Yesterday, it was created by several former executives of OpenAi Thinking Machines Lab(TML) Finally, six months after its inception, its first heavy research paper, Defeating Nondeterminism in LLM Reference (overcoming uncertainty in LLM reasoning)。
According to TML, in the future, blog articles, open source codes and research results will be published frequently, both for the benefit of the public and for the purpose of " optimizing our own research culture". The first of these papers is the opening of a new blog series, Conneactionism。
Before that, you often saw that ChatGPT and others like AI give different answers each time, and this phenomenon has long been seen as common -- and everyone thinks the current AI model is a probability model. And the TML is trying to solve this big problem with this outcome。
IN TERMS OF THE CONTENT OF THE PAPER, THE TML BELIEVES THAT THE AI MODEL IS THE CULPRIT OF RANDOMITY, ACTUALLY THE GPU KERNEL -- THE LITTLE PROGRAMS THAT RUN IN THE BRITISH CHIP -- AND THE WAY IN WHICH THE PUZZLE IS DONE IS PROBLEMATIC。
The core author of the article is the AI Researcher of Thinking Machines Lab, Horace He, who has a very clever solution, and who has the potential to make the output of the AI model more stable and reliable if he can accurately control the implementation process on this level。
In addition, in addition to providing a more scalable AI response, there is a greater value in enabling the AI model to generate repeatable answers — a significant improvement in the effectiveness of enhanced learning (RL) training. Horace He believes that a more consistent response would make the entire RL training process "sliding" a lot。
Mira Murati is known to be the former CTO of OpenAI and currently the founder and CEO of Thinking Machines Lab. In July, she stated that Thinking Machines Lab had completed $2 billion in financing and that the first product would be visible in the coming months, and would be of great benefit to researchers and start-ups that were developing customized models。
ACCORDING TO THE LATEST NEWS, THE TML VALUATION HAS BEEN STRUNG AT $12 BILLION (EQUIVALENT TO 85 BILLION YUAN) IN THE CASE OF "0-PRODUCT-0 USERS"。