E129: The Race to Help Build Custom AI Models

Synthetic Data

Synthetic data offers a unique advantage over traditional data cleaning methods by eliminating the need for user-provided data. Sahil Chaudhary explains that while data cleaning requires a good raw dataset, synthetic data only needs a well-defined use case, allowing users to design their dataset without any initial samples 1. This approach is not yet widely adopted, and Amanda Robson notes the challenge of convincing users of its benefits 2.

We manually onboard all of our customers just to get them to try the platform and show them that this can actually work.

--- Sahil Chaudhary

The focus remains on text-based models, with plans to expand to multimodal models in the future.

   

User Journey

The journey for users adopting synthetic data solutions often begins with those who have attempted to productionize models and faced challenges. Sahil Chaudhary identifies early adopters as those who have validated their product thesis and seek to own their language models 3. The onboarding process is manual and consultative, reflecting the high activation energy required to engage users in this new paradigm 4.

Right now, 100% of the onboarding to our platform is done manually.

--- Sahil Chaudhary

As the industry matures, the aim is to make the platform more self-serving, reducing the need for intensive onboarding.

   

LLMs & Synthetic Data

Synthetic data plays a crucial role in enhancing large language models (LLMs), addressing challenges like hallucinations and data diversity. Sahil Chaudhary notes that while synthetic data techniques have roots in traditional ML, they require adaptation for LLMs due to unique properties like hallucinations 5. Experiments have shown that synthetic data can significantly improve model quality over traditional methods, although issues like hallucinations and lack of diversity remain 6.

The outputs are significantly better than what we could get previously.

--- Sahil Chaudhary

The goal is to refine synthetic data processes to better cover all facets of knowledge required for specific use cases.