QueensGambit

I got $10K GCP bill one month because of ClaudeBot. Its likely some users were trying to replicate our software by using Claude code + chrome plugin. If I was spending $10K on traffic, I can only imagine the money cloners must be spending on tokens.

Factorization is key here. It separates dataset-level structure from observation-level computation so the model doesn't waste capacity rediscovering structure.

I've been arguing the same for code generation. LLMs flatten parse trees into token sequences, then burn compute reconstructing hierarchy as hidden states. Graph transformers could be a good solution for both: https://manidoraisamy.com/ai-mother-tongue.html

Hi HN,

I asked GPT-5 to count the characters in this essay [1]. Here's what happened:

- JavaScript: correct in 1ms

- GPT-5 with Python tool: correct in 26 seconds

- GPT-5 pure reasoning: wrong after 3.2 minutes

So why can't we let AI reason in Python or JavaScript instead of English? In this essay, I explore this question through graph transformers that preserve code structure rather than flattening it into token sequences. Thoughts?

[1] https://github.com/ManiDoraisamy/devforever/blob/c41f69283a8...

From the paper: "sleep insufficiency was significantly associated with lower life expectancy when controlling for traditional predictors of mortality, with only smoking displaying a stronger association."