Tech giants like to boast trillions of AI models that require large, expensive GPU clusters. However, Fastino takes a different approach.
The Palo Alto-based startup says it has intentionally invented a new kind of AI model architecture that is small and task-specific. The model is very small and trained on low-end gaming GPUs totaling under $100,000, says Fastino.
This method has attracted attention. Fastino has secured a $17.5 million seed funding led by Khosla Ventures, the first venture investor in famous Openai, and Fastino tells TechCrunch exclusively.
This brings the startup’s total funding to nearly $25 million. Last November, it raised $7 million in the previous seed round led by Microsoft’s VC ARM M12 and Insight Partners.
“Our models are faster, more accurate and only a fraction of the cost to train while surpassing flagship models on specific tasks,” said Ash Lewis, CEO and co-founder of Fastino.
Fastino has built a small suite of models to sell to corporate customers. Each model focuses on specific tasks that a company may need, such as compiling sensitive data or summarizing corporate documents.
Fastino has yet to disclose early metrics or users, but its performance says it deals with early users. For example, because they are so small, the model can provide the entire response with a single token, Lewis told TechCrunch, flaunting the technology, giving detailed answers at once in a millisecond.
TechCrunch Events
Berkeley, California
|
June 5th
Book now
It’s still a little too early to decide whether Fastino’s approach will catch up. Enterprise AI spaces are busy, and companies like Cohere and Databricks are also promoting AI that are good at specific tasks. Enterprise-centric SATA model makers, including humanity and mistral, also offer small models. It is also no secret that the future of enterprise generation AI is likely to be a smaller, more focused language model.
It may take some time, but an early confidence vote from Khosla certainly doesn’t hurt. For now, Fastino says it is focused on building cutting-edge AI teams. It is aimed at top AI lab researchers who are not obsessed with building the biggest models or breaking benchmarks.
“Our recruitment strategy is very focused on researchers who have an inverse thinking process about how language models are currently being constructed,” says Lewis.