He Built the 'New' AI a Year Ago. A Big Lab Just Got the Credit
The AI world's credit problem, told in one very frustrated Reddit post.
A developer posting as Nandakishor_ml says that back in March 2025 he published a research paper, a working AI model, a downloadable software package, and a training dataset for a new design he calls "Jev." Most AI today writes one word at a time, like someone slowly typing a text message. His design skips that step: it looks at the whole question at once and spits out probability scores — more like a weather forecast giving odds than a paragraph of prose. Roughly a year later, he says, a major AI lab announced essentially the same idea as a breakthrough, without publishing technical papers, open model files, or the data behind it.
His version was built for a very specific job: sales conversations. It was trained using reinforcement learning — the same reward-and-punishment loop that teaches AI to play games — to read a sales chat and predict, turn by turn, how likely a customer is to buy, expressed as a number from 0.0 to 1.0. The lab's version reportedly outputs confidence scores and structured choices instead. Architecturally similar, he argues, just aimed at everything (horizontal) rather than one industry (vertical). He published a second related paper in September 2025.
Two honest caveats. First, similar ideas genuinely do emerge in different places at the same time — the history of science is full of it — so this is one person's account, not proven copying. Second, there's a huge gap between a demo that works and a product used by millions; the lab brings money, computing power, and distribution that one person simply cannot. The frustration here isn't mainly about the idea. It's about the silence: no paper, no open model, no data to check.
Why should you care? Because open research is how outsiders can verify what AI companies claim. When a lab announces a breakthrough without sharing anything, you have to take its word for it. Meanwhile, small builders doing careful work can be quietly leapfrogged by whoever has louder headlines. For you personally, the upside is real: faster, cheaper AI that answers in one shot rather than slowly typing. The downside is that the "brand new" breakthroughs you hear about may already have been built, published, and ignored.
- A one-person team says he published a paper, working AI model, and dataset in March 2025 for an idea a major lab later announced as new.
- The idea itself is simple: instead of writing answers one word at a time, the AI predicts probabilities all at once — potentially much faster and cheaper.
- His version was built for sales chats, predicting how likely a customer is to buy, on a scale from 0.0 to 1.0.
Why It Matters
It shows how small builders can be quietly overtaken — and why open research lets you check AI companies' claims.