Ollama vs ChatGPT - Going Offline During Exams

· 3 min read
offline · cloud

Some proctored exams block all network egress except the proctor's endpoints. ChatGPT obviously fails in that environment. Ollama, running entirely on your Mac, does not. The trade-off is answer quality; the gain is reliability under hostile network conditions.

A proctored exam app may install a network-policy profile that blocks outbound connections to known AI endpoints (chat.openai.com, claude.ai). When that happens, ChatGPT in the overlay shows a connection error. Ollama does not depend on the network at all - prompts and answers route through localhost only. The downside is that on a typical M1 Mac, a 7B model is roughly GPT-3.5 quality, not GPT-4. For most homework explanations that's adequate; for hard problems it isn't. If your exam policy blocks AI endpoints aggressively, Ollama is the fallback.

Key points

How it works

┌── Ollama (offline) ──────────────┐  ┌── ChatGPT (online) ────────────┐
│  Localhost:11434                 │  │  https://chat.openai.com       │
│  Survives blocked exam network   │  │  Blocked if proctor allowlists │
│  ~30 tok/s on M-series           │  │  Faster but blockable          │
│  GPT-3.5-class quality           │  │  GPT-4-class quality           │
└──────────────────────────────────┘  └────────────────────────────────┘

Compatibility on Mac

Network requiredOllama: no / ChatGPT: yes~
Survives blocked networkOllama: yes / ChatGPT: no~
Answer qualityOllama: GPT-3.5 / ChatGPT: GPT-4~
Setup timeOllama: 5 min / ChatGPT: instant~

Common questions

How do I install Ollama before my exam?

Download from ollama.com, run installer, then `ollama pull mistral` from Terminal. Five minutes total. Run `ollama serve` to start the local server.

Does the overlay need extra config for Ollama?

Configure the overlay URL to http://localhost:11434/ or use the built-in Ollama mode if your version supports it.

Will running Ollama drain my battery?

Yes - LLM inference is GPU-heavy. Plug in for exams longer than 90 minutes.