BYTEDANCE SKIPS AI SHORTCUT FOR 10T-PARAM MODEL

At an all-hands meeting last month, Yiming emphasized ByteDance's commitment to developing AI models through full training rather than leveraging distillation techniques that compress knowledge from larger models. The decision contrasts with industry shortcuts. Distillation typically reduces computational costs and speeds up deployment by transferring capabilities from large models to smaller ones. Meanwhile, ByteDance's massive 10-trillion-parameter model dwarfs Moonshot AI's Kimi K3 and approaches estimates for Anthropic's Mythos 5, signaling the company's intent to compete at scale in generative AI. The stance reflects ByteDance's resource advantage and confidence in its infrastructure, though it signals a longer development timeline. The approach prioritizes fundamental capability building over rapid iteration—a bet that raw model size and training quality will outpace competitors relying on efficiency shortcuts. Elsewhere in AI news, design platform Canva slashed revenue growth forecasts after heavy AI feature adoption drove costs up and adoption of ChatGPT among users slowed rollout timelines.