Diogo Almeida (ex-OpenAI) reveals JEV: 200x faster, 400x cheaper, zero hallucinations.

Summary

Diogo Almeida argues that today’s language models often overpromise but underdeliver because they are primarily optimized to assist people, rather than reliably carry out work on their own.

• Assistance versus automation: Chatbots and coding agents can be useful when a human reviews their output, but that oversight makes them assistance tools rather than fully automated systems. Almeida says the distinction matters especially when a system must make decisions or take actions with real consequences.
• Why reliability matters: Models can produce plausible-sounding answers that are wrong, limiting their usefulness for tasks that require dependable decisions. Almeida points to customer-service bots that answer questions but generally do not take action as an example of this boundary.
• Optimization shapes behavior: Models tend to become good at what they are trained to optimize. Almeida argues that training for human preference can favor responses that seem safe and convincing over outputs that reliably complete a task.
• RLHF and “mode dropping”: Almeida describes reinforcement learning from human feedback (RLHF), a common approach to training chat models, as encouraging models to favor safe, likely-looking answers and omit less common possibilities. He contrasts this with pretraining, which he says can preserve a wider range of possibilities. He presents this as one explanation for why answers can sound confident or plausible while still being incorrect.
• Data and task choice: Almeida challenges the idea that more compute alone drives progress. He argues that choosing the right task and data can matter more, and points to the different trajectories of OpenAI and Anthropic as an example of how a clearer optimization target may help a lab advance.
• Caution around consequential actions: He advises against letting current models make high-stakes decisions without safeguards or human oversight. He also raises concerns about systems such as OpenClaw being given too much autonomy.
• A different direction: Almeida argues that the field can move beyond assistance-focused models toward systems designed specifically for automation. He describes his company’s approach as “type-safe”: integrating AI more directly with structured software interfaces rather than relying only on text inputs and outputs. He says the company is focused on automating routine work, not building another human-in-the-loop assistant, and mentions a signup form and hiring.
• Limits of the talk’s evidence: The talk presents a conceptual case for this direction, but the subtitles do not include product demonstrations or verified performance figures. They also do not substantiate the speed, cost, or hallucination claims suggested by the video title.

The talk is primarily a technical and strategic argument; it does not provide a product review, step-by-step guide, or tutorial.

Main Speaker and Sources

• Main speaker: Diogo Almeida, captioned as “Diego” in places, is described as a former OpenAI researcher and co-author of work related to GPT-4, ChatGPT, and RLHF.
• People and work referenced: Yann LeCun and his critique of language models; Rich Sutton’s “bitter lesson”; Google’s FLAN instruction-tuning work; and developments at OpenAI, Anthropic, and Thinking Machines.
• Other voices: Audience members briefly respond during the talk, but no other presenter is identified.