Daily-LLM: A Visual 19-Family Timeline of Deep Learning and LLM Architectures

What Daily-LLM Is For, and Who It Is Not For

Daily-LLM is a curriculum repository. Its README describes it as a visualized learning path for deep learning and large models, running from a 1986 RNN to 2025 reasoning models and organized into 19 architecture families. The audience is the self-directed learner or interview candidate who wants the evolution of these architectures in one place rather than scattered across blog posts. The repository topics confirm the intent: leetcode, pytorch-tutorial, tensorflow-tutorials, nlp, llm.

It is not a library. There is no package to import, no API surface, no released version. The README lists no releases, and the repository has no homepage. If you arrived looking for a model, a serving stack, or a benchmark leaderboard, this is the wrong artifact. What you get is Markdown, notebooks and a local web page, arranged by family and by year.

The bilingual framing matters. README.md is in Chinese and README_EN.md is the English counterpart, so a reader can follow the timeline in either language. For teams where part of the group reads Chinese documentation more comfortably, that is a real practical advantage over English-only course repositories.

The 19-Family Timeline and the Cross-Cutting Foundations Directory

The organizing mechanism is a numbered sequence of directories, 01-cnn/ through 19-recommendation/, each representing an architecture or paradigm family. The README's timeline table gives each family a key year and a one-line positioning statement. CNN sits at 2012, RNN/LSTM/GRU at 1997 and 2014, Word Embedding at 2013, GAN at 2014, Transformer at 2017, BERT-family pretraining at 2018, GPT and scaling at 2018 to 2020, ViT at 2020, multimodal alignment at 2021, diffusion at 2020, PEFT and LoRA at 2021, RLHF alignment at 2022, MoE and efficient inference at 2023, RAG and agents at 2023, reasoning models at 2024, world models and video generation at 2018, graph neural networks at 2017, speech and audio at 2020, and recommendation systems at 2016.

The numbering is not strictly chronological, and the README does not pretend otherwise. Family 16, world models and video generation, carries a 2018 key year while sitting after the 2024 reasoning family. That is a deliberate grouping choice: the sequence is a reading order, and the years are context. A learner who treats the numbers as a strict timeline will be confused; a learner who treats them as chapters will not.

Alongside the families, foundations/ holds cross-cutting material the README names as activations, backpropagation, optimizers, normalization and attention mechanisms. projects/ holds cross-family practical work. TIMELINE.md is described as an automatically generated year-ordered index, which means it is derived from the family content rather than hand-maintained. The _archive/ directory preserves the old timeline/ and tracks/ content as source material for rewriting, which tells you the structure has already been reorganized at least once.

Installing Daily-LLM and Running the Timeline Web App

The README's quick start is two steps. Clone the repository, then install the Python dependencies. Python 3.8 or newer is the stated requirement, per the badge in the README.

git clone https://github.com/zkywsg/Daily-LLM.git
cd Daily-LLM

After cloning, the README gives a single requirements install:

pip install -r requirements.txt

Be aware of what that pulls in. The requirements.txt pins torch>=2.0.0, numpy, scikit-learn and matplotlib for the core phase, then transformers, datasets, peft, trl and sentence-transformers, then faiss-cpu, chromadb and langchain, then vllm, fastapi, mlflow and wandb. Installing everything at once is a heavyweight operation on a laptop. The README anticipates this with a collapsed section offering per-phase installs. For the first two phases, it suggests:

pip install torch numpy scikit-learn matplotlib

For the alignment and fine-tuning phase it suggests peft and trl, and for the systems and production phase it lists sentence-transformers, faiss-cpu, chromadb, langchain, vllm, fastapi, mlflow and wandb. Following the phase split is the practical choice; the full requirements file exists for completeness.

The second entry point is the visualization page under web/, which is a Node project. The README gives these commands:

cd web
npm install
npm run dev

What you should see is a local development server serving a timeline page with two browsing modes, by year and by family. Node pages combine Markdown text with interactive specimens according to the README. The README states the page is currently maintained as a local web page and that a public address will be added after deployment, so there is no hosted URL to visit.

