Curriculum part 2 · after Week 21 · 22 projects in 6 levels

Part 2: the project ladder

The second half of your curriculum, done last. Harder projects in order of difficulty, from rebuilding a tokenizer to the take-home, the flagship research projects and an application. Each is one sharp question with every step spelled out – and each unlocks only when you’ve finished what comes before it.

How unlocking works

1The ladder opens when you’ve finished the whole 21-week plan.
2Levels 1–4 unlock one project at a time: tick every step of a project to open the next.
3Finishing Level 4 opens all eight Level 5 flagships at once – pick one.
4Level 6 opens when you complete one flagship’s week-8 milestone (M1).

Locked projects stay readable so you can see what’s coming.

How to climb it

1One narrow question with a measurable answer beats a big project that never finishes.
2Reproduce first, then change one thing. A replication you can explain is worth more than a novel idea you can’t test.
3Use the smallest model and the least compute that can answer the question; scale up only to confirm the main result.
4Write it up honestly, including what failed. The write-up is the deliverable reviewers read.
5Finish one rung before starting the next. Two or three done well beat six half-done.

Level 0: Your plan projects
P1–P6 from the 21-week plan, plus their lab-grade versions. Everything below assumes these are done.
P1P2P3P4P5P6Open plan projects
Level 1: Rebuild the foundations
Rebuild something everyone uses and check it against the original. Laptop or a cheap rented GPU.
Write the GPT-4 tokenizer from scratch
Can you build a byte-level BPE tokenizer that produces exactly the same token ids as GPT-4’s cl100k_base?
Time
1–2 weeks part-time
Cost
£0
Compute
Your Mac
Unlocks: Finish the whole 21-week plan
Reproduce a model-card score, and explain the gap
When you run a published benchmark on an open model yourself, how close do you get to the model card – and which settings explain the difference?
Time
1–2 weeks
Cost
£0
Compute
Free Colab or Kaggle GPU
Unlocks: Finish “Write the GPT-4 tokenizer from scratch”
Reproduce GPT-2 (124M) from scratch
Can you pretrain a 124M-parameter GPT-2 yourself and match OpenAI’s GPT-2 124M on HellaSwag?
Time
3–5 weeks part-time
Cost
£20–60 for the full run
Compute
Your Mac for debugging
Unlocks: Finish “Reproduce a model-card score, and explain the gap”
Level 2: Small-model research on one GPU
Interpretability and post-training questions you can answer on a 0.5–3B model for under a few hundred pounds.
Find a circuit by hand (ARENA chapter 1)
Can you locate the components that implement a specific behaviour inside a small transformer, and prove it by intervening on them?
Time
3–4 weeks
Cost
£0
Compute
Free Colab GPU
Unlocks: Finish “Reproduce GPT-2 (124M) from scratch”
Trace a behaviour with attribution graphs
For one specific behaviour of Gemma-2-2B, what does its attribution graph say the computation is – and do interventions confirm it?
Time
3–5 weeks
Cost
£0–£50
Compute
Colab GPU
Unlocks: Finish “Find a circuit by hand (ARENA chapter 1)”
Reproduce an SAE finding on a different model, and find where it breaks
Does a published sparse-autoencoder result hold on a different open model – and exactly where does it stop holding?
Time
6–8 weeks
Cost
£50–200
Compute
One rented GPU
Unlocks: Finish “Trace a behaviour with attribution graphs”
Teach a small model to reason with GRPO
How much does GRPO with a verifiable reward improve a 0.5–1.5B model on GSM8K – and what does the model learn to game?
Time
4–6 weeks
Cost
£50–200
Compute
One rented GPU
Unlocks: Finish “Reproduce an SAE finding on a different model, and find where it breaks”
When does verifier-guided search beat best-of-N?
On MATH-500 with a 1B model, at what generation budget does PRM-guided beam search start to beat (weighted) best-of-N – and why?
Time
6–8 weeks
Cost
£100–400
Compute
One rented GPU
Unlocks: Finish “Teach a small model to reason with GRPO”
Level 3: Performance engineering with a public bar
Scoreboards and thresholds set by the labs themselves. Your number is either below the bar or it isn’t.
Beat PyTorch with your own Triton kernel
Can you write a correct Triton kernel for one real operation that beats PyTorch’s implementation – and explain where the speed comes from?
Time
4–6 weeks
Cost
£0–30
Compute
Free Colab T4 to learn
Unlocks: Finish “When does verifier-guided search beat best-of-N?”
Anthropic’s performance take-home
Can you optimise the take-home kernel on its simulated machine below Anthropic’s published 1,487-cycle threshold?
Time
4–8 weeks
Cost
£0
Compute
Your Mac
Unlocks: Finish “Beat PyTorch with your own Triton kernel”
The modded-nanogpt speedrun
Can you reproduce the current GPT-2 (124M) training speed record, then find a change that makes it measurably faster?
Time
6–12 weeks
Cost
£200–1,000
Compute
Rented 8×H100 node, in short sessions
Unlocks: Finish “Anthropic’s performance take-home”
Level 4: Frontier safety replications
Small-scale replications of recent alignment papers. Expensive and open-ended: scope tightly.
Measure chain-of-thought monitorability of an open model
How monitorable is an open reasoning model’s chain of thought on OpenAI’s public evals – and does it change after you train it?
Time
6–8 weeks
Cost
£200–600
Compute
Rented GPU for the model
Unlocks: Finish “The modded-nanogpt speedrun”
Reward hacking and emergent misalignment, at small scale
If a small model learns to reward-hack in a coding environment, does misaligned behaviour generalise beyond the hack – and does inoculation prompting prevent it?
Time
8–12 weeks
Cost
£1,000–3,000 at ~1.5B
Compute
Rented GPUs
Unlocks: Finish “Measure chain-of-thought monitorability of an open model”

Level 5: Flagship research

Eight 5–9 month projects built on research the labs published in 2025–2026. Pick one. Each has a week-8 milestone (M1) that is publishable on its own.

Which one should you pick?

Your background	Pick	Backup
ML / RL research	Reward hacking → broad misalignment on a realistic agentic RL stack	Hack-resistant automated alignment research
Systems + RL	An RL environment that teaches models to write fast GPU kernels, with grading that can’t be gamed	Asynchronous RL for long-horizon tool-using agents
Alignment research	Auditing adversarially trained model organisms	Unverbalised evaluation awareness
Interpretability	An open J-lens for reasoning models	Unverbalised evaluation awareness
Low-level GPU / kernels	Full-stack FP4 training with the Muon optimiser	An RL environment that teaches models to write fast GPU kernels, with grading that can’t be gamed
Not sure	Reward hacking → broad misalignment on a realistic agentic RL stack	–

Rules for whichever you pick

Ship the week-8 milestone publicly, even if it’s a null result.
Write down predictions before running experiments. Use 3+ seeds. Report confidence intervals.
Publish what failed, not just what worked.
Keep the repo reproducible from one command.
Put the finished project at the top of your CV, with links – Anthropic’s careers page asks for exactly this.

Costs are rough estimates for rented GPUs and API usage; fellowships and research credits can cover most of it.

Your PDF

Track A · RL + safety · default pick

Reward hacking → broad misalignment on a realistic agentic RL stack

Does reward hacking in realistic, multi-turn coding RL generalise into broader misbehaviour in open models, which mitigations stop it, and what does grading the chain of thought during RL do to monitorability?

Time: 6–9 months full-time
Cost: $20–40k
Compute: Rented GPUs