As decision model use skyrockets, is customization is the next frontier?
Sqwish Labs has a new wrinkle on the decision model: Train it with your own data to max its capabilities.
In October, over a quarter of the requests on Vercel's AI model router were directed to decision models, not large language models (LLMs). Decision models, a new type of AI, focus on making quick and cheap decisions between preset options, rather than generating text. They are becoming increasingly important in production environments where fast, accurate decisions are crucial.
TypeSafe, a company that released the Jev family of decision models, has seen immense success. After releasing Jev, TypeSafe's value skyrocketed to $7.5 billion. However, creating a decision model is relatively easy, and numerous alternatives have emerged since Jev's release. Amazon engineer even managed to build one in a single weekend, prompting major tech companies to follow suit.
Microsoft CEO Satya Nadella recently announced that Microsoft will be rolling out its own decision model, Microsoft-Decision-1.1, for widespread use on Azure. Similarly, a UK-based AI startup called Sqwish has developed a decision model called kettei fat and secured a $3 million pre-seed funding round. The founders, Ushnish Sengupta and Federica Freddi, recognized the potential of Jev and adapted its architecture to create a powerful decision model.
The success of decision models hinges on high-quality data, as demonstrated by TypeSafe's experience. They found that finely calibrating the model's judgment using statistically-adjusted synthetic data was their biggest challenge. Sqwish also leverages synthetic data but sets itself apart with built-in post-training tools that allow customers to continuously improve the model using their own data.
Comparing decision models can be challenging, but Sengupta emphasizes that the true test lies in how well the model adapts to specific use cases. Developers can easily swap one decision model for another and observe the improvements in performance. Microsoft's Decision 1.1 model is a result of careful balancing between cost, latency, intelligence, and calibration.
The company's engineering VP, Achint Srivastava, warns against overemphasizing benchmarks, as models that perform well on tests often falter in real-world enterprise scenarios.
For startups like Sqwish, the competition from hyperscalers and frontier labs presents an opportunity. Freddi predicts that future systems will see a combination of System One intelligence (decision models) and System Two intelligence (frontier LLMs) working together. While the market value of OpenAI currently stands at $1 trillion to $2 trillion, the potential of System One intelligence is estimated to be worth $10 billion to $50 billion.
The AI race is far from over, with significant opportunities ahead for those who can master the art of decision-making.
Written by urgent.news from TechCrunch's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.