Daily-LLM: a 19-family timeline of deep learning, wrapped in a Next.js viewer

The problem Daily-LLM picks: a timeline instead of a topic list

Most deep learning repositories are organised by task. You get a folder of classification notebooks, a folder of NLP notebooks, and no indication that the second folder's ideas descend from the first. Daily-LLM takes the opposite axis. The README describes the project as a visualised learning route that traces evolution from the 1986 RNN to 2025 reasoning models, sorted into 19 architecture families. The table gives each family a number, a key year range, and a one-line positioning statement. Transformer is 2017. Diffusion is 2020 onward. PEFT and LoRA are 2021 onward. The audience is a developer who already knows what a neural network is and wants to understand why the field moved from recurrent models to pure attention, then from fine-tuning to adapters, then from answering to reasoning. That is a different reader from someone looking for a first PyTorch tutorial. The repository assumes the reader can place a family in time and wants the connecting logic more than a from-scratch walkthrough.

How the repository is laid out: 19 numbered directories plus cross-cutting foundations

The structure is flat and predictable. Directories run from 01-cnn/ through 19-recommendation/, one per family. Alongside them sit foundations/, which the README describes as cross-cutting basics covering activations, backpropagation, optimisers, normalisation and attention mechanisms; projects/, for work that spans families; web/, the visualisation app; TIMELINE.md, described as an automatically generated year index; and _archive/, which holds the old timeline/ and tracks/ content and is described as source material for rewriting family content. That last directory is worth noting. It means the current 19-family layout is a reorganisation of an earlier structure, and the archive is kept deliberately rather than deleted. For a reader, the practical consequence is that some material in the numbered directories may be newer than the archive and some may not. The README does not state which families have been rewritten and which are still pending. The TIMELINE.md file being generated rather than hand-written suggests a script produces it, though the README does not name that script or its location.

The web viewer is the part written in TypeScript

The repository's primary language is TypeScript, which is a consequence of web/ rather than of the learning material, since the curriculum itself is Python and Markdown. The README gives the local run sequence: change into web/, run npm install, then npm run dev. The page offers two browsing modes, by time and by family, and the node pages combine Markdown body text with interactive gold samples. That phrase, interactive gold samples, is the most interesting claim in the README and also the least specified. There is no description of what a gold sample contains, how many exist, or whether they execute in the browser or are static. The README states the page is currently maintained as a local web page only, with a public address to be added after deployment. So there is no hosted demo to inspect before cloning. If the viewer is your reason for adopting this repository, you are committing to running it yourself to find out what the node pages actually render.

Getting it running: two dependency paths, one Python and one Node

The README gives a clone command, git clone https://github.com/zkywsg/Daily-LLM.git, followed by cd Daily-LLM and pip install -r requirements.txt. It then offers a staged alternative inside a collapsible block, which is more useful than the blanket install because the full requirements file is not enumerated in the README. The stages are: torch, numpy, scikit-learn and matplotlib for the early phases covering prerequisites and the vision line; transformers, datasets and sentence-transformers for the language line and scale plus multimodal work; peft and trl for alignment and fine-tuning; and sentence-transformers, faiss-cpu, chromadb, langchain, vllm, fastapi, mlflow and wandb for the systems and production phase. That last group is heavy. vllm in particular brings GPU-oriented serving dependencies, and installing it on a laptop to read about RAG is a poor trade. The web app is a separate install: cd web, npm install, npm run dev. The README states Python 3.8+ in a badge. Nothing in the supplied material pins exact versions of torch, transformers or vllm, so the staged installs are the safer route if you want a working environment rather than a reading environment.

Where the repository is thin: no releases, no pinned versions, no stated test story

The repository has no retrieved releases. There is no versioned artefact to install, no changelog to read, and no way to know whether the 11-peft-lora/ material was written against peft 0.3 or 0.11. The README does not mention tests, continuous integration, or a verification step for the code samples. Combined with the _archive/ directory holding superseded content, this means the reader carries the burden of checking whether a given notebook still runs against the current library API. The staged pip groups make that worse rather than better, because they name packages without version constraints. A second limitation is language. The README's own badge marks the docs as English and Chinese, but the primary README is Chinese, with README_EN.md offered as a parallel entry. The README does not state whether the per-family Markdown bodies are translated or only the top-level navigation. If you cannot read Chinese, verify the English coverage inside a family directory before assuming the route is usable end to end. A third boundary: this is a curriculum, not a library. There is no import to add, no API to call. If you wanted a package that answers questions about model architectures, this is the wrong shape of project.

Alternatives and the actual difference in approach

Two comparisons are useful. The first is Hugging Face's transformers documentation and its course material. That material is organised by task and by model class, and it is versioned alongside the library it documents, so a tutorial for a given model class tracks the library release. Daily-LLM is organised by historical family and is not tied to any library version. The trade is reversed: transformers docs tell you how to do a thing today, Daily-LLM tells you why the thing exists and what it replaced. The second comparison is a general-purpose course such as fast.ai, which is built around running code from the first lesson and working upward through progressively harder problems. Daily-LLM inverts that. It starts from a 1986 to 2025 timeline and expects you to descend into a family when you want detail. Neither approach is a substitute for the other, and the choice comes down to whether your gap is conceptual lineage or hands-on fluency. If it is lineage, Daily-LLM's 19-family table is more legible than a task-indexed doc tree. If it is fluency, the absence of a pinned environment here is a real cost.

Licence and the cost of keeping it current

The project is MIT licensed, and the README links a LICENSE file at the repository root. MIT is permissive: you can reuse, modify and redistribute the material, including commercially, provided the copyright notice and permission notice are retained. That matters if you want to fork the 19-family structure into an internal onboarding path. It does not give you any rights over the third-party libraries the staged installs pull in, each of which carries its own terms. This is not legal advice; check the LICENSE file and the dependency licences yourself if you plan to redistribute. On maintenance cost, the honest read is that the repository is a moving target by design. A curriculum covering 2024 and 2025 reasoning models has to be revised as those models change, and the staged dependency groups name fast-moving packages (peft, trl, langchain, vllm) without versions. Expect to spend time on environment repair before reading, and expect the newest families to age fastest. The MIT licence means you can freeze your own copy at a commit and accept that it will drift from the upstream, which for a reading map is often the right call.

Editorial conclusion

Adopt Daily-LLM if you want a year-ordered reading map of deep learning families and are comfortable reading Chinese, since the README states the main route is Chinese with an English entry at README_EN.md. Do not adopt it if you need a single installable package, a pinned environment, or a course with verified exercise outputs; the repository ships no releases. Before committing, open 11-peft-lora/ and 15-reasoning-o1-r1/ and confirm the Markdown depth matches what you need, then run the web/ dev server locally to see whether the interactive gold samples load without a public deployment.

Official sources

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