The beginning comes out of stealth with a new type of AI model

BeginningA new company based in Palo Alto, launched by computer science professor at Stanford Stefano Ermon, claims that it has developed a new AI model based on “diffusion” technology. Inception calls it a large diffusion -based language or “DLM” briefly.

The AI ​​generative models receiving the most attention now can be widely divided into two types: large linguistic models (LLMS) and diffusion models. LLMS are used to generate text. Meanwhile the diffusion models that power AI systems Midjourney and on Openai Soraare mainly used to create images, video and audio.

The Inception model offers the capabilities of traditional LLM, including the generation of codes and answering questions, but with much faster efficiency and reduced calculation costs, according to the company.

Ermon told TechCrunch that he studied how to apply diffusion patterns Send text for a long time to his laboratory in Stanford. His research is based on the idea that traditional LLMs are relatively slow compared to diffusion technology.

With LLMS “You can’t generate the second word until you generate the first one and you can’t generate the third until you generate the first two,” Ermon said.

Ermon was looking for a way to apply a diffusion approach to the text, since, unlike LLMS, which work consistently, diffusion models begin with a rough evaluation of the data they generate (eg photo) and then adds the data into a focus at one time.

Ermmon is hypotheting the generation and modification of large blocks of text in parallel with diffusion models. After years of experience, Ermon and a student of his achieved a great breakthrough that they describe in detail in a Research paper Posted last year.

Recognizing the potential of progress, Ermon founded Inceptation last summer, eavesdropping on two ex -students, Ucla professor Adita Grover and Cornell Volodimir Kuleshov professor to manage the company.

While Ermon declined to discuss Inceptation’s funding, TechCrunch realizes that the Meifiend Fund has invested.

The creation has already provided several customers, including unnamed Fortune 100 companies, dealing with their critical need for reduced AI latency and high speed, Eon said.

“What we have discovered is that our models can use graphic processors much more efficiently,” Ermon said, citing computer chips, which are usually used to execute models in production. “I think it’s a big deal. This will change the way people build language models. “

Inception offers API, as well as local options for implementing devices, support for fine model settings and a DLMS set outside the box for different use cases. The company claims that DLMS can operate up to 10 times faster than traditional LLM, while costing 10 times less.

“Our ‘small’ encoding model is as good as (Openai’s) GPT-4O Mini Although more than 10 times faster, “a spokesman told TechCrunch. “Our” mini “model is superior to small open source models like (Meta’s) Call 3.1 8B And it achieves more than 1000 tokens per second. “

Tokens is an industrial language for bits raw data. A thousand tokens per second is Impressive speed reallyAssuming that Inception’s claims are detained.