This New Frontier Model Is Not An LLM.

Play the video (opens in a new tab)

Tim Carambat, who builds AnythingLLM, plays TypeSafe’s launch video and then explains the idea with a whiteboard sketch. At the time he had applied for access but had seen results through a friend who had it. His example is “is this email spam?” Asked of a chat model, the question costs reasoning tokens, and the answer can come back as “I think the user should move forward” when the code expected “forward”. A System One model gets the same input plus a fixed list of options. It returns a probability for each, can only pick from that list, and scores the options in parallel instead of one token at a time.

He points out an open question: what the model does when none of the options fit the input. Neither he nor his friend had tested it. For uses, he names email categorization and model routing by prompt complexity. His framing is that LLMs keep the chat and reasoning work while models like Jev handle tool choice, filtering and classification without training a custom classifier. He also asks what hardware Jev needs and whether it could run locally, which TypeSafe had not said.

In a later description update he says he got access a day after posting and calls Jev “just a neat general-purpose classifier”. He disputes the “frontier” label and links Laya as an open-source model that does the same thing.

Open on YouTube