AI Setup — Tools for the Cohort
Tools you’ll use during the cohort. Start with the free tiers; upgrade as you grow.
AI-based IDE — Cursor or Claude Code
We recommend Cursor to start (e.g. $20 plan). It’s similar to VS Code (it’s a fork) and lets you see code changes clearly. Claude Code is also strong but is terminal-based.
Use your AI-based IDE for the first milestone: setting up a Quarto blog and pushing to GitHub.
ChatGPT
You can start with the free tier. As your skills grow, a minimum ~$20/month subscription is useful to use better “thinking” models.
Open Router
Open Router gives you a single API to many models. Check whether signup gives you free credits and use it for experimentation.
GPU Compute for Training
When you reach Step 4 — Training fundamentals, you’ll need GPU access to run distributed training jobs (DDP, nanoGPT speedrun, fine-tuning). Two options:
Modal (highly recommended)
Modal lets you run GPU workloads from pure Python — no Docker, no YAML, no cloud console. Define your function, decorate it, and Modal handles the infra. Sub-second cold starts, instant autoscaling, and scale-to-zero billing.
- Free tier: check whether signup gives you free credits — useful for initial experiments.
- GPU access: H100, A100, A10G, and more — no reservations or quotas.
- Why we recommend it: the fastest path from “I have a training script” to “it’s running on 8 GPUs.” The developer experience is outstanding.
E2E Networks
E2E Networks is an India-based cloud with competitive GPU pricing. Useful if you want dedicated instances, longer training runs, or prefer a more traditional VM-style setup.
- GPU instances (A100, H100) available on-demand or reserved.
- Good option if you’re based in India and want low-latency access and INR billing.
Weights & Biases (W&B)
Weights & Biases is the standard tool for experiment tracking during model training. Log loss curves, hyperparameters, system metrics, and compare runs — all from a single dashboard.
- Free tier: generous for individual use and small teams.
- Why it matters: when you’re doing speedrun experiments or fine-tuning, you need to see what changed between runs. W&B makes this effortless.
- The nanoGPT speedrun scripts and Nemotron training recipes log directly to W&B.