Former ByteDance Seed Engineer: ByteDance Takes Six Months for One Iteration, Google Allegedly Only Three Months

According to monitoring by Dongcha Beating, Zhang Chi, a former engineer from ByteDance’s Seed team and now an assistant professor at Peking University, revealed in the podcast “Into Asia” that it takes about six months for ByteDance to complete one round of large model training (pre-training plus post-training), while Google is rumored to only need three months. He believes that the speed of iteration is one of the core reasons why Chinese companies find it difficult to catch up. Zhang worked at ByteDance for about a year, and his math team’s focus was more research-oriented. He described the team’s positioning as ‘more for publicity,’ which differs from the teams responsible for model delivery in pre-training and post-training. Zhang described the internal ‘benchmaxxing’ culture at Seed: team leaders evaluate performance based on the benchmarks they are responsible for, and everyone is focused on scoring, ‘but this does not translate into a good experience in actual use.’ He stated that on paper, the models of large Chinese companies can match the leading models in the U.S., but in practice, they are ‘not good enough.’ The goal of Seed is to be globally top-notch, ‘but unfortunately, I do not think we have caught up,’ and even the goal of being number one domestically ‘has not been achieved.’ By the end of 2024, Seed believed it had caught up with GPT-4o, but after the release of DeepSeek, the team realized that the gap still existed, and when he joined, the entire group was urgently shifting towards reinforcement learning.