There's no point at which turning your brain off will work
The Futility of the "Brain-Off" Approach to AI
In the early days of 2025, a concerning trend emerged: people began turning off their brains delegating their entire cognitive process to Large Language Models (LLMs).
The pattern was simple: an LLM would perform a task—perhaps summarizing a document or generating a script—and the user would simply assume it was correct. In early 2025, this was usually a recipe for disaster, often resulting in absurdly incorrect outputs.
The Rise of the "Meat Proxy"
Some users attempted a slightly more sophisticated approach. If the AI's first attempt failed, the human would act as a middleman, feeding the error back to the LLM and asking it to fix the problem. Niklas Gruhn describes this specific role as being a "meat proxy."
Essentially, the human becomes a biological for loop:
By September 2026, this "meat proxy" method has become surprisingly effective. Some software developed this way actually functions. However, this creates a dangerous paradox for the employee.
The Economic Trap
If an LLM improves to the point where a "brain-off" human can produce average or even great software, the fundamental question becomes: Why does the company need the human?
| Scenario | Role of Human | Outcome for Employee |
|---|---|---|
| Low LLM Capability | Active Architect/Coder | Essential |
| Medium LLM Capability | "Meat Proxy" (Looping) | Redundant/Replaceable |
| High LLM Capability | None (Fully Automated) | Laid Off |
The Bottom Line: There is no threshold of AI improvement where the "meat proxy" methodology becomes a sustainable career strategy. If the AI can do the work via a simple loop, the company will eventually just automate the loop and remove the human.
High-Value Work vs. Low-Value Tasks
As noted by Luke Burton, the ability to work "brain-off" reveals a lot about the nature of the task. If a task is low-value and can afford to fail, a meat proxy might suffice. But for high-value engineering, the math changes.
The probability of an LLM "one-shotting" a complex, high-value task is remarkably low:
For critical work, the human must step up as the:
- QA Lead (Testing for edge cases)
- Engineering Manager (Overseeing direction)
- Architect (Ensuring long-term viability)
Without this, a poorly phrased prompt could lead to a fundamental architectural error that is incredibly costly to undo later. Furthermore, because LLMs increase throughput, the standard for what is shipped should actually increase. Instead of shipping a basic MVP, we should use agents to polish and explore edge cases—though they will fail to do this unless explicitly pushed.
The "Bazel" Case Study: Unknown Unknowns
Consider the complex task of converting a legacy system to build with Bazel. This isn't a simple translation; it's riddled with intangible requirements.
The "Meat Proxy" Checklist (What they think they're doing):
- Prompt: "Convert this to Bazel"
- Run command
- Fix syntax error
- Ship
The Actual Engineering Checklist (What is actually required):
- Identify why the code functions this way (hidden dependencies)
- Determine if a specific customer is relying on a quirk of the current build
- Negotiate with stakeholders to remove deprecated elements
- Manage "unknown unknowns" that the LLM cannot see
If you simply "walk away" after a prompt, you are ignoring the decision points that define professional engineering.
Out-of-Distribution (OOD) Failures
The gap between human intuition and AI becomes glaring when dealing with "out-of-distribution" data—things the AI hasn't seen enough of during training.
- Obscure Languages: Agents are significantly worse at niche programming languages compared to Python or JavaScript.
- Modern Board Games: In games like Dominion or Lost Cities, a state-of-the-art model is often worse than a human who has never played the game before but understands general gaming heuristics.
The Dominion Experiment
I witnessed a new player try to use ChatGPT to learn Dominion. The result? The AI was roughly 50% accurate. The dangerous part was that the "wrong" 50% sounded confident and steered the player toward worse strategies than if they had just guessed.
In contrast, a human spending just 30 minutes reading the rules could potentially reach the 99th percentile of players. This highlights the massive gap in how humans and agents approach novel problems.
Final Thoughts
If you find that agents are "nailing" your tasks with zero effort, you must ask yourself:
- Have I been coasting in my role already?
- Why am I not pushing the agents toward tasks they actually struggle with?
The goal is not to be the loop; it is to be the mind that directs the loop.