Training only the readout layer of a fixed random dynamical system turns out to be dramatically cheaper than training a full network. That property has moved reservoir computing from a theoretical curiosity into a serious candidate for edge devices, photonic hardware and neuromorphic chips. The field has labs, conferences and a growing commercial edge, particularly wherever power budgets matter and running a full transformer is out of the question. What it lacks is a canonical address.

In a market where every company is fighting for a name, the last two characters here are doing real work. ML reads as machine learning to the entire technical audience without a word of explanation. That’s a meaningful head start in a crowded category. It puts the subject in the address without spending the whole name on a descriptive phrase that will date, and the IK prefix keeps it distinctive and trademarkable rather than generic.

Job adverts for this role have multiplied faster than anyone can fill them, and the title itself is still settling. Machine learning engineer, ML developer, MLOps engineer. The abbreviation is stable even where the phrasing isn’t. Two of them, joined, in five characters. Practitioners read it instantly and nobody else needs to. The work is genuinely distinct from both data science and conventional engineering, with its own toolchain: experiment tracking, model registries, feature stores, deployment pipelines and production monitoring.

Almost nobody builds on a single model any more. Routing between providers, ensembling outputs, falling back when one fails, comparing results and managing cost across vendors has become structural to how AI systems work. The companies solving this are a real and fast-growing category of model gateways, routers and orchestration layers, and what the category lacks is a plain-English name. This is it. The name suits a routing or gateway product, a comparison and benchmarking service, an API aggregation platform, or a research resource tracking model performance.

The word did not arrive with the current wave. Education researchers have used multimodality for decades to describe learning that arrives through several channels simultaneously, and there is a substantial literature under the term. Machine learning then converged on the same idea from the opposite direction, and the interesting results now come from combinations rather than from any single input type. Two established fields sharing one plural noun is unusual and useful.

Buyers evaluating this layer have a genuine information problem, since the products differ in ways that only become apparent in production. A comparison resource would be used constantly. The plural names the category rather than a single product, which is the right shape for a directory, a comparison site, or a publication covering the space. Model aggregation became structural to how AI systems are built: routing between providers, ensembling outputs, managing fallbacks, and comparing cost and quality across vendors.

The vendor-neutral form matters. This name covers every model family rather than tying a brand to one provider’s naming, which makes it more durable as the market shifts between providers. Inference is the operational half of machine learning and the half costing money every day. Training is a capital expense; inference is the running cost, and optimising it absorbs a great deal of current engineering effort. The name suits an inference service, an optimisation product, a hardware company, a benchmarking resource, or a cost management tool.

Organisations routinely discover their inference spend after it has already become a problem. Cost management is genuinely underserved here. Training gets the attention, but the recurring cost of running models in production determines whether an AI business works, which has made inference optimisation one of the most valuable specialisms in the field. The plural suits a service or volume business: many inferences, billed or measured, which is literally how the market works, priced per token and per call.

Engineers evaluating options search in precise vocabulary rather than in marketing language, which is why exact-term naming is an advantage for a developer-facing product. The singular suits a product rather than a service volume. One inference: the unit of work being optimised, measured and billed. Inference became the operational centre of the AI business. Running cost determines margin, and an entire supply chain formed around making it cheaper and faster: specialised hardware, quantisation, caching, batching, routing and serving frameworks.