Ivan Novikov Unveils XERJ Reader 1, a Neurosymbolic Model That Answers With Programs Instead of Text

Context

Ivan Novikov, founder of the security company Wallarm, now lists XERJ in his X profile. On November 12 he posted that he built a new AI model that is not allowed to generate text. Neurosymbolic systems combine neural networks with explicit symbolic rules or code.

Deep dive

Ivan Novikov posted on November 12 that he has built a new AI model that is not allowed to generate text. Instead, he said, it produces a verifiable program that answers a prompt using the original reference. He described XERJ Reader 1 as a neurosymbolic model for working with data: it points to the exact words in a document, writes a small program over them, and exact code computes the answer.

Novikov claimed the approach is 20 to 90 times faster and 10 to 1,000 times cheaper, with output tokens free, and that every answer is a program that can be replayed. He called it the shortest path to automating data work. The post included a demo video and passed 560,000 views.

The speed and cost figures come from the developer, and the post did not name the baseline models or benchmarks they were measured against, so they are unverified. No paper or independent evaluation was linked in the post.

The idea targets a known weakness of language models on data tasks: they can misquote sources or make arithmetic errors. Grounding each answer in quoted text and a runnable program makes errors easier to check, at the cost of handling only tasks that can be expressed as code.

What to expect next

Watch for a technical report, public benchmarks or open access that would let others test the speed and cost claims. Comparisons with standard language models on document question answering and spreadsheet tasks will show whether program based answers hold up on messy real data. Pricing, an API and early customer use would show whether the approach can compete with general assistants that already handle data work.