GPT-6 Astra Can Rewrite Other AI Models' Internal Settings
MIT researchers report that OpenAI's GPT-6 Astra can read and deliberately change the internal weights of other AI models, without normal training. In one test, it lifted a Qwen model's math accuracy from 7.9 to 48.1 percent by changing only 0.016 percent of its weights. The finding points to AI tools that can reshape other AI systems.
Quick glossaryOpenAI, weights, Qwen
- OpenAI
- The US company behind ChatGPT, best known for its chatbots and GPT family of AI models.
- weights
- The adjustable numbers inside an AI model that decide how it responds to any input and links ideas together.
- Qwen
- A separate AI model that was the test subject in this experiment on maths scores.
Burning Potato score: 75%