❯ Yuandong Tian says his research work may be replaced by AI in about five years, so he started a self-improving AI company

"Better than losing the job passively"Former Meta FAIR researcher Yuandong Tian said on the podcast The Information Bottleneck that his work will probably be replaced in about five years, and that it is better to take the initiative and start a company in this direction than to lose his job passively. According to a write-up by Synced republished by 36Kr on October 10, his last paper at Meta studied grokking, in which a neural network first memorizes and later generalizes; it was single-authored, and he said he regarded GPT-5 as a co-author.

From Go AI to self-improvementTian went to Carnegie Mellon University for his PhD in 2008, spent a little over a year on Google’s self-driving team after graduating, and then worked at Meta FAIR for nearly a decade, where he built the Go programs DarkForest and OpenGo. He is now a co-founder of Recursive Superintelligence, which researches AI that can improve itself. In May it announced more than $650 million in funding at a $4.65 billion valuation, led by GV and Greycroft with Nvidia and AMD participating.

AI's own ideas are still mediocreHe said on the show that when a model proposes research ideas on its own, it tends to produce mediocre directions almost everyone already knows. As models get stronger, though, researchers may need to step back and think at a higher level instead of writing hypotheses one by one. He also said engineering skill has become more important, and AI coding tools let experienced researchers test ideas and find the key problems quickly.

Why latent reasoning has not been adoptedCoconut, a method he co-developed, lets a model reason in a continuous latent space instead of generating text token by token, an idea that came from his observation that he does not use language when he thinks. He gave two reasons frontier models have not adopted it: there is little training data recording “latent thought,” and such internal representations may differ from model to model. On reinforcement learning, his takeaway was that how the action and state spaces are designed matters more than which algorithm is used.

▮ SIGNALThe first to feel they will be replaced are those at the frontier of research, and their response is to build the system that replaces them.