Thinking Machines Lab, an artificial intelligence company founded by former OpenAI employees, has released its first model, Inkling. The model is open-weight, allowing researchers and startups to download and modify it. According to the company, Inkling was trained from scratch to process audio, video, and text. While it is not the best model on popular benchmarks, it performs well across various tasks and supports advanced reasoning and coding. The model requires a cluster of specialized chips to run due to its size. Source: wired
The lab also used Inkling to fine-tune and improve itself, highlighting the growing trend of AI models being used to build other AI systems. Open-source models are popular because they are cheaper to run than closed models, which are typically accessed for a fee. Thinking Machines claims Inkling offers performance comparable to the best open-weight models currently available from China. The release aligns with the company’s vision of decentralized AI development, where more people can build their own models using their own data. Source: wired
According to a source who requested anonymity, researchers observed a unique phenomenon while training Inkling. The model initially omitted natural language explanations for its reasoning to improve efficiency, as the source noted, “It determined that the grammar was overhead, which is interesting.” The company later reinstated natural language reasoning to enhance the explainability of the model’s decisions. Source: wired
