In a recent posted on the blog, the lab introduced an AI model construction project with consistent feedback and recycling.
The article is titled Fighting uncertainty in LLM reasoning ( LLM is a language model with the ability to grasp the language creation and other natural language processing tasks. LLM has this ability by learning statistical relationships from documents during self-supervision and semi-supervision training with high complexity). According to the community, this is inherently the nature of AI, but Thinking Machines considers it a problem that can be solved.
Horace He - a researcher at the laboratory, believes that randomness originates from the arrangement of the reasoning process. He argued that if the coordination layer (GPU) was strictly controlled, the model could be more decisive, thereby producing stable results.
In addition to helping businesses and scientists receive more reliable feedback, stable jobs also improve enhanced learning (RL). Because RL needs compliments for correct answers, but the data will be affected if the response is different. According to Thinking Machines, consistent feedback will help RL go smoother. The company also plans to use RL to customize AI models for businesses, according to sources from The Information.
Murati, former technology director of OpenAI, revealed that Thinking Machines' first product will be released in the coming months and will be useful for researchers and startups developing custom models. However, it is unclear what this product is and whether research on regenerative feedback will be applied.
The laboratory also affirmed that it will regularly publish blog posts, source codes and research documents to benefit the public and improve internal research culture. This approach is reminiscent of OpenAI's early days as it pursued open research, before becoming more confidential with large-scale development.
