Yang Zhilin, founder of Moonshot AI—the company behind Kimi—recently told a blockchain media outlet that running a team is like training a neural network. Use Reinforcement Learning (RL) for the explorers, Supervised Fine-Tuning (SFT) for the rule-followers. As a Layer2 research lead who has spent 23 years staring at smart contract bytecode and ZK proofs, I recognize the surface appeal. But when I unpack the analogy using the same rigor I applied to the Bancor V2 weighted constant product formula, I find a structure that is mathematically elegant yet operationally brittle. The deeper issue is not whether RL or SFT works in management—it is that the reward function design, the very core of any RL system, remains undefined in human organizations. And undefined reward functions are exactly the vulnerability that causes DAOs to fork, DeFi protocols to lose millions, and Lightning Network channels to stay half-empty for seven years.
Context: The AI Training Metaphor
In machine learning, Supervised Fine-Tuning means you give a model labeled examples—‘this is the correct output.’ Reinforcement Learning means you define a reward signal and let the model explore actions to maximize cumulative reward. Yang’s translation: SFT is for junior employees who need explicit instructions, RL is for senior talent who should self-organize around objectives. The interview frames this as a new management philosophy for AI-native companies. The logic is intuitive, and it has already resonated with the crypto community—after all, blockchain governance often mirrors the same tension between rigid smart contracts (SFT) and adaptive DAO proposals (RL). But intuition is not verification. I need code, not vision.
Core: The Structural Flaws in the Analogy
1. The Reward Function Gap
In RL, the reward function is a mathematical expression—usually a scalar measured from environment state. In human teams, rewards are complex composites: salary, promotion, peer recognition, autonomy. Yang’s analogy collapses these into a single ‘freedom to explore’ signal, which is the equivalent of a reward function that only detects whether a transaction was included in a block, but ignores gas costs, MEV extraction, and finality delays. When I audited Aave’s interest rate model in 2022, I saw exactly this failure: the protocol rewarded lenders with fixed linear yields, ignoring real-time supply-demand dynamics. The market gamed it. Yang’s management RL will suffer the same fate unless a precise, multidimensional reward metric is defined—one that cannot be hacked by employees inflating PR counts or cherry-picking easy problems.
2. Sparse Reward and Credit Assignment
AI RL training struggles with sparse rewards—when the reward only comes after 10,000 steps, the model cannot tell which action contributed. In management, this is even worse. A developer might spend two months building a feature that only yields user growth after three quarters. Yang’s model offers no mechanism for intermediate feedback. Compare this to the way I verified zk-Rollup logic in 2020: the circuit constraints had explicit intermediate proofs for each step. No step, no final proof. Management without step-level reward calibration leads to either burnout (employees never see progress) or cheating (employees manipulate short-term proxies).
3. The Scalability Cliff
RL works best with small, homogeneous agents. Large-scale multi-agent RL is still an open research problem—reward conflicts, non-stationary environments, credit assignment complexity. Yang’s analogy applies to a team of 20, but when Moonshot AI grows to 500, the RL mode will fracture into competing local optima. I saw this pattern in the modular blockchain data availability audit I led for Celestia in 2022: ten thousand nodes each optimizing their own blob propagation strategy caused a latency bottleneck that required manual SFT-style intervention (hardcoded leader selection). The same will happen inside Moonshot AI.
Based on my experience decomposing protocol risks, the core insight is this: Yang’s analogy borrows the language of RL but discards its safeguards—the exploration cap, the entropy regularization, the value function baselines. Without these, ‘RL management’ is just another name for ‘let the strongest survive,’ which is not a strategy but a cultural drift.
Contrarian: The Hidden Symmetry Between SFT and Smart Contracts
Most commentators view SFT as rigid, RL as free. But from a cryptographic perspective, SFT is equivalent to a deterministic smart contract—every input has a defined output. RL is equivalent to a probabilistic state machine with off-chain oracles and governance. The blockchain industry has already learned that pure permissionless environments (RL) without base-layer rules (SFT) lead to chaos: see the 2016 DAO hack, where the contract’s governance process (RL-like on-chain proposals) was exploited because the underlying rule of ‘no reentrancy’ (a SFT invariant) was missing.
Yang’s interview acknowledges that ‘pure RL leads to gaming,’ but offers no concrete constitutional constraints. This is the same blind spot I identified in Lightning Network routing failures: the protocol assumed nodes would self-organize to optimize routing, but without a formal reward function for channel rebalancing, the network collapsed into low liquidity. Complexity is the enemy of security. Yang’s management RL is complex because it has no formal grammar. Every team member becomes a potential exploit vector.
The contrarian angle, then, is that the analogy may actually harm Moonshot AI’s long-term stability rather than help it. By branding the culture as RL-first, the company might attract high-variance risk-takers who optimize for personal reward at the expense of team cohesion. The opposite, an SFT-first culture with explicit protocols (like Netflix’s culture deck), has proven more scalable precisely because it defines the reward function in unambiguous terms.
Takeaway: Vulnerability Forecast for Crypto-AI Fusion Teams
Yang’s philosophy will likely produce a burst of innovation, much like the early days of Uniswap’s permissionless liquidity model. But without a verifiable reward function—something we can audit—the medium-term risk is reward hacking, talent alienation, and eventual governance crisis. I predict that within 18 months, Moonshot AI will need to introduce explicit SFT guardrails for project prioritization, performance reviews, and conflict resolution. If they don’t, the cultural entropy will become the smart contract vulnerability that brings down the system.
Check the math, not the roadmap. Audits are snapshots, not guarantees. And complexity is the enemy of security. Yang’s interview is a snapshot of a vision. The real audit begins when the team hits 500.