Repost: Yage's Hot Takes on Overhyped AI Coding Metrics

Why Repost This

This comes from the Trial and Iteration channel in Superlinear Academy, a community created by Kedaibiao (课代表立正), a Chinese YouTuber focused on AI education.

Yage (Yan Wang, aka 鸭哥) is described by Kedaibiao as the strongest coder he has ever met – a frontier engineer with deep hands-on experience in AI Agents and AI Coding.

About Yage

Yage runs his own blog at yage.ai, covering Agentic AI and AI Coding engineering practices. Every article is packed with insights. Recent highlights include:

• Key Decisions for Agentic Workflows – A real case of using AI to add SEO summaries to 300 blog posts in 2 minutes, demonstrating five critical engineering decisions
• OpenClaw Deep Dive – In-depth analysis of why OpenClaw went viral, plus his own alternative engineering approach
• Beyond Tutorial Thinking – Analyzing four dropout stages for AI learners, proposing engineered platforms to eliminate friction

Each article also contains valuable links and open-source project references well worth exploring. Highly recommended.

Original Content

Below is Yage’s full post from the Superlinear Academy community:

There are quite a few trending concepts in AI coding lately. Here are my hot takes:

1. How many hours AI can work continuously doesn’t matter

Because speed varies wildly. Opus often takes half an hour on something that Composer 1.5 knocks out in two minutes, with the same quality. Opus is just slower and overthinks unnecessarily. (Caveat: there are counterexamples – maybe 5% of the time, what Opus produces in 30 minutes is genuinely better than what Composer 1.5 does in 2 minutes.)

2. How many tokens you burn doesn’t matter

This is because plugins like Oh My OpenCode have a tendency to over-engineer – spinning up multiple background agents in parallel even for simple problems. I once burned through Antigravity’s entire week of Claude quota with a single-sentence prompt. (Caveat: the order of magnitude of token usage does have reference value – many lessons can’t be learned without burning through a certain threshold.)

3. Using multi-agent teams to work

My feeling is this isn’t very useful. It helps somewhat for breadth-first research tasks, but for work that demands depth, or where the bottleneck isn’t coding speed but rather the human element (which is most work), it’s largely useless.

Conclusion

So my stance on these trending concepts is: these metrics have good intentions, but they’re too easily gamed. Look at the evidence. What really matters is how many products you’ve shipped, how much growth and profit you’ve generated. Not “I use agent teams and you don’t, so I’m better.” Or “I burned 500 million tokens today and you only burned 300 million, so I’m better.” These can be references, but there’s no need to treat them as standalone metrics to chase or flex.

Source: Superlinear Academy - Trial and Iteration · Author: Yan Wang (鸭哥)

Related Links:

• Kedaibiao YouTube Channel
• Yage’s Blog - yage.ai