Are SKILL.md files the Quantum Error Codes of Industrial AI?

ELI5/TLDR

AI companies have a trust problem: their models are probabilistic (read: they sometimes make things up), and industry needs results it can bet money on. The current fix is wrapping the AI in hundreds of deterministic “skill” files that tell it exactly what steps to follow — like building bamboo scaffolding around a wobbly building. The speaker argues this is structurally similar to how quantum computing uses error correction codes to tame chaotic qubits. His pitch: stop piling on scaffolding and invest in making the core AI itself more reliable.

The Full Story

The Trust Gap

Industry runs on reproducibility. Finance, aviation, medical pipelines — these are domains where a 1% error rate across millions of operations is not a rounding error, it is a catastrophe. AI models are statistical. They have a non-zero hallucination rate baked into their architecture. This is not a bug that gets patched. It is a feature of how the math works.

So industry looks at AI and says: interesting, but no thanks. A McKinsey report from March 2026 found that 74% of industrial-sector respondents flagged inaccuracy as a highly relevant risk. Not “somewhat concerning.” Highly relevant.

The Scaffolding Strategy

The current answer from companies like OpenAI and Anthropic: wrap the probabilistic core in deterministic scaffolding. Skill markdown files. Python scripts. Step-by-step workflows numbered 0.1 through 0.17. The AI doesn’t freelance — it follows instructions.

“We have now just in one model more than 460 defined skills.”

The speaker frames this as a cost-saving measure too. Training a smarter model is expensive. Bolting on external scripts that handle the predictable stuff is cheap. So the intelligence stays in the middle, and the scaffolding does the heavy lifting on the outside.

The problem he keeps circling back to: the execution of those “deterministic” steps is itself handled by a probabilistic model. The scaffolding looks solid from the outside. From the inside, it is still running on vibes.

The Quantum Analogy

Here is where the video gets ambitious. The speaker draws a structural parallel between AI scaffolding and quantum error correction (QEC).

A qubit — a quantum bit — lives in a two-dimensional complex Hilbert space. That is a fancy way of saying it holds more information than a regular bit (which is just 0 or 1), but that extra information is fragile. Three problems make quantum error correction hard:

No-cloning theorem. You cannot copy a qubit. In classical computing, error correction means “keep backups.” Quantum physics says no.
Wave function collapse. The moment you measure a qubit to check for errors, you destroy the quantum state you were trying to protect. The act of looking breaks it.
Continuous errors. Classical errors are simple — a bit flips from 0 to 1. Quantum errors are continuous, meaning they can drift by any amount in any direction.

The breakthrough (circa 1991-1995): you do not actually need to correct for infinite continuous errors. Any single-qubit error can be written as a combination of three Pauli matrices — a bit flip (X), a phase flip (Z), or both (Y). When you measure using the right error-checking operator, the continuous error collapses into one of these discrete categories. The measurement itself makes the problem manageable.

The speaker’s claim: this is structurally what skill files do for AI. They take the continuous, unpredictable output space of a language model and force it through discrete, manageable steps. The scaffolding is the error correction manifold.

His Actual Position

Despite spending most of the video building the analogy, the speaker is not a fan of the scaffolding approach. He thinks it is a band-aid driven by commercial pressure — AI companies need revenue, VCs need returns, and industry needs to be told the system is safe. So companies build elaborate scaffolding and sell confidence.

His preferred direction: invest in making the core model smarter and more reliable. He cites an experiment where a 4-billion-parameter model, tiny by current standards, was trained on domain-specific knowledge using supervised fine-tuning and reinforcement learning. It outperformed a Gemini 3 Pro model on that domain.

“I’m not interested here to limit the intelligence of an AI system with more and more scaffolding to elude the industry that all these AI systems are safe. They are not safe.”

He wants the scaffolding loosened. Let the chaotic intelligence out more. Make the core better instead of wrapping it in duct tape.

Claude’s Take

The quantum-AI analogy is visually suggestive and structurally hollow. Yes, both quantum computing and AI have a noisy probabilistic core surrounded by stabilizing infrastructure. But that describes… most engineered systems. A car engine is a controlled explosion wrapped in metal and cooling systems. The analogy works at the level of “both have a wild middle and a tame outside,” which is not saying much.

The specific parallel — Pauli error decomposition maps onto skill-file discretization — sounds precise but is not. Quantum error correction has rigorous mathematical guarantees. Measuring a syndrome collapses a continuous error onto a discrete one in a formally provable way. Skill files have no such property. They are prompt engineering. The “deterministic” steps still run through a model that might hallucinate step 0.7 into something creative. The speaker actually acknowledges this (“the execution of the deterministic step is not deterministic, but nobody tells it is”) but then keeps building on the analogy anyway.

The McKinsey stat (74% flagging inaccuracy as a risk) is real and well-placed. The observation that scaffolding is partly a commercial confidence trick rather than a technical solution has teeth. The argument that we should invest in core model intelligence rather than external guardrails is a legitimate position in the field, though it glosses over the fact that both approaches are being pursued simultaneously, not as either/or.

The presentation style is rambling. The speaker circles the same three ideas for 20 minutes, with frequent references to his previous videos and AI-generated images. Signal-to-noise ratio is low.

Score: 4/10. One genuinely interesting observation (the scaffolding-as-confidence-theater angle) buried under a stretched analogy and a lot of repetition. The quantum physics explanation is competent but the bridge to AI does not hold weight. Worth a skim if you are thinking about enterprise AI trust. Not worth the full runtime.