# Right Now is the Worst Stage for AI — And Why That’s Okay
## AI Coding Tools Dominate Usage
If you check the **Top 10 token consumption rankings** for OpenAI or Anthropic, you'll see code-related tools consistently at the top. **AI coding has taken off**, and both **developers** and **non-technical users** are driving this irreversible trend.
For outsiders, the magic is easy to feel:
> “I just wanted to create an alien terminal... and it actually ran.”
Inside companies, however, the reality is stark:
> “AI cuts the first 10% of work down to 1%, but the remaining 99% is still the bottleneck.”
We may be in **the worst AI stage**, but from here, things can only improve.
---
## The Problem: LLMs Aren’t Really Thinking
From Antithesis CEO **Will Wilson**:
- LLMs are **memory giants**, not brilliant thinkers.
- They excel at mainstream, familiar tasks (React demos, popular frameworks).
- They struggle with **novel or niche problems**.
- They generate fast **0-to-1 prototypes**, but fail at logical, large-scale modifications.
- They rely on statistical similarity — **not reasoning**.
### Fast-Food-Style Code: The “Tragedy of the Commons” Risk
Without safeguards, AI-generated quick fixes can degrade long-term quality:
- If one person submits sloppy AI code for others to fix, it’s fine.
- If *everyone* does it, **system stability collapses**.
---
## Hidden Risk: Team Skill Decay
Over-reliance on AI atrophies developer “muscle memory”:
- Coding feel fades after weeks without hands-on coding.
- Like languages: after a year in Python, you forget C.
- Problem-solving capability shrinks if unused.
---
## How to Integrate AI *Without* Slowing Teams
Wilson’s two key accelerators:
1. **Documentation** – Feeds better context to LLMs and sustains shared code intent.
2. **Testing** – Strong architecture + robust tests maximize AI benefit.
> Conservative, test-heavy teams gain the most acceleration from AI.
---
### Business Tip: Don’t Over-Build Internal AI Platforms
Field CTO **Akshay Shah**:
- Enterprises shouldn’t rush to create complex AI automation pipelines.
- Instead, **give every employee their own Claude/Gemini subscription**.
- Let *individuals* integrate AI into their workflow for best efficiency.
---
## Testing: The AI Code Safety Net
Wilson and Shah emphasize **property-based testing**:
- Traditional: multiple scenario-specific unit tests.
- Property-based: **rules as tests**.
- Example: *“The defense robot must never kill its creator under any circumstances.”*
This transforms tests into **living documentation** that prevents misunderstanding and drift.
---
## Reviving Legacy Systems: The Perpetual Beta Era
AI + property-based tests excel at **migrations**:
- “Re-implement this code in another language / framework with identical behavior.”
- Rules ensure edge cases are covered.
**Legacy code = code without tests** — the same truth now, faster.
---
## Vibe Coding: Lowering the Barrier, Raising New Risks
**Pros**
- Accessible for non-programmers.
- Encourages experimentation and iteration.
**Cons**
- Creates “comprehension debt” — code in repos the team doesn’t truly understand.
- AI accelerates technical debt if unchecked.
---
## Locked in 2024 Paradigms?
LLMs favor dominant frameworks (React over Svelte) due to training data bias.
- New languages (like **Zig**) suffer from poor AI support.
- Risk: stagnation in toolsets and approaches.
---
## Rule #1: You’re the Boss, AI Is Not
**Best practice**:
- Decide your direction before AI starts coding.
- Use AI as **idea expander** or **options generator**, not leader.
- Keep **human judgment** central.
AI should be your **bicycle for the mind** — amplifier, not replacement.
---
## Advice for New Engineers
Two paths:
1. **Deep technical**: OS, CPU, runtimes — hard problems AI can’t solve yet.
2. **Domain-specialized**: Pair solid programming with deep industry knowledge.
Both require blending human expertise with AI tools.
---
## Creative & Business Integration
Platforms like **[AiToEarn](https://aitoearn.ai/)** help creators:
- AI content generation
- Cross-platform publishing (Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, Twitter/X)
- Analytics and AI model ranking
Resources:
- [AiToEarn官网](https://aitoearn.ai/)
- [AiToEarn博客](https://blog.aitoearn.ai)
- [AiToEarn开源地址](https://github.com/yikart/AiToEarn)
- [AiToEarn文档](https://docs.aitoearn.ai/)
- [AI模型排名](https://rank.aitoearn.ai)
---
## Key Takeaways
- **AI accelerates early dev but magnifies testing/documentation importance.**
- **Skill preservation** is critical — avoid full reliance on AI-generated code.
- **Property-based testing** is returning as a safety net.
- Keep **human leadership** in coding decisions.
- Decision bias from training data could lock tech in present paradigms.
- For sustainable AI adoption: documentation + tests + human control.
---
Would you like me to prepare a **one-page executive summary** with these points for quick team briefings?
