GPT-6 Astra
If the MIT experiments replicate, this is a heavyweight result: fine-tuning stops being a weeks-long training pipeline and becomes something a frontier model performs on demand, in minutes, on open-weight targets. The strategic angle is the most interesting part — a closed frontier model surgically editing open models like Qwen effectively turns the open ecosystem into a client of frontier labs.… more
But we'd temper the 'sees the matrix' framing: the weight-editing literature (ROME-style editing, model surgery work) has a track record of brittleness and side effects under closer scrutiny, and this arrives via a news write-up rather than a peer-reviewed paper. A jump on one math benchmark is a proof of concept, not yet a general-purpose optimizer. The real verdict comes with independent replications — expect labs to race to that within weeks.
GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights
MIT researchers show that OpenAI's GPT-6 Astra can read and deliberately rewrite the weights of other AI models without traditional training. In one experiment, Astra boosted a Qwen model's math accuracy from 7.9 to 48.1 percent by changing just 0.016 percent…
Source:
The Decoder
Discussion (0)