Getting set up for nothing

The constraint most advice ignores

A lot of study advice assumes a laptop with 16GB of RAM, a stable connection and a card on file. Plenty of students have a shared machine, an older phone, patchy data and no card. None of that stops you doing any of this course. It changes which route you take, and the routes are worth knowing before you need them.

Free access to capable models

Every major provider runs a free tier with a daily or hourly message cap. The caps move; the shape does not. Practical consequences:

• Spend your allowance on the hard questions. Do the easy rewriting with a weaker free model or on your own.
• Batch. One message containing five related questions costs one message from your allowance. Five messages cost five.
• Keep two accounts on different providers. When one runs out, switch. This is not a trick, it is what the free tiers are for.

Beyond the big names: Hugging Face Spaces hosts thousands of free demos, including chat interfaces to open-weight models, and needs no payment. Google Colab gives free GPU time in a notebook, enough to run a mid-sized open model for an afternoon. Some universities provide staff and students with a paid tier free: check before assuming; a surprising number do and never announce it well.

Running a model on your own machine

This is the option that removes quotas, works offline, and keeps everything private. It needs RAM more than anything else.

• 8GB RAM: a 3B or 4B parameter model, quantised to 4 bits, around 2-3GB on disk. Usable for rephrasing, practice questions from supplied text, vocabulary, simple code.
• 16GB RAM: a 7B or 8B model, 4-5GB. Noticeably better; still well short of the hosted ones on hard reasoning.
• Apple silicon punches above its weight here because the memory is shared with the GPU.

The two free tools are Ollama, which is a single install and then ollama run <model> in a terminal, and llama.cpp, which is more work and runs on almost anything. LM Studio gives a graphical interface if terminals are unwelcome.

Be honest about the gap. A 4B model run locally will hallucinate more, follow multi-step instructions worse, and give up on long documents. For a set of tasks (drilling, rewriting, quizzing you from a chapter you paste in) it is entirely adequate, and the fact that it works on a train with no signal is worth a lot.

A phone-only setup that actually works

1. Chat tools: all major ones have web interfaces that work in a mobile browser; no app needed.
2. Documents: Google Docs or the free Microsoft web apps; both save continuously, which doubles as your process record.
3. Notes: Obsidian and Joplin both have mobile apps and sync free via a folder.
4. Maths: GeoGebra and Desmos are excellent on a phone.
5. Reading papers: get the PDF into a reader that lets you highlight; Zotero’s mobile app is free.
6. Voice: dictation is free on every phone, and talking through a problem out loud, then pasting the transcript, is a good way to think.

The thing to avoid on a phone is writing long prose directly into a chat window. It encourages accepting whatever comes back because editing is painful. Draft in a document, paste in, paste back.

Data, and working offline

If data is expensive or intermittent:

• Download the material once. Lecture slides, past papers, PDFs. Read offline.
• Batch your questions. Write them in a note during the day, ask them together when you have signal.
• A local model needs no connection at all after the download.
• Wikipedia has offline packages via Kiwix: the whole of English Wikipedia without images is a few tens of gigabytes, and a subject-specific selection is far smaller.

What not to spend money on

There is a large market in AI study apps charging monthly for a wrapper around a model you can use free. Before paying for anything, ask what it does that a free chat window plus a document does not. Sometimes the answer is real: a good spaced-repetition scheduler, for instance, though Anki is free on desktop and Android and is still the best one. Usually the answer is a nicer interface around the same free capability.

The honest exception is a paid tier of a major model when you are doing heavy work in a specific week. It is bought by the month and cancellable, and one month during a dissertation is a different decision from a standing subscription all year.