The AI Revolution’s New Math: Customization vs. One-Size-Fits-All

The latest development in artificial intelligence is a significant shift towards customization, rather than relying on one-size-fits-all models. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released its first open model, Inkling, which marks a new approach to AI development. By making this model available for outside developers and companies to modify directly, Thinking Machines is challenging the conventional wisdom that centralized, proprietary AI systems are the only way forward.

Inkling uses a fraction of its 975 billion parameters for any given task, making it faster and cheaper to run than larger counterparts. Training on 45 trillion tokens of text, image, audio, and video data has allowed Thinking Machines to create an AI that reasons natively across all four domains, with outputs limited to text, including code, styled artifacts, and structured data.

Unlike OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini, which have opted for a more centralized approach, Thinking Machines is betting on customization. This means organizations can adapt Inkling to suit their specific needs without relying on a single, all-encompassing model. Companies like Bridgewater Associates are already seeing the benefits of this approach.

Bridgewater’s project with Thinking Machines was particularly notable. Researchers from both companies took an existing open-source model and trained it further on Bridgewater’s own financial expertise. The result was impressive: the custom-trained model scored 84.7% on financial reasoning tests, beating top proprietary AI models while costing roughly a fourteenth as much to run.

This development raises important questions about the economics of AI. Thinking Machines’ efficiency-driven approach may prove to be a game-changer in the long run, especially if it can replicate its results with other organizations. The company’s willingness to use pre-trained models and distillate early post-training data from competitors’ outputs has also sparked debate.

Some have questioned whether this is just a way for Thinking Machines to sidestep the costs of training large models from scratch. However, the company’s own materials suggest that it used other open-weight models to help generate some of its early post-training data before large-scale reinforcement learning took over. The next model will use fully self-contained post-training instead.

One thing is clear: Thinking Machines’ approach marks a significant departure from traditional one-size-fits-all models. It’s an acknowledgment that AI development is not just about creating powerful models, but also making them useful to organizations in their specific contexts. This shift towards customization may ultimately prove more effective and efficient than relying on centralized systems.

As the AI revolution continues to unfold, it will be interesting to see how Thinking Machines’ approach compares with its competitors. Will its efficiency-driven approach pay off in the long run? Can it replicate its results with other organizations? These are questions that only time will answer.