Most crypto governance discussions start with token votes and quorum thresholds. They obsess over proposal formats and delegation curves. They rarely ask the question that determines whether a protocol survives its first liquidity crisis: Are we managing this network like a Reinforcement Learning agent or a Supervised Fine-Tuned model?
For the uninitiated, the terms come from AI training. Supervised Fine-Tuning (SFT) means you feed a model thousands of labeled examples. You tell it exactly what output is correct. Reinforcement Learning (RL) means you let the model stumble around a environment, reward it for good outcomes, and let its strategy emerge naturally. The analogy has been applied to team management by Moonshot AI founder Yang Zhilin. But the same dichotomy haunts every decentralized system.
Context
Crypto protocols are fundamentally governance machines. They encode rules about who can propose changes, how funds move, and what behavior gets slashed. These rules are either rigid (SFT) or flexible (RL). Most early blockchains, Bitcoin included, are heavily SFT: the consensus rules are hardcoded, upgrades require near-unanimity, and the behavior of miners is constrained by a strict reward function (block subsidy + fees). There is no exploration; only execution.
Ethereum shifted the balance. Smart contracts allowed for programmable logic, which introduced RL-like freedom: developers could experiment with new financial primitives, and the market would reward the successful ones. But Ethereum’s base layer remained SFT. The real RL explosion came with DeFi. Uniswap’s constant product formula is a rigid SFT rule, but the ability to create any pool, set any fee, and add any token is an RL environment. The protocol does not tell you which pair to launch; it rewards you with fees for providing liquidity to whatever the market demands.
Based on my 2017 audit of early ICO distribution mechanics, I saw the conflict early. Golem claimed a fair distribution, but my Python script revealed a 15% discrepancy between their emission schedule and actual tokens released. That was an SFT failure: the rule was there, but the verification was not enforced. The market did not punish them immediately; it took years for the trust to erode. The ledger remembers what the bubble forgets.
Core
The core insight is that crypto projects do not consciously choose between SFT and RL. They drift. A DAO launches with an SFT-heavy governance process (proposal -> vote -> execution), but as the community grows, the founders start rewarding exploration (grants, bounties, hackathons). The tension arises when the SFT constraints clash with RL incentives.
Take Aave V2. In 2020, I modeled its systemic risk under a 30% ETH price drop. The protocol had a rigid SFT rule: borrowers must maintain a certain collateralization ratio. But the reward function (interest rates) was RL-driven, responding to supply and demand. My model showed that 40% of users were technically undercollateralized in the simulated crash. The SFT rule was there, but the RL exploration of leverage had pushed the system to the edge. Liquidity is not depth; it is just delayed panic.
Now apply Yang’s management metaphor. He argues that RL (let employees explore, reward outcomes) is superior for innovation, but SFT (give explicit instructions) is needed for safety. Most crypto protocols invert this: they use SFT for trivial things (like token decimal limits) and RL for critical things (like oracle design). The result is a system that encourages risk-taking without adequate guardrails.
Layer2 scaling is a textbook example. Dozens of rollups have launched, each promising to scale Ethereum. But they are slicing already-scarce liquidity into fragments. From a management perspective, each L2 operates as its own RL agent, optimizing its own TVL and transaction volume. The Ethereum foundation acts as an SFT supervisor, setting standards and bridging rules. But the reward function is misaligned: L2s are incentivized to capture liquidity, not to contribute to the overall health of the ecosystem. The result is not scaling; it is fragmentation.
Based on my 2024 regulatory deep dive into ETF custody, I observed the same pattern at the institutional level. Traditional finance runs on SFT: clear rules, audits, compliance checks. Crypto’s RL culture rejects that as bureaucracy. Yet when an institution like BlackRock enters, it demands SFT guarantees. The clash produces hybrid products like ETFs that look like crypto but behave like traditional securities. The market rewards them with inflows, but the underlying RL agents (miners, validators, DeFi protocols) are still exploring in a environment where the reward function keeps shifting.
Contrarian
The contrarian angle: the crypto community worships RL as the path to innovation, but it is precisely the RL mindset that produces the worst exploits. Flash loan attacks are a perfect example of reward hacking. The attacker explores the environment, finds a sequence of transactions that yields a profit, and executes it until the protocol patches the loophole. That is the same behavior Yang warns about in human management: “completely RL may lead to gaming the system.”
Most crypto projects respond by adding more SFT: circuit breakers, rate limits, whitelist checks. But this pushes the system toward centralized control. The real solution is not a binary choice but a layered constitution: a small set of immutable SFT rules (the Bitcoin consensus model) that constrain the RL exploration space. Yang’s metaphor misses this distinction. He frames RL and SFT as complementary, but he does not address how to design the “constitutional layer” that prevents reward hacking without killing creativity.
In crypto, that constitutional layer is the base layer’s security and finality. Bitcoin’s SFT rules (21 million cap, PoW difficulty adjustment) create an environment where RL can safely operate on top (Lightning Network, Runes). Ethereum’s SFT is weaker because its governance is more malleable, allowing RL agents to capture value through MEV and governance attacks. The industry must stop romanticizing decentralized exploration without acknowledging the need for a supervisory fine-tuning layer.
Based on my 2022 analysis of algorithmic stablecoin de-pegging, I saw the failure of pure RL. Terra’s Luna had no SFT guardrails; it relied entirely on a market-driven reward function. When the exploration hit a negative feedback loop, the system collapsed. The market did not punish the explorers; it punished the holders who trusted the RL framework. The audit trail never lies.
Takeaway
The next cycle will separate protocols that understand this management metaphor from those that ignore it. The winners will be those that hardcode a small set of SFT principles (finality, immutability, auditability) while leaving maximum RL freedom for application-layer innovation. The losers will be those that chase pure RL — infinite exploration without constitutional constraints. The ledger remembers what the bubble forgets. Build accordingly.