As AI models become more advanced, companies are showing growing interest in autonomous AI agents that can carry out specific tasks on their own without people. Yet a paradoxical scene is still playing out, with grumbling heard even in Silicon Valley that automation driven by AI is moving slowly. That suggests the gap remains wide between expectations for the future and what people actually experience in the field.

More recently, some have pointed out that existing large language models, or LLMs, are structurally unsuited to automation. Diogo Almeida (디오고 알메이다), a former OpenAI researcher who is now trying to develop an AI model with different characteristics from LLMs as CEO of TypeSafe AI, is among those making that case.

According to a report by The Information, he took the stage at an event in San Francisco and posed a provocative question: "Where is all that automation?" He stressed that LLMs are not inherently optimised for automation.

He said all the recent LLMs focus on supporting humans rather than delivering self-directed automation because they are based on reinforcement learning from human feedback, or RLHF. RLHF is a method of training models based on ratings given by humans. Because models learn to satisfy people, they may be good at assisting but weak at automation that runs without humans, he said.

At another event earlier, he said most success stories in the news are support work in which humans intervene in the process. He said the work that does not deliver results is automation work that AI must handle on its own. "Automation work looks simpler, but the results are the opposite," he said. He said most customer service chatbots cannot execute actions directly and stop at finding and guiding users to documents because models can be wrong and are not given execution authority.

He also drew a line under expectations that AI can achieve both human-in-the-loop assistance tasks and automation. He said that if two goals pull a model in different directions, it can produce "bumpy intelligence" with inconsistent performance.

TypeSafe AI recently introduced Jev as an alternative to LLMs. Almeida’s criticism that LLMs have limits for automation could also be rhetoric aimed at packaging Jev as a good fit for automation.

Jev is an AI model focused on decision-making rather than generating text like an LLM. If developers predefine options such as yes or no, scores or categories, it makes decisions based on each option’s probability and confidence score. Speed and cost are also strengths. Jev’s response time is 0.07 to 0.5 seconds. It charges $0.042 per 1 million input tokens, and output is free. TypeSafe AI says its own tests show Jev is up to 193 times faster than Anthropic’s Claude Sonnet 5 and about 444 times cheaper than Claude Opus 5.

TypeSafe AI calls Jev a "System One" model, a term that appears to be borrowed from Nobel economics laureate Daniel Kahneman’s book "Thinking, Fast and Slow," which divides human thinking into System 1 and System 2. According to Kahneman, System 1 is fast, intuitive and operates automatically. System 2, by contrast, is slow, conscious thinking used for complex calculations or logical reasoning.

The Information reported that other AI industry leaders, in addition to Almeida, are trying to spread automation more widely in software engineering organisations but acknowledge the change is not easy.

Cat Wu (캣 우), product lead for Anthropic’s Claude Code, said, "A lot of automation is actually staying at the semi-automation stage," adding, "We are encouraging people to find repetitive tasks and ask why humans are still in that process."

Because AI automation is not easy, for now human-in-the-loop is seen as a methodology that can complement AI’s limits and minimise risks by giving AI a role like a human’s while leaving room for humans to intervene.

Some also point out that the burden of checking AI-produced results is a factor that eats into productivity gains from AI.

A recent New York Times report said Workday surveyed 3,200 employees and executives and found respondents spent an average of about 37 percent of the time saved by AI fixing and rewriting low-quality outputs. A survey by Google economists of 600 scientists showed similar results. Scientists saved an average of 7 hours a week using AI, but nearly 90 percent spent one-tenth of the time saved checking AI outputs. Some 46 percent said they spent more than a quarter of the time saved on that task.

A structure with humans in the loop does not always lead to good results. Some say humans are more troublesome than AI. In concept, a combination of people and AI is attractive, but in reality people may too easily compromise on AI outputs or become lazy.

According to The Information, Dax Raad (댁스 라드), a well-known open-source developer who created OpenCode, said through The Information, "I am worried engineers these days are not reviewing enough of the code written by AI." He added, "I am looking for a way to get team members to ask questions properly. I am also thinking about how to make them feel accountable when code that they did not fully review goes out as is."