Post

two days ago a former OpenAI researcher (worked on ChatGPT's RLHF) released jev. it's a model that can't talk back. here's the problem it solves:

Thomas - sourcetms

@sourcetms.com

18 Sept

every time you ask ChatGPT to categorize something or make a decision, you get back a paragraph. then you have to read it, interpret it, and hope it's in a format you can actually use. jev doesn't write paragraphs. it just answers questions with structured data.

Thomas - sourcetms

@sourcetms.com

18 Sept

you give it a customer support ticket and ask: is this urgent? what category? what's the sentiment? it returns: urgent (87% confident), billing issue, frustrated. there's no explanation. no longwinding essay. just decisions your systems can act on immediately. the difference:

Thomas - sourcetms

@sourcetms.com

LLMs write and explain. jev decides and scores. use ChatGPT to draft your email. use jev to route 10,000 tickets while you sleep. their numbers: 194x faster, 445x cheaper than LLM baselines on classification workflows.

18/9/26, 6:44 am0 reposts0 likes

Thomas - sourcetms

@sourcetms.com

18 Sept

$0.042 per million tokens. 70-500ms latency. output tokens are free because it doesn't write prose. if judgment becomes cheap enough to run on every ticket, every agent step, every tool call... the amount of AI in software (again) goes up, not down. that's the jevons paradox bet.