Tutorials¶

The series is organized by capability bottlenecks, not by model-release chronology. One question runs through every phase:

Can a system understand a new goal, build a useful model of its situation, plan and act over long horizons, learn from feedback, and improve with decreasing human supervision?

flowchart LR
  A[Represent\ntext · image · video · state] --> B[Predict\n token · latent · future · action]
  B --> C[Act or generate]
  C --> D[Evaluate\nloss · preference · verifier · environment]
  D --> E[Improve\ndata · search · policy · system]
  E --> A

Each page changes one part of this loop. A release earns a permanent row only when it changes a durable concept or supplies a reproducible experiment. Model names that do neither belong in the Frontier Radar.

A page answers: what was missing, the minimal mechanism, one controlled comparison, where it fails, the current judgment, and the next test. Experiments are optional; a page can ship before its implementation exists.

Phase I · Language¶

External anchors: Stanford CS336, Karpathy — Zero to Hero, Hugging Face LLM Course.

No.	Tutorial	Historical bottleneck	Minimal experiment
01	Transformer and language modeling	Recurrent sequence models were hard to parallelize and scale	Train a tiny decoder-only Transformer; separate memorization from generalization
02	Scaling laws, data, and compute	A working architecture did not tell us how to allocate a training budget	Fit a small scaling curve under a fixed compute budget
03	Supervised fine-tuning, preference optimization, and language-model RL	Pretraining did not directly optimize instruction following, human preferences, or task outcomes	Hold the base model and prompts fixed; compare SFT, direct preference optimization, and a small verifier-guided RL update
04	Reasoning, search, and test-time compute	A single forward sample was brittle on multi-step problems	Hold the model fixed; vary sampling, search, and verifier budget
05	Retrieval, memory, and long context	Parameters and context windows were poor substitutes for persistent memory	Compare parametric recall, retrieval, and explicit episodic memory

Tutorial 03 follows the post-training signal from demonstrations through preference data and outcome feedback. Treat specific algorithms as case studies; the durable question is what feedback is available and how reward hacking can make an apparent gain misleading.

Phase II · Perception¶

External anchors: Stanford CS231n, course notes.

No.	Tutorial	Historical bottleneck	Minimal experiment
06	Visual representation and contrastive learning	Language-only models had no direct perceptual grounding	Train a small CLIP-style model and inspect retrieval failures
07	Connector-based multimodal LLMs	Image and language representations were not directly usable by one decoder	Compare pooling, an MLP projector, and query tokens under a token budget
08	Spatial, temporal, and document grounding	Caption alignment did not imply precise perception or reasoning	Build adversarial tests for counting, spatial relations, OCR, and evidence citation
09	Native multimodal and omni models	Separate encoders and objectives limited cross-modal generation and interaction	Compare discrete and continuous representations conceptually and empirically

Phase III · Generation¶

External anchors: Hugging Face Diffusion Course, Stanford CS25.

Open problem: when does a video model become an interactive world model? Attractive video is not enough; the tests are action-conditioned prediction, long-term state consistency, and whether the state is useful for planning.

No.	Tutorial	Historical bottleneck	Minimal experiment
10	VAE, score matching, and diffusion	High-dimensional continuous generation was unstable or low fidelity	Train a toy diffusion model and visualize the denoising field
11	Latent diffusion, DiT, and flow matching	Pixel-space generation was expensive and difficult to scale	Compare objectives and sampling cost on one small dataset
12	Conditioning and controllability	High-quality samples did not guarantee useful control	Measure prompt, structural, and reference adherence separately
13	Video generation and temporal consistency	Image models did not preserve state through time	Test object permanence, identity, causality, and long-horizon drift
14	World models and interactive simulation	Plausible video was not necessarily useful for planning	Compare pixel, latent, and state prediction in a controlled environment

RL bridge¶

Complete this before Tutorials 16–18. Tutorial 03 covers RL as language-model post-training; this bridge is the general machinery for agents, control, and embodied learning.

External anchors: Berkeley Deep RL.

Central question	Minimal comparison
How should an agent improve decisions when actions change future observations and rewards?	In one small environment, compare behavior cloning, a model-free update, and model-based planning; diagnose shift and reward misspecification

Phase IV · Action¶

External anchors: Hugging Face Agents Course, Hugging Face Robotics Course.

Open problem: do VLAs need explicit long-term memory? Success on short tasks does not imply success on tasks that span hours. Separate memory, state estimation, planning, and the action policy.

No.	Tutorial	Historical bottleneck	Minimal experiment
15	Tool-using agents	Models produced answers but could not reliably alter external state	Build an agent with typed tools, traces, budgets, and recovery tests
16	Imitation learning, action representations, and diffusion policies	Language-model post-training and discrete token generation did not explain how to learn high-dimensional, multimodal actions from demonstrations	Hold the demonstration data fixed; compare behavior cloning with single-step, action-chunked, and diffusion-policy outputs
17	Vision–language–action models	Robot policies did not transfer language and visual knowledge cleanly	Train or adapt a small policy in simulation; evaluate task and embodiment shift
18	Planning, memory, and online adaptation	Short demonstrations did not solve long-horizon recovery	Hold the policy fixed while varying planner, memory, and feedback

Tutorial 16 stays on action representations and supervised policy learning. Do not turn it into a second general RL survey.

Phase V · Improvement¶

External anchors: Full Stack Deep Learning for evaluation discipline, Stanford CS25 for current disagreements.

Open problems: can verifiers replace large-scale human feedback? How can self-generated tasks avoid distribution collapse? How can automated discovery prove that progress is real (independent holdout, preregistered metrics, rollback, compute accounting)?

No.	Tutorial	Historical bottleneck	Minimal experiment
19	Synthetic data and automatic curricula	Human-generated training tasks did not scale indefinitely	Generate tasks near the capability boundary and detect diversity collapse
20	Verifiers and process supervision	Self-generated outputs could not be trusted as their own labels	Measure verifier false positives, gaming, and held-out transfer
21	Coding and research agents	Tool use did not imply reliable multi-hour improvement	Run an auditable propose–execute–evaluate–revise loop
22	Discovery loops and open-ended search	Optimization remained confined to objectives designed by people	Search over a bounded method space with preregistered evaluation and rollback
23	Safety and epistemics of self-improvement	A stronger optimizer can exploit an incomplete evaluator	Red-team the evaluator and define stop, audit, and containment conditions

Three systems to reuse¶

Prefer extending these over starting 24 disconnected demos:

Tiny foundation model — begins as a decoder-only Transformer.
Small interactive world — begins as a predictable simulator.
Auditable discovery loop — begins as the research ledger.

Current sprint¶

Complete 01 · Transformer and Language Modeling.
Run Experiment 001 if you want a number behind the memorization-versus-composition claim.
Write one note in the research ledger.