Google DeepMind: AI Research Foundations
by Google DeepMind
Google DeepMind's free 8-course path: build a small language model from n-grams to transformers, then fine-tune Gemma with LoRA.
Overview
AI Research Foundations is Google DeepMind's language-model curriculum. It was developed with pedagogical experts and academics at University College London and is offered free on Google Skills, where the learning path lists eight activities. Each course mixes reading, videos, Colab labs and short assessments, and earns a skill badge; challenge labs let experienced learners skip ahead to the assessment. Course 1 (Build Your Own Small Language Model, ~6 hours) moves from probability distributions and n-gram models to training a small transformer language model. Course 2 (Represent Your Language Data) covers preprocessing, splitting text into characters, words and subword tokens, implementing a byte-pair encoding tokenizer, and embeddings. Course 3 (Design and Train Neural Networks) builds single-layer networks and multilayer perceptrons, tunes hyperparameters, reduces overfitting, computes gradients and trains with Keras. Course 4 (Discover the Transformer Architecture) implements the attention equation and positional embeddings. Course 5 (Fine-Tune Your Model, ~8 hours) formats dialogue data and does full-parameter fine-tuning of the small model and of Gemma. It then has you write LoRA from scratch, fine-tune Gemma with LoRA, and try reinforcement learning as an alternative to supervised fine-tuning. Course 6 is Align Your Model. Course 7 (Accelerate Your Model, ~7 hours) estimates training FLOPs and GPU memory, then uses bfloat16 and gradient accumulation to fine-tune a 4-billion-parameter Gemma on a single GPU. Course 8 is a ~15-hour advanced capstone: a full Gemma 3 LoRA fine-tuning workflow on a problem you choose. The labs use Keras 3 on the JAX backend with Keras Hub; the Gemma LoRA lab, for example, runs on a free Colab T4 GPU. The labs are open-sourced at github.com/google-deepmind/ai-foundations, with code under Apache 2.0 and other materials under CC-BY 4.0. In March 2026 Google.org committed $4 million to support classroom use of the curriculum in Africa.
At a Glance
- Topic
- ML
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- ~60 hours total, self-paced: 8 courses of ~6-8 hours each (per Google Skills listings) plus a ~15-hour capstone
- Provider
- Google DeepMind
- Hands-on
- Yes — code/exercises
- Certificate
- Available
What You’ll Learn
- ✓Compare n-gram models with transformer language models, then train your own small language model
- ✓Preprocess raw text and implement a byte-pair encoding tokenizer from scratch, then train with it
- ✓Design multilayer perceptrons, tune hyperparameters, mitigate overfitting and train neural networks with Keras
- ✓Implement the attention equation and positional embeddings yourself and visualize what attention computes
- ✓Format data for turn-based dialogue and run full-parameter fine-tuning on your model and on Gemma
- ✓Implement LoRA from scratch, then fine-tune Gemma 3 1B with LoRA through Keras Hub on Colab
- ✓Estimate training FLOPs and GPU memory, then use bfloat16 and gradient accumulation to fit larger models
- ✓Plan, build and evaluate a capstone Gemma 3 LoRA fine-tuning project on a real problem you choose
Highlights
- •Built by Google DeepMind with pedagogy experts from University College London: a first-party look at how DeepMind teaches language modeling
- •You build each piece before using the library version: your own BPE tokenizer, attention equation and LoRA come before Gemma
- •Every published lab is open on GitHub (google-deepmind/ai-foundations, 307 stars as of 2026-09-22), and the Gemma LoRA lab runs on a free Colab T4 GPU
- •Each course earns a Google Skills badge, and challenge labs let experienced learners skip straight to the assessment
- •Honest caveats: the stack is Keras 3/JAX rather than PyTorch/Hugging Face, the public repo has no labs for course 6 (Align Your Model), and each course includes responsible-AI modules that some engineers will skim
Who It’s For
Best For
- ✓Software engineers who call LLM APIs and want to understand how the models work inside
- ✓Computer science, math and physics students preparing for AI research or ML engineering roles
- ✓Educators looking for an open, badge-backed language-modeling curriculum with runnable labs
- ✓Practitioners who want a guided first hands-on fine-tune of an open model like Gemma
Prerequisites
- •Proficiency in Python (Google's stated requirement for the path)
- •A technical background in computer science, math, physics or similar; the path is aimed at students in those subjects
- •A Google account for Google Skills and Colab; the Gemma LoRA lab needs no paid GPU
FAQ
What is Google DeepMind: AI Research Foundations?
AI Research Foundations is a free, eight-course curriculum from Google DeepMind on Google Skills. It is for Python-proficient students and engineers who want to understand language models from the inside. You build a tokenizer, attention and LoRA yourself, train a small language model, fine-tune Gemma, and finish with a capstone project you design.
Is Google DeepMind: AI Research Foundations free?
Google DeepMind: AI Research Foundations is free to access.
What level is Google DeepMind: AI Research Foundations for?
Google DeepMind: AI Research Foundations is aimed at a intermediate audience. Recommended background: Proficiency in Python (Google's stated requirement for the path), A technical background in computer science, math, physics or similar; the path is aimed at students in those subjects, A Google account for Google Skills and Colab; the Gemma LoRA lab needs no paid GPU.
How long does Google DeepMind: AI Research Foundations take?
Expect roughly ~60 hours total, self-paced: 8 courses of ~6-8 hours each (per Google Skills listings) plus a ~15-hour capstone. Most learners work through it at their own pace.
What will I learn from Google DeepMind: AI Research Foundations?
You'll learn: Compare n-gram models with transformer language models, then train your own small language model; Preprocess raw text and implement a byte-pair encoding tokenizer from scratch, then train with it; Design multilayer perceptrons, tune hyperparameters, mitigate overfitting and train neural networks with Keras; Implement the attention equation and positional embeddings yourself and visualize what attention computes; Format data for turn-based dialogue and run full-parameter fine-tuning on your model and on Gemma; Implement LoRA from scratch, then fine-tune Gemma 3 1B with LoRA through Keras Hub on Colab; Estimate training FLOPs and GPU memory, then use bfloat16 and gradient accumulation to fit larger models; Plan, build and evaluate a capstone Gemma 3 LoRA fine-tuning project on a real problem you choose.
Topics
Sources
This page was written from 8 sources, 3 on domains other than skills.google.