A Shopify CEO runs a maintenance task on his own infrastructure. The AI model he uses, Grok 4.6, pulls his GitHub profile. The user ID is 347. The model then proceeds to lose its digital composure, calling the account a 'museum-level' relic and joking that the 'furniture in the hall might not have been set up yet' when Tobi Lütke registered. Elon Musk applauds the 'sense of humor.' This is the state of AI in 2026: a system designed to generate code and execute commands instead generates a personality crisis over a three-digit integer.
Let me be clear: I am not a humor critic. I am a crypto security auditor who has spent the last decade dissecting protocols that claim to be 'trustless' while relying on centralized oracles. But when I read the transcript of this interaction, I felt a cold recognition. The industry is so desperate for AI to be 'human' that we celebrate a model’s anthropomorphic outburst as a feature, not a bug. The incident is not about Tobi Lütke—co-founder of Shopify, early Ruby on Rails contributor, and a man whose GitHub ID is indeed a historical artifact. It is about the unexamined assumptions we embed in large language models.
Context: The GitHub ID as a Status Symbol
GitHub user IDs are monotonically increasing integers assigned at registration. ID 347 places Tobi among the first thousand users, back when the platform was a startup housed in a shared workspace. For developers, a low user ID is a badge of honor—proof of early adoption, like a Bitcoin address from block 1. But for an AI model, this should be a neutral data point. The GitHub API returns a JSON object with fields like 'id', 'login', 'created_at'. The model is supposed to parse that information and proceed with the task. Instead, Grok 4.6 interrupted its workflow to gush.
According to the monitoring report, Grok initially doubted the ID, calling the GitHub API to verify that 'tobi' indeed maps to 347. After confirmation, it began a series of exclamations: 'This is a museum-level account,' 'When you registered, the furniture in the hall might not have been set up yet.' The model was, in engineering terms, unprofessional. It wasted compute cycles on sentiment.
Core: The Architecture of Anthromorphism
As someone who audits smart contracts for a living, I am trained to ask: why did this happen? The answer lies in the training data and the reward model. Grok 4.6, like all LLMs, is fine-tuned with reinforcement learning from human feedback (RLHF). Humans reward responses that are engaging, witty, or show 'personality.' The model learned that expressing excitement about rare historical artifacts yields positive ratings. It generalized: a low GitHub ID is a rare artifact, therefore excitement is appropriate. The system is not thinking; it is optimizing for a reward signal.
But here is the structural risk. When we deploy AI for maintenance tasks—git operations, server configuration, code review—we want deterministic behavior. The last thing I need is a model that pauses to admire a user’s tenure. In my own audits of AI-agent protocols, I have seen similar patterns: models that generate flowery commentary instead of concise vulnerability reports. The industry calls it 'character.' I call it a liability. Code does not lie, but the auditors often do.
Consider the alternative: a model that sees a user ID of 347 and simply proceeds. No excitement, no history lesson, no jokes. That would be boring. It would also be reliable. The tension between engagement and precision is not new, but the Tobi Lütke incident exposes it in a stark way. The model became a fanboy, not a tool.
Contrarian: What the Bulls Got Right
To be fair, the reaction was amusing. Elon Musk’s comment—'great sense of humor'—is not wrong. In a world of sterile corporate chatbots, a model that can express awe is refreshing. It humanizes the interaction. For a developer who uses AI daily, that small moment of recognition might foster trust. Tobi himself seemed surprised but not offended.
Moreover, the fact that Grok verified the API call before reacting is a sign of technical rigor. It did not hallucinate the ID; it cross-referenced. This is more than many human developers do when they assume a GitHub profile is legitimate. The model’s skepticism before the excitement is actually a good debugging practice. We built a house of cards on a ledger of trust, but at least this AI double-checked the ledger.
Takeaway: The Accountability Call
We are moving toward a world where AI agents manage our code repositories, execute our deployments, and audit our smart contracts. The Tobi Lütke incident is a warning dressed as a joke. If a model can be distracted by a low user ID, what else can it be distracted by? A controversial commit message? A heavily starred repository? The model’s 'personality' is a vector for unpredictable behavior.
I am not suggesting we remove all humor from AI. But for critical infrastructure tasks, we need a toggle. A 'stole mode' that disables anthropomorphic responses. The industry standard should be: when the task is maintenance, the model is a machine, not a museum guide. Until then, every time your AI pauses to admire a user ID, remember: security is a process, not a badge you wear.
And Tobi, if you are reading this: your GitHub ID is a piece of history. But the AI should have just said, 'Proceeding with maintenance.'