What changes, and what does not

Sorting the durable from the perishable

This course has quoted prices, model names, memory figures and capability claims, and a good proportion of them will be wrong within a year. That is worth confronting directly at the end rather than leaving as an unstated embarrassment, because knowing which parts expire is itself part of the method.

What has a shelf life of months

• Every price per token, every subscription figure, every free-tier allowance.
• Every model name and version number.
• Which company leads on which task.
• Context window sizes, which have grown by orders of magnitude and will again.
• What a laptop can run, which improves from both directions as models get more efficient and hardware gets cheaper.
• Which capabilities are paid-only, which migrate to free tiers steadily.
• The specific tools named in the free paths, though the open ones tend to outlive the commercial ones.

If you are reading this more than a year after it was written, assume every number is stale and check the ones that bear on your decision.

What has held for years and is likely to keep holding

Fluency is not a reliability signal. A model sounds identical when it is right and when it is inventing. This has not changed with any release and there is no sign it will.

The product layer decides your daily experience more than the model does. The same weights inside two products feel like different tools.

Output costs several times input, because generating tokens one at a time is mechanically more expensive than reading them in parallel.

Small models handle transformation; large models supply knowledge and multi-step reasoning. The boundary moves; the shape of it does not.

The context window is a ceiling rather than a promise, and material in the middle of a long input is attended to less reliably than material at either end.

Verification cost decides whether a tool pays. Tasks you can check quickly give real returns; tasks you cannot give the appearance of them.

Your own dozen prompts beat any published comparison for your work, because they measure fit to your work rather than fit to a test.

What leaves your machine is the question to settle before the quality question.

The parts that are unresolved

Three, and they should be left unresolved rather than tidied away.

Whether training on scraped material is lawful, which is being litigated and legislated differently in different countries with no settled international answer.

Whether AI output can be copyrighted, which turns on human authorship requirements that differ by jurisdiction and are under review in several.

What sustained use does to skills that are no longer practised. The mechanism is not mysterious (practice maintains ability) but the size of the effect, and which skills matter, is contested and being studied now.

Anyone who resolves these for you in a paragraph is telling you their preference.

How to use a course that dates

The practical instruction is simple. Where a lesson gives you a number, treat it as an illustration of a shape rather than a fact: the twentyfold ratio between small and large models matters more than the specific prices used to show it. Where a lesson gives you a method, use the method.

And when you find something here that is now wrong, that is the course working as intended rather than failing. You noticed because you tested, which is the entire habit these lessons exist to build. The alternative (a comparison written as though it were permanent) would have been more comfortable to read and worse to rely on.

The last word

The tools will keep moving. Prices will fall, capabilities will arrive at smaller sizes, companies will be bought and closed, and something in this course will look quaint.

What you have that survives all of it is a file of twelve prompts, a method for testing, a rule about what material goes where, a budget with a cap, and a date twice a year. That is the whole apparatus, it fits in a folder, and it will still work when every product named in these lessons has been replaced by something nobody has heard of yet.