The vocabulary around AI is enormous and mostly noise. Underneath it there is one question that orders everything: does the model's own set of numbers change, or not? Answer that and you know what a project will cost you, what data you need, and whether it is even possible from where you are sitting.
Three questions, answered before anything else on this page will make sense. Any word with a can be tapped.
An AI model is not a program someone wrote line by line. It is a huge list of numbers — often billions of them — called {{weights}}, plus a fixed recipe for multiplying your input against them.
“Training” means slowly adjusting those numbers while showing the model millions of examples. When training stops, the numbers are frozen in place.
The old way: to solve a task you collected labelled data for that exact task and trained a model from scratch. Every new task meant starting over.
The new way: one organisation spends enormous money training a single very general model on a huge amount of unlabelled data. Everybody else then adapts that one model to their own tasks, cheaply.
Training a model like this can cost millions of dollars and thousands of specialised chips. Almost nobody can do it.
But using one can cost almost nothing. That gap is the single most important fact about AI right now — including for who gets to participate and who does not.
Every use of a foundation model is one of five things. They are ordered by a single property — do you change the model's numbers, and by how much? That one axis predicts the cost, the data you need, the hardware you need, and whether the model's owner can stop you.
Tap any rung to read what it means in practice. Your choice carries into the next section, where it highlights every model you can reach that way.
Each coloured bar is one model, and its position is the argument: bars on the left can only be asked questions; bars stretching right can be retrained. A single square means you can do exactly one thing with that model, and nothing else — usually because its weights were never published.
Tap any bar for what the model does and the catch attached to it. Tap a rung in Section 1 to dim everything not reachable that way.
Scroll the table sideways on a narrow screen. Names in this field change every few weeks — treat the names as a snapshot and the five columns as the part that stays true.
This lesson was adapted from an earlier draft. Before publishing it we checked every claim against a primary source. Three did not survive, and rather than quietly fix them we are showing you the corrections — because how they were caught is the part you can reuse, and because a page about evaluating AI claims that hid its own errors would be worth nothing.
The most useful habit in this whole field is cheap: before believing a foundation model helped, check it against something deliberately stupid — a straight line, or “find the most similar example we have already seen.” Three independent 2026 evaluations did exactly that, and the results should change how you plan a project.
Five single-cell foundation models were tested against pre-registered linear and nearest-neighbour baselines across the dimensions a “virtual cell” is supposed to handle.
Two of the seven could not be tested end-to-end at all. Of the five that could, the dumb baselines matched or beat every foundation model on four.
Genomes tell a subtler story, and the subtlety matters. GENEB compared 40 genomic foundation models across 100 tasks in 13 categories. Scale does show a substantial association with performance in aggregate — so “bigger is useless” would be the wrong lesson. But it is an unreliable predictor of any particular category, architecture and pretraining alignment frequently offset large size differences, and the model that wins under full supervision reranks in 8 of the 13 categories once you move to a 10-shot budget. The leaderboard you read is not the leaderboard for your task.
Pathology has the most actionable version. Independent studies found these models encode which site and which scanner a slide came from more strongly than they encode the disease. Standard stain normalisation reduces the dominant part of that signature but not the subtle part, which stays recoverable by a simple linear probe.
Everything above is general. This part is ours. Rung 05 — pretraining from scratch — is out of reach here for the reason the compute gap explainer measures: Evo 2 took 2,048 H100 GPUs, and the African regions are the most expensive on Earth to rent that class of hardware. That is real, and it is not the interesting part.
The interesting part is that rung 04 is not only reachable — for our languages it is where the winning is.
The WAXAL corpus, released by Google Research, contains about 1,250 hours of natural conversational speech in 19 African languages. Four of them are spoken in Ghana: Akan, Dagbani, Ewe and Dagaare. A 2026 study then asked the question that matters for anyone building here: is it better to prompt a big model, or to fine-tune a small one?
Word error rate, macro-averaged across all 19 languages. A 26.9 percentage-point gap.
The winning systems were Whisper Tiny (39M), Whisper Small (244M) and MMS-300M — 3 to 40× smaller than the baselines they beat. The smallest is 39M parameters, which runs on a phone.
Read that carefully, because it inverts the assumption most people arrive with. On our languages, the frontier model asked politely (rung 01) loses to a model small enough to run offline on a handset, once somebody does the unglamorous work of fine-tuning it on real speech (rung 04). The binding constraint is not compute and it is not model size. It is transcribed audio in the language, and that is a problem Ghanaians are uniquely positioned to solve and outsiders are not.
Every term used on this page, in one place. The full site glossary goes much wider — browse every term →
Everything the main page deliberately left out, with references. Nothing here is needed to understand the sections above — this is for the reader who wants the original work.
Six quick ones, or the full set. Every answer explains itself, and your first answer is the one that counts.
Each figure on this page is held in website/scripts/build-foundation-data.py alongside the URL it was read from, and check-foundation-claims.py fails the build if the prose and that table ever disagree. Verified .
If you take one thing: most useful AI work sits on rungs 02 and 03, needs no GPUs, and is bounded by careful thinking rather than money. And for Ghanaian languages, rung 04 on a small model beats the frontier model — the missing ingredient is transcribed audio, not compute.
This is one of the interactive explainers we make at AgenticGHX — talks, learning tracks, and research out of Ghana.
More from AgenticGHX Learn →