He grew up from the ghost with a new type of artificial intelligence model

beginningA new company based in Pallo Alto, which was started by Professor of Computer Science at Stanford Stefano Ermun, claims to have developed a new model of artificial intelligence based on “spreading” technology. It is called an Inception as a large -based language model, or “DLM” for a short period.

Tructured artificial intelligence models that now receive most of the attention can now be to two types: large language models (LLMS) and proliferation models. Llms, built on TransformerIt is used to generate the text. Meanwhile, prevalence models, which are similar to artificial intelligence systems Midjourney And Openai’s SuraIt is mainly used to create images, videos and sound.

The Inception Model offers traditional LLMS capabilities, including the generation of the code and the answers of questions, but with significantly faster performance and computing costs, according to the company.

Tell Ermon Techcrunch it is studying how to apply Proliferation models To a long text message at the Stanford Laboratory. His research was based on the idea that traditional LLMS is relatively slow compared to spread technology.

With LLMS, “You cannot create the second word until you create the first word, and you cannot create the third word until you create the first two.”

ErMon was looking for a way to apply the spread of the text because, unlike LLMS, which works in a sequential manner, the spread models begin with an approximate estimate of the data it generates (for example, a picture), then bring the data to focus once.

Ermon was supposed to generate and modify large blocks of the text in parallel with the prevalence models. After years of trying, Irmun achieved a major penetration, they detailed in a Search paper Posted last year.

With the realization of the possibilities of progress, Ermon Inception was founded last summer, and he benefited from two former students, Professor UCLA Aditya Gover and Professor Cornell Volodimir Colicov, to participate in the company’s leadership.

While Ermon refused to discuss Inception, Techcrunch realizes that the Mayfield box has invested.

Emron said that Inception has already guaranteed many customers, including the Fortune 100 companies whose name has not been revealed, by addressing its decisive need to lower the time of artificial intelligence transmission and increased speed.

“What we found is that our models can benefit from the graphics processing units more efficiently,” Ermun said, referring to the computer chips used to operate the models in production. “I think this is a big problem. This will change the way people build language models.”

Inception provides an application programming interface as well as edge publishing options, accurate models support, and a group of DLMS outside the box for various cases of use. The company claims that DLMS can reach 10x faster than the traditional LLMS while it costs 10x less.

The small “coding” model like (Openai’s) GPT-4O Mini While more than 10 times the speed, “told the company Techcrunch. Call 3.1 B. More than 1000 symbols are achieved. “

“Symbols” are the language of making parts of raw data. A thousand symbols in the second Really great speedAssuming that the establishment of the founding was resurrected.