TLDR: we used DSPy-style recursive prompt optimization to train a browser agent that solves Brett Adcock's 30-step challenge in 265 seconds at $0.06 per run, with GPT-OSS 120B matching frontier performance after training. 2 weeks ago, Brett Adcock posted a public browser agent challenge; the website
TLDR: reinforcement learning cut our monthly inference costs by 98.32%, improved performance on our internal benchmark, and reduced latency by 66%. Freesolo builds people search infrastructure, which involves indexing hundreds of millions of LinkedIn profiles and training models to first query and t
Originally when Freesolo was first started when it was still called Linkd, there was one database for each school with approximately 10k profiles per school. That means that there is a lot of room for inefficiency. However, when we decided to build people search for the entire world and started work