tech
Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling
It's the company's first public proof point after a year and a half spent building AI infrastructure largely out of public view.
TL;DR
- Thinking Machines Lab, founded by Mira Murati, has launched its first open-weight AI model called Inkling.
- Inkling is a mixture-of-experts system with 975 billion parameters, trained on diverse data types, and reasons across text, image, audio, and video, though outputs are currently text-based.
- The model is designed to be adaptable by organizations, allowing them to fine-tune it via the Tinker platform, contrasting with closed models from major AI labs.
- Inkling aims for well-rounded performance and calibrated answers, allowing users to adjust 'thinking effort' for speed-vs-accuracy trade-offs.
- The company's strategy is based on the belief that AI shaped by individual organizations will outperform centrally trained, fixed models.
- Industry figures like Microsoft CEO Satya Nadella and Hugging Face CEO Clem Delangue suggest a shift towards open-source or private alternatives for enterprise AI.
- A project with Bridgewater Associates reportedly showed an open-source model further trained on proprietary data outperforming top proprietary AI models at a fraction of the cost.
- Thinking Machines claims rapid development time for Inkling compared to competitors.
- The model used some distillation from other open-weight models during its pre-training phase, with future models planned to use fully self-contained post-training.
- The company's revenue is expected to come from its Tinker platform for training and fine-tuning, rather than direct model usage fees.
- Thinking Machines currently employs around 200 people and emphasizes a culture of continuity over individual personalities.