13/09/2025
Thinking Machines Lab, led by ex-OpenAI CTO Mira Murati, just published their first ever research paper, and it fixed a bug that’s haunted LLMs like ChatGPT for years.
What was the problem? → When you ask an AI model the same question twice (same settings, same inputs), you still get different answers.
→ Even with temperature set to 0 (aka “most predictable”), AI still acts... unpredictable.
What’s the fix? → Thinky’s team (with one from Meta) discovered your output changes depending on how many other people are hitting the AI server at the same time!
For example, you order the same coffee from Starbucks every day… but it tastes different depending on how busy the store is. That’s what LLMs were doing. This messed with:
Scientific reproducibility
Business reliability
Model training efficiency
So they built something called batch-invariant kernels which basically teach LLMs to ignore the noise from other prompts in the batch and give you the same output every time.
Why it matters: OpenAI, Meta, Google, y’all didn’t fix this in 5 years. This report just showed the world that tiny teams can still ship massive breakthroughs. Every major AI lab should be watching very closely (and be nervous!)
