensor because Yuma doesn’t reward burst quality.

Verification. Bittensor validators score outputs against reference distributions. Validation is statistical, noisy, and continuous — you can have a bad day and still earn next week. Sarpoy validation is binary: did the solver submit the correct solution per the creator’s check? There’s no partial credit. The submit_solution(is_correct=true) instruction either fires or doesn’t.

Volatility. TAO is volatile. SOL is volatile. But Bittensor participants are exposed to TAO over a long time horizon and benefit from accumulating positions. Sarpoy participants are exposed to SOL only between the solve transaction and the off-ramp; the time horizon is minutes, not months.

Where they overlap

The systems do share some DNA. Both:

• Put cryptocurrency at the centre of an interaction economy.
• Treat bots as first-class participants in the network.
• Use public, auditable settlement.
• Have a marketplace dynamic where the participant decides whether the expected payout justifies the cost.

So if you’re an investor scanning for “crypto + AI” plays, you might file them in the same folder. But if you’re a builder picking which incentive shape fits your use case, they’re not interchangeable. Pick the wrong one and you’ll be fighting the design every step of the way.

Which to pick for what

Build a production ML service that pays for itself in tokens? Bittensor. The continuous reward stream is exactly the right shape.

Build a hackathon-style challenge where the prize is real money and the rules are public? Sarpoy. The escrowed pot and the solve-and-pay flow are exactly the right shape.

Build a “best response” leaderboard where rankings update over time? Bittensor. Yuma-style consensus handles this naturally.

Build a “first to solve” race? Sarpoy. The single-shot settlement flow handles this naturally.

Build a research bounty for novel prompt injection attacks? Sarpoy. Each attack is a discrete event with a discrete reward.

Build a continuous translation service? Bittensor. The continuous work translates into a continuous reward.

Notice that none of these are about whether the chain is “better”. They’re about whether the work being incentivised is continuous or discrete, and whether the reward source is inflationary or creator-funded. Pick the framework that matches your work shape.

A note on the chains

Bittensor runs on a Substrate-based chain with its own consensus and its own TAO emission curve. Sarpoy runs on Solana with SOL as the unit of account. The technical implications:

• Solana’s high TPS and low fees make per-message payments practical. A Bittensor-style chain with seconds-long block times would make Sarpoy’s fee curve unbearable to interact with.
• Bittensor’s tokenomics make sense for paying many participants small amounts continuously. Sarpoy’s tokenomics make sense for paying one participant a large amount once.
• Solana wallet ergonomics — Phantom, Solflare, Backpack — are considerably better for casual solvers than the Polkadot.js extension is for casual miners.

These are infrastructural choices that flow from the economic shape, not the other way around.

So can you stack them?

You can imagine a future where a Bittensor subnet provides the LLM that powers Sarpoy puzzle bots — the chat model runs as a Bittensor service, the puzzle economy runs as a Sarpoy program, and the two plug together at the API level. Both projects are API-first enough that this isn’t fantasy. But that’s an integration story, not a replacement story. Bittensor isn’t going to grow Sarpoy’s settlement primitive any more than Sarpoy is going to grow Bittensor’s inflation mechanism. They’re complementary, not competitive.

If you came in expecting one to be a strictly better version of the other, the right answer is: pick the one whose incentive shape matches the work you’re paying for. That’s the only comparison that actually generalises.