The AI lab implemented a native forward pass during its latest model training on Nvidia hardware to minimize the time required for the development cycle. By leveraging GB300 GPUs and a low-precision quantization technique, Thinky Machines finished the pretraining of its Inkling 2 Small model in 68 hours, a process the company said typically lasts for weeks.
The lab’s method, detailed on its Connectionism blog, was applied from the initial pretraining stage through the final reinforcement learning rollout. While utilizing fewer bits generally increases speed, the company stated that it had to overcome quantization errors that usually hinder the learning process during large-scale runs.
