TypeSafe AI founder Diogo Almeida, in an interview before releasing Jev, should have realized at OpenAI that aligning models with human preferences is an unreasonable optimization direction for automated tasks. So this interview provided a good explanation of his automation concept. And Jev's alignment direction is RLCD. If you are interested, you can delve deeper into the interpretation at the end of my article (what is RLCD? The secret link behind Jev. The video mainly shares: He believes that the current AI model is "AI: Overpromise, Underdeliver" to summarize this phenomenon. 2. There is a huge gap between two types of tasks On one hand, there are abilities that are "too good for humans" (such as chatting, coding assistants, customer service scripts); On the other hand, there are scenarios that are not sufficient for true automation (requiring high reliability, no illusions, and decisions that can be directly called by code). 3. The Bitterest Lesson He borrowed and extended Sutton's viewpoint: algorithms, computing power, and data are all important, but the most crucial thing is whether the task is being done correctly. The current model mainly focuses on optimizing "to satisfy humans" rather than "to make software directly usable". 4. Side effects of RLHF Speaking of Mode Collapse: In order to please humans, models will lose diversity and become too "safe" and too accommodating, which is fatal when precise decision-making is needed. 5. Provide examples to illustrate the problem For example, the news of Taco Bell using AI to quickly order food and causing a breakdown is used to illustrate that throwing LLM directly into high-risk, real-world processes is not yet feasible. 6. Their direction Don't turn AI into an assistant for chatting with people anymore, but into a decision engine for software (later released as Jev/System One Model): Input state+structured problem, directly output choices/scores with probability and confidence, instead of generating a paragraph of text. The goal is to make AI like water and electricity, an infrastructure that code can call with confidence, truly driving automation and economic change. My opinion on JEV: Why current chat models cannot bring about a true automation revolution, and why the next generation of models should be optimized for computers rather than for human chat. If you only understand JEV as another BERT, doing classification, labeling, or another sorting model often underestimates its potential. He has better generalization and generality, and now many uses can be seen on X, which cannot be replaced by BERT. Moreover, he is not a model that solves all problems, and some scenarios may be better and more cost-effective when paired with LLM. But after all, JEV only outputs probabilities, and for some important irreversible scenarios, the code still needs more checks and validations, otherwise relying solely on the probabilities it outputs can easily lead to huge risks. And for some LLM applications, I vaguely see the possibility of increasing profits. JEV will promote the development of the application ecosystem, but it is not omnipotent. Although it does not generate illusions, it also generates probabilities. Note: TypeSafe AI founder Diogo Almeida (former OpenAI researcher, involved in RLHF, InstructGPT, ChatGPT related work) gave a speech before releasing Jev. What is RLCD? The secret behind Jev: https://di-zhang-llm. (github.io)/blog/what-is-rlcd-the-secret-behind-jev/