Recent
Pi Coding Agent: N=1 Comparison of App Building Behavior in InklingSmall and GLM-5.3-Flash
GLM-5.3-Flash is a very good and cheap model -- how does it compare to InklingSmall a similarly priced model from Thinking Machines? Turns out the behavior is really hard to compare at N=1, even if the small scale makes human interpretation easier.
A Brief Review of Behavior of InklingSmall for Data Science
A review of a InklingSmall's performance on a recent visualization task I gave it.
Automated Alignment Researchers Can Reliably Mitigate Alignment Failures
Followup by Anthropic on Weak-to-Strong Automated Researchers (May 2026) and Building and evaluating alignment auditing agents (July 2025). Can Claude autoresearchers solve alignment issues on their own? Sort of yes! On 10 different misaligned behaviors a Claude autoresearchers successfully reduced the incidence of said behavior to varying degrees. The real story is one of efficiency though -- the agents did considerably better than humans in an 8 hour period, largely by finding and implementing literature-based solutions to misaligned behavior faster than the humans good. That makes automated researchers a useful tool, but not a general solution for alignment. Worth a read!