Track 02 / AI / SI
AI / SI
How modern AI is built, trained, evaluated and constrained, from gradient descent to the superintelligence question.
- Duration
- 4 weeks
- Pace
- Self-paced
- Modules
- 8
- Live sessions
- 8, optional
- Price
- $24.99 founding, $49.99 later
- Seats
- Limited to 25
Starts in December and runs into the first week of January, with a break over the holidays.
Free to join. Nothing to pay now.
01 / The climb
What a joule buys once it is spent on training.
K = 0.73
02 / Who it's for
Who it's for
- 01Engineers who use AI tools every day and want to know what is underneath.
- 02Researchers and analysts who work on AI strategy or policy and need the technical basis.
- 03Builders who are done calling an API and want to train and evaluate models themselves.
03 / Prerequisites & time
Prerequisites and time
- 01Python you can write without looking things up.
- 02Calculus and linear algebra at first-year university level. There is a refresher.
- 03Basic probability and statistics.
- Duration
- 4 weeks
- Pace
- Self-paced
- Note
- Starts in December and runs into the first week of January, with a break over the holidays.
04 / Syllabus
Four weekly blocks
Week 1
M01
Learning from data and backpropagation
Models, loss functions, gradient descent, generalisation, layers, activations and automatic differentiation.
Proof of work
A small autodiff engine and a network trained with it, checked against numerical gradients, with a held-out evaluation report.
M02
Sequence models and attention
Tokenisation, embeddings, attention and the transformer architecture.
Proof of work
A minimal transformer implemented and trained on a small text corpus.
Week 2
M03
Training and scaling language models
Pre-training objectives, data pipelines, mixed precision, scaling relationships, compute-optimal training and energy per token.
Proof of work
A reproducible training run across a few model sizes, with a fitted scaling curve and a compute and energy estimate.
M04
Post-training and alignment methods
Fine-tuning, preference learning, reward modelling and where each breaks.
Proof of work
A fine-tuned model with a documented preference-learning loop and failure-case analysis.
Week 3
M05
Evaluation and interpretability
Benchmarks and their flaws, probing, circuits-style analysis and red-teaming.
Proof of work
An evaluation harness plus an interpretability investigation of a behaviour in your own model.
M06
Agents and tool use
Planning, tool calling, memory, long-horizon tasks and measuring reliability.
Proof of work
A tool-using agent with a test suite of tasks and measured success and failure modes.
Week 4
M07
Superintelligence and safety
Capability trajectories, specification and misalignment, oversight, governance and open problems.
Proof of work
A written threat model for a hypothetical advanced system, with proposed mitigations and their limits.
M08
Capstone: a verified AI system
Design, build and evaluate a system of your choice, and defend it.
Proof of work
A reproducible repository and evaluation report, reviewed with the platform and, if you want, in a live session.
05 / Format
Format
01
Self-paced
Four weeks. Work whenever suits you.
02
Eight optional live sessions
A kickoff session and seven more with a human expert in the field. You can come to all of them, some, or none.
03
Small cohort
Limited to 25 people per course.
06 / Outcomes
Outcomes
- 01A transformer language model that you implemented and trained.
- 02A habit of estimating the compute and energy cost of a training or inference job.
- 03Evaluations that test what you say they test.
- 04A technical basis for talking about advanced AI and its risks.
07 / FAQ
FAQ
Is this about using AI tools?
No. It is about how the systems are built, trained, evaluated and constrained. Tool use gets one module.
Do I need a GPU?
Not for most modules. Where a bigger run helps, we will point you to a small, cheap setup.
Does the track take a position on superintelligence?
It lays out the main arguments and the open problems, and asks you to build your own view from them.
When does it start?
The first cohort starts in December and runs into the first week of January, with a break over the holidays. It is limited to 25 people.
02 / AI / SI
Reserve your place
Free to join. Nothing to pay now. Founding price $24.99 per course. Later cohorts are $49.99.