Keep pulling the thread on Edwin Chen.
Human data startup Surge surpassed $1 billion in revenue in the last year.
Surge CEO Edwin Chen reports that some AI researchers have told him their VPs require them to prioritize improving their model's rank on the LMSYS Chatbot Arena, even at the cost of degrading performance in factuality and instruction following.
Surge CEO Edwin Chen describes the LMSYS Chatbot Arena as a "giant plague on AI" because its evaluation method incentivizes models to be superficially appealing rather than factually accurate or good at following instructions.
According to Surge CEO Edwin Chen, a thousand pieces of high-quality human data can be more valuable for model training than 10 million pieces of synthetic data, of which 99% might be useless.
Surge plans to begin publishing external research to educate the public on the comparative capabilities of different AI models, including strengths in coding and instruction following, and weaknesses like hallucination rates.
Surge CEO Edwin Chen predicts that human feedback for training AI models will never become obsolete, even as models achieve superhuman performance.
Surge CEO Edwin Chen believes there is "no ceiling" on the useful diversity and richness of reinforcement learning environments, as more complexity allows models to learn more over longer time horizons.
Surge CEO Edwin Chen believes many founders in Silicon Valley, including some from Y Combinator, raise money primarily for social validation, like getting a headline on TechCrunch, rather than to solve a specific problem.
Surge was founded in 2020 and has been bootstrapped since its inception.
Surge CEO Edwin Chen has observed a top-tier frontier language model that randomly outputs Hindi and Russian characters in the middle of its English responses approximately 10% of the time.
Surge has found that the easiest way to improve a model's rank on the LMSYS Chatbot Arena is to increase the length of its responses.
Surge CEO Edwin Chen characterizes many competing data labeling companies as "body shops" that deliver human labor without underlying technology, in contrast to Surge's focus on a technology platform for quality control.