Connected Papers vs nanochat

Connected Papers and nanochat both help people understand research, but they solve very different jobs. Connected Papers is a literature-discovery product: give it a seed paper and it maps related work, prior work, and derivative work. nanochat is Andrej Karpathy’s open-source LLM training reference: a readable codebase for learning how tokenization, pretraining, supervised tuning, RLHF, evaluation, inference, and a minimal chat UI fit together.

That means this is not a normal “which research assistant is better?” comparison. Pick Connected Papers when the problem is finding the right papers. Pick nanochat when the problem is understanding how language models are trained.

Quick Answer

Choose Connected Papers for literature review scoping, citation-neighborhood mapping, and quickly finding foundational or adjacent papers from one known seed. Choose nanochat if you are an engineer, student, or educator trying to learn the full LLM training pipeline from inspectable Python code. Connected Papers is a usable research discovery app. nanochat is a learning and experimentation repo, not a hosted research assistant.

Decision Snapshot

Where Connected Papers Wins

• It is an actual research workflow tool. Paste a DOI, title, URL, or arXiv ID and get a graph you can use immediately during literature review.
• It reveals the shape of a field. Prior Works and Derivative Works help separate foundational papers from newer follow-on work.
• It is easier for non-engineers. No local setup, no GPU rental, no Python environment, and no model-training knowledge required.
• It pairs well with other research tools. Use it before Elicit, Consensus, Scite, or Semantic Scholar when you need to understand what to search next.
• Its paid tier is inexpensive. The main reason to upgrade is graph volume, not a complicated platform bundle.

Where nanochat Wins

• It teaches the machinery. nanochat exposes the full LLM pipeline in a compact repo instead of hiding training behind a managed product.
• It is free and inspectable. The license and source make it useful for courses, workshops, and serious self-study.
• It covers more than pretraining. The repo includes tokenization, pretraining, supervised fine-tuning, and a minimal chat surface.
• It is better for experiments. Researchers can modify code paths directly, run ablations, and see how choices affect small models.
• It complements nanoGPT. nanoGPT is the smaller pretraining reference; nanochat is the broader system walkthrough.

Key Differences

The most important difference is product shape. Connected Papers is a hosted application for people doing literature review. It assumes you already have at least one relevant paper and want to map the surrounding neighborhood. The value comes from visual structure: clusters, influence, prior work, and derivative work that may not surface through keyword search alone.

nanochat is not trying to help you find papers. It is a reference implementation for building and studying small language-model systems. The value comes from readability and completeness: you can trace the path from tokenizer to training loop to alignment step to inference without jumping across a production-scale framework. That is powerful for learning, but it is the wrong tool if your deadline is “find me the best papers on this topic by Friday.”

Pricing also works differently. Connected Papers has a normal SaaS-like free and paid graph workflow. nanochat is free open source, but meaningful training experiments still cost compute, setup time, and engineering attention. The sticker price is lower; the operational effort is higher.

Who should choose Connected Papers

Choose Connected Papers if you are entering an unfamiliar research area, preparing a related-work section, checking whether a paper sits in the right conceptual neighborhood, or building a reading list from one trusted seed. It is especially useful for researchers who think visually and want a quick map before deeper extraction or synthesis.

Who should choose nanochat

Choose nanochat if your goal is educational or technical: learning LLM internals, teaching a model-training course, running small ablations, or comparing the clean reference code against heavier training stacks. It is best for engineers and students who want to understand systems, not for general researchers who want a polished app.

Bottom Line

Connected Papers is the practical pick for literature discovery. nanochat is the practical pick for learning how LLM training works. They can coexist in a technical research workflow, but they are not substitutes: one maps papers, the other teaches model-building.

FAQ

Which is cheaper? Connected Papers has a free tier and a low-cost Academic plan. nanochat is free open source, but meaningful experiments may require paid GPU time. Treat nanochat’s real cost as compute plus learning time, not subscription price.

Which has better output quality? They produce different outputs. Connected Papers is judged by whether its graph surfaces useful related papers. nanochat is judged by whether its code clearly teaches the LLM pipeline and supports small experiments.

Can I use both? Yes, but usually for different phases. A researcher might use Connected Papers to map the literature around LLM training, then use nanochat to understand and modify a training pipeline.

Sources

• Connected Papers website
• nanochat documentation