The AI tool list is overwhelming and 80% is noise for a learner. Here is the honest map of what to actually invest time in.
Models (the brains)
- Closed APIs: Claude (Anthropic), GPT (OpenAI), Gemini (Google) — best quality, pay per token, zero setup. Start here to build.
- Open models: Llama, Mistral, Qwen — run them yourself, free, private, customisable. Learn once you need control/cost savings.
Frameworks (the tools)
- PyTorch — the standard for building/training deep learning models. Learn this over TensorFlow in 2026.
- scikit-learn — classic ML (regression, trees, clustering). Where every ML learner starts.
- Hugging Face — the "GitHub of models"; download and run any open model in a few lines.
- LangChain / LlamaIndex — glue for building LLM apps (RAG, agents). Useful, but learn the raw API first.
Data & infra
- NumPy / Pandas — array + table crunching. Non-negotiable foundations.
- Vector databases — Pinecone, Chroma, pgvector — store embeddings for RAG/semantic search.
- Jupyter / Colab — notebooks for experiments. Colab gives free GPUs.
Beginner's stack: Python + NumPy/Pandas → scikit-learn → PyTorch → an LLM API + a vector DB. That covers classic ML, deep learning and modern GenAI. Skip the rest until a project demands it.