Keep pulling the thread on Jan Stoica.
Elamarena, a company co-founded by Jan Stoica to help other companies with LLM evaluations, has raised $100 million.
Jan Stoica states that the best open-source large language models are currently being developed in China.
Jan Stoica warns that the United States faces a structural disadvantage in AI development compared to China due to the siloed nature of its research and the limited role of academia.
Jan Stoica believes it is "very likely" that the current massive build-out of AI data center infrastructure by hyperscalers will result in overbuilding, similar to what occurred during the dot-com bubble.
Jan Stoica predicts that China will successfully build out its own domestic AI compute infrastructure within the next few years, overcoming current limitations.
The United States has lost its leadership position to China in multiple high-tech industries including solar cells, car batteries, drones, electric cars, and industrial robotics.
Jan Stoica stated that viable alternatives to NVIDIA's dominance in AI hardware have not emerged as quickly as he initially expected over the last year.
Jan Stoica was surprised that the most significant progress in open-source models over the past year came from China rather than the US.
Jan Stoica predicts that China, and specifically Huawei, will invest heavily in developing alternatives to NVIDIA's GPUs because U.S. export controls leave them with no other choice.
While competitors like Google (TPUs), Amazon Web Services (Trainium), and AMD have developed very good hardware, their biggest challenge in competing with NVIDIA remains the software stack.
The Vicunia model was created by fine-tuning the first version of Meta's Llama model using data from ShareGPT.
The "LLM as a judge" evaluation technique was pioneered by a team at UC Berkeley using GPT-4 to evaluate the Vicunia model.