Tutoring & Education

1. Install Ollama → ollama pull llama3.1:8b (5 GB) or ollama pull qwen-3-30b-a3b (18 GB — MoE, stronger reasoning for tutoring).
2. Tutoring works best with a system prompt that constrains the model to Socratic teaching:

ollama run llama3.1:8b
/set system "You are a patient, encouraging tutor. Never give away the answer directly. Instead: (1) Ask what the student already knows, (2) Guide them with hints and questions, (3) Confirm understanding before moving on, (4) Praise effort, not just correctness. Use the Socratic method."

1. Student: "I don't understand how binary search works." Model: "Great question! Let's start with something familiar — when you look up a word in a printed dictionary, do you start on page 1 and read every word? [No...] Right! What do you do instead?"
2. First tutoring interaction in 2-5 seconds. The model adapts to the student's level and asks guiding questions.
3. For STEM tutoring: use reasoning models (DeepSeek R1 distillations) with Socratic prompting for math/physics problems — the CoT trace helps explain the reasoning steps.
4. For language tutoring: ollama pull aya-expanse:8b — multilingual, patient, can explain grammar rules in the student's native language.

Tutoring is VRAM-light. Llama 3.1 8B runs at 50-80 tok/s on a used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb) — fast enough for real-time conversation. For a homeschool family or self-study setup: $400 handles all K-12 tutoring subjects with an 8B model. Pair with Ryzen 5 5600 + 16 GB DDR4 + 512 GB NVMe. Total: ~$360-405. For CPU-only: Llama 3.2 3B at 20-40 tok/s on a $300 laptop handles basic tutoring conversations. Tutoring is a use case where latency matters (students hate waiting) — the GPU makes conversations feel natural.

Used RTX 3090 24 GB (~$700-900, see /hardware/rtx-3090). Runs Qwen 3 30B MoE at 25-40 tok/s or DeepSeek R1 Distill Qwen 32B at 15-25 tok/s — these models tutor advanced STEM topics (linear algebra, organic chemistry, algorithms) with far fewer errors than 8B models. For a tutoring platform serving 10-50 concurrent students: the 32B model provides reliable Socratic guidance without hallucinating incorrect explanations. Total: ~$1,800-2,200. Tutoring quality jumps at 32B — the model catches when students make subtle errors (sign errors in algebra, misunderstanding of theorems) that 8B misses.

The mistake: Using a standard chat model without a tutoring system prompt, resulting in the model giving direct answers ("The answer is 42") instead of teaching the student to find the answer. Why it fails: Chat models default to "helpful assistant" mode = give the answer. This is anti-tutoring. The student copies the answer, learns nothing, and becomes dependent on the AI to solve every problem. The fix: Always set a Socratic system prompt. The prompt should instruct the model: "Never give the full answer. Break the problem into steps. Ask the student what they've tried. Give a hint, wait for their attempt, then give the next hint. Only reveal the answer after the student has demonstrated understanding." Test the prompt: give the model a math problem and see if it resists giving the answer. If it blurts out the solution, iterate the prompt. A good tutor talks 30% of the time and listens 70%. A bad tutor (default chat model) talks 100% of the time. Your system prompt is the difference.