Modern AI Tools

1A tool family is an object, not a brand

A chat model, a model hub, an image generator, a code completer and an AutoML search are five different objects. GPT-class systems are large next-token predictors. Hugging Face is a place that hosts model files and the code to run them. Stable Diffusion is a generator that turns a text prompt into an image. Copilot-class tools are next-token predictors specialised to code. AutoML is a search that tries model families and settings on a table you already framed.

Brand names move. The objects do not. This lesson uses the names a current syllabus prints so you can recognise them, and then immediately talks in the object language: predictor, hub, generator, completer, search. If a newer name replaces one of these, the job test stays — what does it return, and what can it not certify?

2Chat models predict text

A GPT-class chat model is trained to continue text. Given a prompt, it samples a plausible next token, then another, and the paragraph that appears is that chain. It is not a knowledge base, not a minimax player, and not a calibrated classifier. It can draft the overdue-book email, outline a section, or rephrase a paragraph you already believe. It cannot certify that a citation exists, that a rule follows from a KB, or that a probability is calibrated.

The failure that matters is fluency: a wrong fact in correct grammar. The evaluation and leakage lessons already taught you not to trust a number that looks finished. The same habit applies to a paragraph that looks finished. Use the chat model as a drafter; keep the earlier lessons' tools for anything that has to be true.

3A hub is not a model; a generator is not a fact

Hugging Face is a hub: a catalogue of published model files, datasets, and a library that loads them. Saying 'we used Hugging Face' is like saying 'we used the library' — it does not name the book. The object you owe a reader is the model card: which weights, which task, which licence. The hub also does not run your clinic rules; it is not an expert-system shell.

A Stable Diffusion-class generator produces an image from a text prompt by iteratively denoising in a learned latent space. That is a picture, not a photograph of the lab and not a diagram of the AND-OR graph from the AO* lesson. Use it for a slide that needs an illustration you will label as generated. Do not use it as evidence that a device looks a certain way, and this course will not ask you to generate one here.

4Completers and AutoML still need a job

A Copilot-class completer is a next-token predictor that has seen a lot of code. It can propose a loop, a test name, a pandas snippet. It does not run the snippet, does not know your label definition, and does not inherit the A* optimality proof. Read every suggestion the way the hybrid lesson reads an LLM tool call: a proposal, then a check. The check is still yours — a compiler, a test, a reviewer.

AutoML searches a space of model families and settings on a table you provide, and returns a fitted model plus a score. It is a real search, closer in spirit to the search lessons than a chat box is. It cannot invent the decision, the label window, or the split. If the overdue-book flag is actually 'returned after a reminder' mixed with 'never returned', AutoML will happily maximise a meaningless accuracy. Frame first — the learning-setup habit from the ML course, and the evaluation lesson here — then let a search spend the table.

5Pick the object the job already named

The course you just finished already named objects that these families do not replace. A cheapest path is still A*. A required pair of subtasks is still AO*. A safe move against an opponent is still minimax. A sentence that must follow is still entailment. A leftover 'I do not know which' is still a Dempster mass on \Theta, not a chat hedge. A constraint that must never bend still belongs in a rule layer, as the hybrid lesson said.

Reach for a modern tool when the job is drafting, illustrating, proposing code, loading a published model, or searching a framed table. Reach past it when the job is a guarantee. The project that flags overdue books can use a chat model for the email and a table model — AutoML or otherwise — for the flag, and it still needs a definition of overdue that no tool will invent.

Figure. A cheapest path is still A*. A required pair of subtasks is still AO*. A safe move against an opponent is still minimax. A sentence that must follow is still entailment. Reach past a modern tool when the job is a guarantee.