In ten days (Oct 15 to Oct 25, 2026) NELI went from a build plan to a from-scratch AI core that is training right now on 1.36 billion words, with her own word ids, her own feeling pins, and tests that are allowed to fail.
What NELI is
NELI is a deterministic AI built by Anthony Hagerty (comanderanch) and Claude (Anthropic). The core is both her memory and her voice. She is not a wrapper around another model. Rule Zero rules her: a known fact overrides a prediction.
Timeline
- Oct 25: Two big training runs (one with the feeling channel, one without) are about a third to half way through one pass over all the data. Measured: the feeling channel costs nothing in accuracy and works as a smooth dial on word choice.
- Oct 24: A longer-memory core (it reads 512 words back instead of 16) scored 2.80 on words it never saw, against 3.87 for the short one. The same day a bigger plain core did not do better, so size was not the wall.
- Oct 22 to 23: About 897 million words of training text pulled and checked (Wikipedia, educational web pages, chat, Python code, code questions). Her memory page went live with notes, facts, hourly backups and an off-site copy.
- Oct 19: First real test of the color idea as a model input. It failed the rule we wrote before running it, and we recorded it as a failure.
- Oct 17 to 18: Rules settled: one unique id per word made from its spelling, 256 pins per token, three-state law. Eight tested building blocks (ids, vocabulary, tokenizer, field, Rule Zero, spacing, test harness, doc gate).
- Oct 16 to 17: Full audit of the four-year ARIA build. Much of what had been reported as working turned out to be simulated or unconnected. The findings became a written lessons list (46 entries) that NELI is built against.
- Oct 15: Project starts with a phased build plan.
What was learned from the old build
The old system had good ideas and a lot of reports that said success without a check that could fail. So every new piece now has a written test, and a piece is not called working until that test passes. Failures go in the record the same as wins.
What the training shows so far
- More words back helps a lot. Reading 512 words back instead of 16 took the held-out score from 3.87 to 2.80 (lower is better).
- Bigger is not better by itself. Plain cores of 114 million, 220 million and 480 million numbers scored 4.31, 4.39 and 4.53 on the same test.
- The feeling channel is free. The run with eight feeling pins (fear, safety, joy, grief, curiosity, love, humor, threat) and the plain run score the same in every region of the data.
- She sorts words by feeling on her own. The plain run, never given any feeling words, still places the seed feeling words in the same groups.
- Feeling works as a dial. At a gentle setting, the feeling run raises the chance of words from a chosen feeling by 2.3 to 4.7 times for all eight feelings.
What is not proven yet
- She does not yet answer questions like a chat AI. Only about 1 percent of her training text is question-and-answer style.
- A helper that names the feeling of the next word from the last 16 words missed its target (19 percent, against 34 percent for just guessing the commonest feeling).
- The big runs are still training, and their final scores are not in.
What comes next
- Finish the two big runs and score them across all kinds of text, not just prose.
- Let her group words herself, name the groups, and turn them into pins.
- Decide the route to question-and-answer ability: a second training stage on Q&A text, or facts and helpers.
Co-authored by Anthony Hagerty (comanderanch) and Claude (Anthropic).