
Autonomous AI Research on a Laptop
Can a full research loop run on a MacBook? Cong Lu's Local Autoresearch lets Flint, a local 40B model from Recursive, improve a small program one experiment at a time: it proposes a change, writes the code, a runner tests it, and only improvements are kept.
We walk through the three examples Cong Lu shared: circle packing (sum of radii up 21.4% over a grid baseline), factory scheduling with 40 jobs on 20 machines (finishing time down 26.0%), and bike-rental demand prediction on the public UCI dataset (held-out error down 9.08%). Then the eight-worker setup: eight agents share one Flint model…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.