AI Programming Agent Crisis: George Hotz denounces "Eternal Waste September", academic papers confirm that structural constraints are attenuated by 30%

Core conclusion

On May 24-25, 2026, two pieces of news hit the top of the Hacker News list at the same time, pointing to a question that has been hotly debated for a long time: Can AI programming agents write real production code? **

• George Hotz (founder of tinygrad) published a long article "The Eternal Sloptember", saying that AI programming agents "will never really write code", they are just getting better and better at creating "waste" that is difficult to detect.
• On the same day, the arXiv paper "Constraint Decay" system tested 80 back-end generation tasks, proving that when structural constraints (ORM, database mapping, framework specifications) are superimposed, the pass rate of LLM Agent drops by an average of 30 percentage points, and weak configurations directly return to zero.

The two pieces of news are based on experience and the other is empirical research, but the conclusions are surprisingly consistent: **AI can help you quickly run through prototypes, but it is still far from being a reliable "production-level programmer". **

Key Points

• George Hotz asserts: Using AI Agent development will be the most costly mistake in the history of software engineering
• Academic research confirms: Agent's performance under complete constraints is 30+ points lower than that of "casual writing"
• Huge framework sensitivity: Flask task is close to full score, Django/FastAPI task plummets to zero
• Biggest pain point: Data layer defects (wrong query combination, ORM runtime violations) account for 60%+ of errors

Background: Two forces point to the same conclusion at the same time

George Hotz: "Eternal Scrap September

George Hotz started with his actual experience using AI to program Agent in the past 6 months. He tried using Agent to write code in tinygrad, and also tried to use Agent to reverse engineer USB chips. Each time, the Agent performed amazingly in the "prototyping stage" - quickly generating a large amount of code - but in the detail polishing stage, it became a "slot machine lever". You kept pulling it, and it occasionally gave the correct result, but never completely.

He specifically warned of the risks in large organizations: high-performing programmers can use experience to judge whether AI output is reliable, but lower-level employees lack this self-checking ability. Agent has increased the "output" of low-level employees by 10 times, but the code quality has declined sharply - a large amount of "waste" code is being merged into the production environment, and these wastes are becoming increasingly difficult to identify.

"When you see an artifact, you automatically assume that its creator possessed a (basically) normal human state of mind. This assumption is no longer true."

Academic paper "Constraint Decay

The paper co-authored by Francesco Dente, Dario Satriani, and Paolo Papotti designed a rigorous dual evaluation system: 80 new back-end generation tasks + 20 function implementation tasks, covering 8 Web frameworks. Each task is first tested with "loose specifications" (only functional requirements), and then tested with "full specifications" (plus ORM, database, and architectural constraints).

The results are shocking:

The paper names this phenomenon "Constraint Decay" - as structural constraints are superimposed, the performance of the Agent declines systematically rather than linearly.

Key Impact: Practical Implications for Developers and Enterprises

Adaptation suggestions

1. Quality gate must be placed in advance: Do not let the Agent directly generate the final code. Let it generate a constrained skeleton first, and then the developer can complete the key logic in the skeleton. This "human complement AI" model is more time-saving than "AI writing and human modification".

2. Choose a low-magic framework: If you must use Agent to generate back-end code, give priority to explicit frameworks such as Flask, and stay away from high-magic frameworks such as Django/Rails, which prefer convention over configuration. Framework selection may have a greater impact on Agent success rate than model selection.

3. Data layer manual review: The paper points out that ORM runtime violations are the largest source of errors. All Agent output involving database queries and mapping must be manually reviewed line by line.

4. Don’t treat Agent as a programmer: George Hotz’s advice is very straightforward - Agent is a smarter Google, a rapid prototyping tool, but not your colleague. Use it to explore options and then use your judgment to decide whether to adopt them.

Task List

• Review your current pipeline: What quality gates does the code produced by the Agent pass through?
• Check your framework choice: How does the agent perform under your current framework?
• Establish a manual review mechanism for the data layer, especially ORM related code
• Establish a decision tree for the team on "when to trust the Agent and when to have human review"

Reference sources

• George Hotz: The Eternal Sloptember — HN score 352
• Constraint Decay: The Fragility of LLM Agents in Back End Code Generation — HN score 245
• HN ：Eternal Sloptember
• HN ：Constraint Decay

Tool entry

This article covers the following tools: Claude Code, OpenAI, ChatGPT, DeepSeek, n8n, OpenClaw. The tools themselves are not the problem—the problem is how we understand their boundaries. Good tool users know when to trust the output and when to be skeptical.

Internal link guidance

• Want to add a quality gate to the AI Agent output? Watch: How to add quality gates to your AI automation workflow: A practical guide from output to trustworthy results
• Want to systematically learn the correct usage of AI Agent tools? Watch: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes
• Someone uses specification-driven development + AI code review to earn over 10,000 yuan a month: He earns over 10,000 per month by relying on AI code review + specification-driven development: a practical review of a freelance developer
• Real income of independent developers using Agent tools to automate workflows: Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo