TypeSafe AI CEO Says LLMs Are Not Built for Automation
Diogo Almeida told a San Francisco event that LLMs assist rather than act autonomously, as TypeSafe AI pushes its Jev decision-making model.
Diogo Almeida, CEO of TypeSafe AI, argued at an event in San Francisco that large language models are structurally unsuited for automation. According to The Information, which reported on his remarks, Almeida said LLMs are inherently not optimized for automation because they are trained through reinforcement learning from human feedback to assist rather than act autonomously.
Almeida's company is promoting an alternative. TypeSafe AI recently introduced Jev, a model focused on decision-making rather than text generation that selects from predefined options based on probability and confidence scores. The company says Jev has a response time of 0.07 to 0.5 seconds and is priced at $0.042 per million input tokens, with output free of charge.
TypeSafe AI claims 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. Anthropic Claude Code product lead Cat Wu said much automation remains at a semi-automated stage.
The gap between promised and delivered automation shows up in user data. A Workday survey of 3,200 employees and executives found respondents spent an average of 37% of the time saved by AI fixing and rewriting low-quality outputs, as reported by The New York Times. Google economists surveyed 600 scientists who reported saving an average of 7 hours per week with AI, but nearly 90% of that saved time involved checking AI outputs.
Dax Raad, the developer behind OpenCode, said engineers often do not sufficiently review AI-generated code. Almeida is a former OpenAI researcher who now leads TypeSafe AI.
Quick answers
What is Jev from TypeSafe AI?
Jev is an AI model focused on decision-making rather than text generation. It selects from predefined options based on probability and confidence scores, with a response time of 0.07 to 0.5 seconds and a price of $0.042 per million input tokens with free output, according to the company.
Why does Diogo Almeida say LLMs are not suited for automation?
Almeida argues that large language models are trained via reinforcement learning from human feedback to assist rather than act autonomously, making them inherently not optimized for automation.
How much time do workers spend fixing AI output?
A Workday survey of 3,200 employees and executives found respondents spent an average of 37% of the time saved by AI fixing and rewriting low-quality outputs.