Korva is a scroll of the internet’s best ideas.
Creativity is not verifiable, and you can't put an easy number on creative ideas. When two models are used together, one trained on a different data set than the other, a person is more likely to get creative ideas than from either alone. If one model ever fulfilled everyone's desires, someone would build a different one trained on completely different data, and that creates demand for both.
Companies that train a model on their own data are surrounded by an ecosystem doing the same thing, and the data those other companies acquire is potentially valuable to them. Trying other models can cut costs, improve margins, or lift performance when models are merged together.
No one is going to win all of this market, so a company is better off building a model known for a specialty that matters to its business. Its brand is a big part of its moat, and the model becomes a way the brand carries around.
How can combining models improve results?
Together AI's 2024 Mixture-of-Agents paper reported that layering several open models, each refining the others' answers, topped GPT-4o on the AlpacaEval 2.0 benchmark. Models trained on different data make different mistakes, so their errors partly cancel out. Combining independent estimates is a long-standing idea in machine learning, known as ensembling.
How does merging models work?
Model merging combines the weights of several fine-tuned models that started from the same base model. A 2022 paper by Mitchell Wortsman and colleagues, called Model Soups, showed that averaging the weights of fine-tuned models could improve accuracy. The merged model runs at the same cost as a single one.
Why does training data shape a model's strengths?
A model learns patterns from what it reads, so a model trained heavily on code, legal text, or conversation will be strongest in those areas. Different companies hold different data, such as their own documents, customer interactions, or licensed archives. That difference is a reason models trained by different companies end up with different specialties.
More from 20VC with Harry Stebbings
