TypesafeAI, founded by Diogo Almeida (디오고 알메이다), has released a new AI model called Jev that is drawing attention in the tech community.

A new AI model unveiled by a ChatGPT co-developer is in focus in the developer world.

Jev is designed to output probabilities rather than text. Foundation Capital partner Jaya Gupta (자야 굽타) said Jev can be seen as a model that isolates judgment. Given complex situations, narrow questions and a list of possible answers, Jev provides decisions and probabilities instead of sentences.

"What software needs are decisions like whether something is urgent, whether to continue or stop, whether to allow or block," he said. "But now a language model writes a sentence first and then turns that sentence back into a decision. Much of what an agent repeats internally does not need to write sentences," he said.

A 'Tech Insight' looks at an ambitious scenario driven by the unbundling of giant intelligence.

TypesafeAI is positioning Jev as an alternative that can overcome limits of existing large language models (LLMs).

Almeida, the CEO, worked at OpenAI for 4 years and 6 months before founding TypesafeAI. He co-participated in developing GPT-4, InstructGPT and ChatGPT and was also involved in developing reinforcement learning from human feedback (RLHF).

Almeida said two narratives coexist in current debate around AI: that it seems excessively good at some tasks and useless at others.

At an event recently, he cited customer service as a representative case, saying, "Most success stories in the news are assistance tasks where humans intervene in the process. Tasks that do not deliver results are automation tasks that AI must handle on its own. Automation tasks look simpler, but the results are the opposite."

"Most customer-service chatbots cannot execute actions directly," he said. "They stop at finding documents and guiding users because the model can be wrong, so it is not given authority to execute."

Almeida said he also sees coding agents, broadly, as assistive tools because code is closer to a language for people than for machines. He presented as an example of that view the question of whether coding agents can be trusted without version control.

He said the reason this happens is RLHF, which focuses on training models to produce answers that people like. "A model trained with RLHF chooses the option that looks safer when there are two possibilities, so the answer almost always looks plausible. Wrong answers also look plausible," he said. "As precision rises, LLMs become less useful," he said, adding, "Machine learning is the opposite, where higher precision rapidly increases usefulness."

He also drew a line under expectations that AI can both assist human work (human in the loop assistance tasks) and achieve automation. He said if the two goals pull models in different directions, performance fluctuates and becomes what he called "bumpy intelligence."

"To achieve trustworthy automation, accuracy needs to be at a level with 'multiple nines,'" he said.

He said TypesafeAI focuses on automation rather than assistance functions. That means Jev may not suit companies seeking to build assistive services with human involvement. Almeida said he tells companies, "Do not leave tasks involving risk in decisions to AI," adding, "Before asking whether to use Openclo for new-user onboarding, you should first check whether there are risk factors and whether there are security safeguards to prevent harm."