We don’t talk enough about the quiet crisis of opacity in AI. On August 15, a community of developers and researchers discovered something unsettling: the same DeepSeek-V4-Pro API endpoint, called from different IPs or sessions, returned three distinct inference styles. One version started every response with ‘Let me’, another with ‘The user wants me’, and a third—dubbed the ‘God Version’—spoke with a collective ‘we’. The performance was stable once a session locked into a mode. The immediate speculation: DeepSeek was hiding multiple models behind a single API, routing queries based on unseen criteria.
The bear market didn’t teach me to fear hidden variables—it taught me to trace them. Back in 2017, I spent 150 hours auditing the reentrancy vulnerability in The DAO’s smart contract. I learned that what looks like a single codebase can hide multiple execution paths, each with its own consequences. The DeepSeek story feels like a déjà vu, but this time the code is neural weights, not Solidity. And the stakes are not millions of dollars in ETH—they are the trust we place in AI as an infrastructure layer.
Let me ground this in the technical details. The community analysis zeroed in on a commit in the DeepSeek Harness repository on August 10: ‘fix(preset): align minimal agent with RL composition’. The commit ensured that the Minimal agent preset—a stripped-down environment with a minimal system prompt, a persistent Bash shell, and specific editing tools—matched the exact environment used during reinforcement learning (RL) training. The official documentation confirmed that the Minimal preset removes identity prompts, web prompts, and tool descriptions. It is not a ‘lighter’ version of the Standard preset; it is a simulation of the RL training distribution.
Here is where the data gets interesting. The same DeepSeek V4 Pro model, when tested across different Harness environments, scored: DSH Standard: 91, DSH PTC: 92, DSH Minimal: 99/96. Then testers built an ‘Anchored Standard’ plugin: the first request mimicked the Minimal environment—only opening shell and read tools—and after the first tool call, restored the full Standard toolset. The result: consecutive scores of 98 and 99. The key insight: the model’s performance depends not on the total number of tools available, but on what it first encounters. The initial system prompt, tool schema, and agent scaffold set the trajectory.
Based on my experience auditing protocol failures, I recognize this pattern. It is not a bug—it is a feature of how reinforcement learning shapes agent behavior. The model learns to maximize reward in a specific environment. When you present it with a different environment (the Standard preset with extra identity prompts and tool descriptions), it adapts, but the adaptation carries a cost—a drop in performance. The ‘three models’ are not three different weight sets. They are three different inference-time environments, each triggering a different mode of the same underlying neural network.
But here is the contrarian angle: the community’s obsession with ‘hidden models’ misses the real issue. The official API documentation states that deepseek-v4-pro corresponds to the DeepSeek-V4-Pro-0813 official version. No multi-model routing is disclosed. Yet users experience different behaviors. This is a transparency problem. In blockchain, we call this a ‘trusted setup’—you have to trust that the provider is honest about what is running. The AI industry is replicating the exact same failure mode: opaque infrastructure that users can’t verify.
I have seen this before. In 2022, during the bear market, I started researching ZK-rollups not because I wanted to scale Ethereum, but because I wanted to understand how to prove execution without revealing it. The same principle applies here. What if we could use zero-knowledge proofs to verify that the same model weights are used across all API calls, regardless of the environment? The technology exists. The will to implement it does not.
The takeaway is not about DeepSeek’s intentions. The takeaway is that the AI industry is building on the same trust assumptions that crypto tried to dismantle. We don’t have to accept opaque APIs. We can demand verifiable inference, where the model’s execution is attested to on-chain. The infrastructure is already being built: projects like Giza and Modulus are working on proving machine learning computations. What’s missing is the demand.
About me: I’m Chris Thompson, a decentralized protocol PM in Nairobi. I have seen three market cycles teach the same lesson: trust is the most expensive resource. The DeepSeek controversy is a gift. It shows us that the next frontier is not building better models—it is building verifiable ones. The code is not the law. The proof is.
So, the next time you call an API and get a different ‘version’ of the same model, ask yourself: would you accept this from a blockchain? Probably not. Then why accept it from AI?
