by its parts Paper 01 · Natural language · Co-first author Open paper 01 →
- +9.8 Structure F1 on characters the tokenizer cannot represent, where token-only models fall to near chance.
- Submitted to ACL Rolling Review 2027. Under review; no decision yet.
- Costs 2.5M parameters against BERT-base's 110M, and leaves the backbone untouched.
Liqian (Eric) Yan · Shanghai, China
Language is infrastructure. I study how machines learn to read it, and what that does to the work people are hired for.
million job adverts Paper 02 · Economics · First author Open paper 02 →
- 100M postings across 72 countries, 2015–2024, scored one at a time by a distilled language model.
- Regeneron ISEF 2026 Finalist. Still collecting 2025 postings and looking for an external check on the index.
- −2.1 percentage points of routine hiring per unit of AI intensity, with flat pre-trends.
Two papers, one question.
A Chinese character is built out of parts, and a tokenizer throws every one of them away before the model sees anything. A job advert is a firm saying, in its own words and at its own expense, what work it needs done next, and almost nobody reads them at scale. Both papers are about the same thing: structure that is sitting in plain sight, discarded because the standard way of measuring cannot see it.
I am a high-school student in Shanghai. Most of what is here started as something I wanted the answer to on a weekend and then refused to leave me alone. Both manuscripts are published on this site in full, not summarised, because a claim you cannot check is not worth much.
Unicode already publishes how every Han character decomposes. This paper parses those descriptions into trees, encodes them with a small recursive network, and folds the result into BERT's embedding layer before the first transformer block runs. On characters outside the vocabulary the effect is large; on ordinary text it is close to nothing, and the paper says so.
The AIFE Engine Read the argument, then the manuscript →A frontier model reads a sample of postings and scores them for how much real AI work they describe. Human labels correct its drift, a much smaller model learns the corrected scores, and that model reads all hundred million for $410 rather than $98,000. What comes out is a firm-by-year panel nobody had before.
Certified · every claim with its source
The record, with receipts.
The other half of the week.
I am co-founder and CFO of LinkScape, a student-run nonprofit fiscally sponsored by Hack Club, where I look after budgets, sponsorships and compute costs. Its flagship, Resonaite, reads eight EEG bands once a second and re-mixes the audio in response. Seven stages of filtering stand between that idea and something a person would willingly listen to, which is most of what the work turned out to be.