Key part of ByteDance all-hands meeting: They see a widening gap with US models and are focused on competing long-term in AGI foundational capabilities rather than near-term shortcuts like distillation. “Regarding the widening gap between ByteDance’s large language models and overseas state-of-the-art (SOTA) models, Tan Dai said that the most important thing right now is to get the fundamentals right, so that over the long term the company can enter the top tier without taking shortcuts. Liang Rubo added that the most important thing is not to let their actions become distorted by short-term pressures. Mentally, the team needs to accept that its large language models may lag behind for a period of time, while continuing to develop the technology in-house and optimize for the long term. “Yesterday I saw outside reports saying that we are insisting on developing our models in-house because of external pressure. That’s pure speculation. The real reason is that we want the team to delay gratification and build a strong technical foundation. That is critical to achieving AGI over the long term.”