Discovering what comes after the transformer.

Eight years since the transformer. Refinements, no paradigm shift.

That could mean transformers are globally optimal. It could equally mean the field's search procedure — humans publishing one legible, incremental move at a time — can't reach what's next: the bitter- lesson pattern, where hand-crafted priors lose to scaled search.

The enabling capability missing in 2020 — an LLM strong enough to be a useful mutation operator over ML principles — exists now. That puts automated discovery, not another round of human-designed refinements, in position to produce the next paradigm. The competitive advantage is not any one architecture we publish; it's owning the engine that produces them.

Hyper is the engine; Code Republic is how it compounds.

Hyper searches ML concepts, turns selected hypotheses into runnable experiments, scores them, and ablates the winners down to reusable designs. Code Republic supplies the implementation memory that keeps that search from rebuilding the same pieces. The CR / Hyper page carries the shared-layer story; the architecture page maps Hyper's loop.

Hiring engineers for Stage 1.

The team's research direction and experimental rigour are owned; the open Stage-1 gap is software engineering. The filter is taste for the problem — familiarity with the nearest prior art (FunSearch, EoH, AlphaEvolve, Darwin Gödel Machine) helps but isn't a prerequisite.

To apply for either role — or to make an introduction — write to ilya@covenance.ai with two or three sentences on which design decision in the engine you'd most enjoy or most disagree with.