Tech companies like to be proud of the trillion parameter AI models, which require large and expensive GPU clusters. But Fasteno Taking a different approach.
Startup, based in Paulo Alto, says it has invented a new type of AI model architecture that is deliberately smaller and related to work. Fasteino says the models are so small that they are at a low -finally trained GPU, which is worth less than 100,000.
The method is attracting attention. Fasteno, headed by Fasla Ventures, has received funds of 5.7.5 million, the famous Openi’s first venture investor, Fasteino, especially told Tech Crunch.
This reduces the total startup funding to about $ 25 million. It raised $ 7 million last November, headed by Microsoft’s VC Arm M -12 and Insight partners.
“Our models are faster, more accurate and spend a part of training by performing well on specific tasks,” says Ash Lewis, co -founder and co -founder, Ash Lewis.
Fasteino has created a sweet model that sells it to the enterprise consumers. Each model is focused on a specific task that the company needs, such as repeating sensitive data or summarizing corporate documents.
Fasteno is not yet disclosing the initial matrix or consumers, but says its performance is producing early users. For example, since they are very small, so their models can produce a full response to the same token, Lewis told the Tech Crunch, Tech was shown together at the same time as a detailed response in the same time.
Taxkarnch event
Berkeley, ca
|
June 5 June
It is just a little quick to tell if Fastino’s vision will continue. Enterprise is a crowd in AI space, in which companies like Kohir and Data Berks also help AI, which takes the lead in certain tasks. And the enterprise -based SATA model makers, including anthropic and masters, also offer small models. It is also no secret that for the enterprise is the future of productive AI It is likely in small, more concentrated language models.
Time can tell, but the initial vote of confidence in Khosla does not bother. For now, Fasteino says its focus is on the construction of a modern AI team. It is targeting researchers in the high AI labs who do not have the passion to create the biggest model or defeat the benchmark.
“Our services strategies are very focused on researchers that it is probably the process of contradictory thinking about how language models are being built right now,” says Lewis.