Where Daily-LLM Breaks Down: Dependency Weight and Unfinished Edges

The most concrete limitation is the dependency surface. A repository whose primary content is explanatory Markdown ships a requirements file containing vllm, mlflow and wandb. vllm in particular expects a GPU environment and a supported CUDA stack; installing it on a CPU-only machine or an Apple Silicon laptop is likely to fail or to consume a long build. The README's per-phase section mitigates this, but the default path a new reader copies first is the full install. If you only want to read the Transformer chapter, you do not need any of it.

The second limitation is the web app's deployment state. The README says the timeline page is maintained locally and that an official access address will be added after public deployment. There is no live site. Anyone evaluating the project on the strength of the visualization has to run npm install and npm run dev themselves before seeing anything.

The third is structural churn. The presence of _archive/ holding the previous timeline/ and tracks/ directories, described as material for rewriting family content, means parts of the repository are in transition. A reader who lands on an archived page through a search engine may be reading content the maintainers have already replaced. The README points to TIMELINE.md as the generated index, which is the safer entry point than a stale file path.

Finally, there are no benchmarks here. The repository teaches architectures; it does not evaluate them. Anyone expecting measured comparisons between the families will not find them documented in the README.

Daily-LLM Compared With a Single-Topic Tutorial Repository

The obvious alternative approach is a focused repository: one that does PyTorch tutorials only, or LeetCode solutions only, or a single model family only. The topics list on Daily-LLM includes leetcode and leetcode-solutions alongside pytorch-tutorial and tensorflow-tutorials, which signals breadth as the design goal.

The difference in approach is real. A single-topic repository optimizes for depth in one skill and is easy to finish. Daily-LLM optimizes for the connective tissue between families: why attention replaced recurrence, how pretraining gave way to alignment, where retrieval and agents entered. That is the value of the 19-family numbering and of foundations/ sitting outside the numbered directories. The cost is that no single family gets the exhaustive treatment a dedicated repository would give it, and a learner who wants to master CNNs specifically will likely outgrow the 01-cnn/ directory faster than they expect.

A second alternative is a video course or a book. Those are linear and paced; Daily-LLM is a directory tree you navigate by interest. If you need someone to tell you what to do next, the repository's structure will not do that for you, though TIMELINE.md offers a year-ordered path.

Licence, Maintenance and the Cost of Upgrading

Daily-LLM is MIT licensed. That is permissive: you can reuse, modify and redistribute the content, including commercially, provided the copyright notice and permission notice are retained. Note that MIT covers the repository's own content; the dependencies it installs carry their own licences, and torch, transformers, langchain, vllm, mlflow and wandb are separate projects with separate terms. Checking those is your responsibility, and this is not legal advice.

The last push to the default branch was on 2026-08-24, which is recent. The repository is not archived. There are no releases, so there is no versioned upgrade path: you track the master branch, and any reorganization lands directly on your clone. The _archive/ directory is evidence that such reorganizations happen. If you fork this for a course or an internal reading group, pin your fork to a commit rather than pulling master, because a family directory can be rewritten underneath you.

Upgrade cost is mostly dependency drift. The requirements file uses lower bounds (torch>=2.0.0, transformers>=4.35.0, langchain>=0.1.0) rather than pins, so a fresh install months later can resolve to substantially newer versions than the notebooks were written against. LangChain in particular has moved quickly across its 0.x line, and the README's own install example still shows langchain>=0.1.0. Expect to debug imports if you run older notebooks against a current environment.

Editorial conclusion

Adopt Daily-LLM if you want a structured, bilingual reading path across 19 architecture families and are willing to install dependencies per phase rather than all at once; the requirements.txt pulls in torch, transformers, peft, trl, faiss-cpu, chromadb, langchain, vllm, fastapi, mlflow and wandb together, which is a large footprint for a repository that teaches rather than runs. Skip it if you need a maintained library, an API, or reproducible benchmark numbers, because the README documents no releases and no published results. Before committing, clone the repository, open the family directory closest to your current work, and check whether its Markdown and notebooks are complete enough for your level; then run npm install and npm run dev inside web/ to confirm the timeline page renders on your machine.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. zkywsg/Daily-LLM on GitHub