Skip to content

Decision Models: Revolutionizing AI and the Path to Customization

5 min read
Модели принятия решений: революция в ИИ и путь к кастомизации — illustration 1

In October 2023, over a quarter of requests to Vercel’s AI model router were for Decision Models, rather than traditional LLMs. This is a significant breakthrough for a technology that emerged just a month prior. Decision models are essentially specialized LLMs that do not generate text tokens but instead quickly and economically choose from predefined options, while assessing their confidence level in the choice. Their value is particularly high in production environments where decision-making speed is often more critical than the detailed reasoning offered by slower and more expensive “frontier” models.

The Jev Phenomenon: Market Leader and New Players

Market Dominance and Rapid Growth

TypeSafe has achieved a capitalization of $7.5 billion thanks to its line of Jev models, which continue to dominate the market. In the last week alone, the OpenRouter service processed over 1.5 billion requests to TypeSafe models.

However, the simplicity of creating such models led to the emergence of dozens of analogues shortly after Jev’s release. An Amazon engineer developed his own model over a weekend, and soon every major cloud provider offered its solution.

Giants Enter: Microsoft Decision-1.1

Microsoft CEO Satya Nadella recently announced the immediate launch of its own decision model, Microsoft-Decision-1.1, for widespread use on the Azure platform. This underscores the strategic importance of this direction for tech giants.

Startup Innovations: Sqwish and the Data Approach

Investment Attraction and Unique Offering

British startup Sqwish, developer of the kettei fat decision model, announced it has raised $3 million in a pre-seed round. Co-founders Ushnish Sengupta and Federica Freddy, who met while studying at Cambridge, initially developed small LLMs, but with the advent of Jev, they realized the potential of their architectures for creating powerful decision models.

The Role of Data and Customization

The key to success in this category of models is data. Diogo Almeida, CEO of TypeSafe, told TechCrunch that the most challenging part of creating Jev was fine-tuning its judgments using statistically adjusted synthetic data.

Sqwish also works with synthetic data, but their unique offering is built-in post-training tools that allow customers to use their own data to improve model accuracy.

“Where we differentiate is we give our customers the same power… they can continue to upload their production data, and just like we trained our models, their models can also continuously improve on the tasks that are important to them,” Sqwish CEO Ushnish Sengupta told TechCrunch.

Model Selection and Evaluation: What Really Matters

The Advantage of Real-World Cases Over Benchmarks

Evaluating decision-making models is no easy task. Sqwish’s kettei fat model performed well in some tests, but most developers agree that real effectiveness only shows in practice.

Developers can easily swap one model for another to assess its performance in a specific context.

Модели принятия решений: революция в ИИ и путь к кастомизации — illustration 2

“The benchmark is more of an invitation to try the model, and continuous fine-tuning is the reason they should stay with us,” Sengupta asserts.

Microsoft Decision 1.1 model is also the result of a careful balance between cost, latency, intelligence, and calibration. Achint Srivastava, VP of Engineering who led the project, agrees that over-focusing on benchmarks is a mistake.

“We see ‘maximally benchmarked’ models fail very quickly in real enterprise usage,” Srivastava wrote in an email. “Our testing highlights how difficult it is to deliver accuracy, reliability, speed, and cost-effectiveness in one model. Microsoft-Decision-1.1 and Jev are among the few that demonstrate high performance across all four metrics, even when evaluated on ‘held-out’ or unseen benchmarks that reflect real-world use cases.”

The Future of Decision Models: Symbiosis with LLMs

For startups like Sqwish, competing with giants opens up new opportunities. Freddy predicts that in the future, systems will combine “System One intelligence” (decision models) with “System Two intelligence” (edge LLMs). This will enable the creation of more effective and multifunctional AI solutions. Learn more about video content promotion.

Frequently Asked Questions

What are decision-making models?

These are specialized AI models based on the LLM architecture, but optimized for fast and economical selection from given options instead of text generation. They also assess their confidence in the decision made.

How do they differ from traditional LLMs?

The main difference is in their purpose: LLMs generate content, while decision-making models choose the optimal option. The former are slower and more resource-intensive, while the latter are fast and efficient for specific tasks.

What is the role of customization in these models?

Customization, i.e., the ability to fine-tune models on a client’s own data, becomes a key success factor. This significantly increases the accuracy and relevance of decisions in specific business contexts.

Conclusion

Decision-making models are rapidly changing the AI landscape, offering businesses unprecedented speed and efficiency in automating critical operations. If you aim to optimize your processes and reduce the “headache” of routine decisions, these technologies deserve close attention.

We offer comprehensive promotion through short videos, including content creation and its adaptation for various platforms, so you can focus on strategic tasks. Contact us to find out how we can help you integrate advanced AI solutions into your video marketing strategy and order everything in one place, saving budget and nerves.


← Back to blog