Open vs closed, big vs small

Two questions from the news cycle are left standing, and they travel together: why does every AI announcement brag about size — parameters, billions, trillions — and what is an "open" model, which people discuss with either religious fervour or corporate dread? Both answers fall out of machinery you already understand, and both are worth having straight before the next headline reaches you.

Parameters, in one honest sentence

All those adjustable internal settings this course keeps calling "dials" and "the blur" — the numbers that pretraining nudged trillions of times (lesson 5)? Each one is a parameter, so a "70-billion-parameter model" is a model with seventy billion of them: the resolution of the blur. More parameters mean more capacity to absorb pattern and nuance — scale genuinely was the story of the last few years, and small models are noticeably blurrier. But the honest caveats have grown as fast as the models: capability per parameter keeps improving (training data quality and technique matter enormously — a well-trained small model as of mid-2026 beats a sloppily-trained giant of two years earlier), vendors have largely stopped disclosing counts for their frontier models, and past a point, parameters buy less than the marketing implies. Treat the number, when you even get one, the way you treat megapixels on a camera: correlated with quality, gameable, and never the spec that settles the question.

Size buys capability; size also costs. Every token generated runs the whole parameter stack (lesson 1's loop), so bigger means slower and pricier per word — which is why the market splits into families: frontier models for hard reasoning, mid-size for the everyday bulk, small models where speed and volume rule — including ones small enough to run on your phone, which is what "on-device AI" means: nothing leaves the handset, works in airplane mode, blurrier but private and instant. Matching model size to task, rather than defaulting to the biggest, is half of what "AI cost engineering" means in practice.

Open weights: the recipe question

Now the fervour. A closed model (the frontier ChatGPT, Claude, Gemini tiers) lives on its maker's servers; you send requests, answers come back, the parameters stay locked in the building. An open-weight model — Meta's Llama line, Mistral, DeepSeek, and a fast-growing field as of mid-2026 — has its parameter file published: anyone can download the actual blur, run it on their own hardware, fine-tune it (lesson 12) into a specialist, and never send a byte outside. ("Open-weight" rather than "open-source" is the pedantically correct term — you usually get the trained parameters, not the training data or full recipe — a distinction that matters in licensing arguments you can now follow.)

Why it inspires fervour: open weights mean AI capability isn't exclusively rented from three companies — a hospital can run models entirely inside its walls, a startup can build without per-token tolls, researchers can actually inspect what they study, and India-focused teams can tune models for Indic languages the frontier labs underserve. Why it inspires dread: a published blur cannot be un-published, and lesson 6 taught you that an assistant's manners and refusals are a trained layer plus a rulebook — layers that whoever holds the weights can strip or retune. Both camps are reasoning correctly from the same fact. The gap between frontier closed and best open models keeps narrowing without quite closing, and where it stands any given quarter is exactly the kind of question Snack exists to keep you current on.

Try this in your chatbot: ask — "If a 3-billion-parameter model runs on a phone and a frontier model needs a data centre, walk me through what the extra parameters are buying, with examples of questions each handles well or badly." You'll get lesson 5's blur, argued back to you with specifics — a good test of whether the mental model holds.