๐ฌ๐ผ๐๐ฟ ๐๐ ๐ช๐ฟ๐ถ๐๐ฒ๐ ๐๐๐ป๐ธ, ๐ฎ๐ป๐ฑ ๐ฌ๐ผ๐ ๐ฃ๐ฎ๐ ๐ณ๐ผ๐ฟ ๐๐ ๐ง๐๐ถ๐ฐ๐ฒ
AI models charge you by the token. This creates a massive problem.
Last week, an AI agent wrote ten tests for a function. Two of them were useless. They re-checked things the code already guaranteed. They added zero value.
I paid for the tokens to create them. Then, I paid more tokens to keep them in the chat history for every future turn. The model had no reason to stop. The company billing me had no reason to stop.
The issue is simple: We pay for what AI produces, not for the value it delivers.
This misalignment creates several hidden costs:
- The Overthinking Tax: Smaller models often use massive amounts of tokens to mimic complex reasoning. One benchmark showed a model using 26 times more tokens to reach similar accuracy.
- The Verbosity Trap: Since models charge per output token, they have an incentive to say more. We even build tools to make agents talk like cavemen just to save money.
- The Feedback Loop: When an agent writes both the code and the tests, the tests often just repeat the bugs found in the code. The tests pass, but the logic remains broken.
This pricing model changes how you build software. You start capping history, caching aggressively, and routing tasks to smaller models just to manage costs. You are no longer designing for quality. You are designing for the invoice.
How do we fix this? I propose three rules for fair AI pricing:
- Charge only if the code compiles.
- Charge only if the tests pass.
- Cut the price if the agent ignores your specific instructions.
Real value lives at the end of the process, not at the start. Current pricing only measures the very beginning.
To fix this, we need an external referee. A verifier cannot ask the AI how well it did. It must watch the boundary. It must check the build, the tests, and the architecture independently.
A green light is cheap to produce but expensive to verify. Stop trusting the self-reported meter.
Would you pay per outcome for generated code? What signal would you trust to verify it?
Source: https://dev.to/tibtof/your-ai-writes-junk-and-you-pay-for-it-twice-185c
Optional learning community: https://t.me/GyaanSetuAi