Why not ChatGPT, or a coding agent like Claude Code?
General AI tools are built for everything. Lemma is built for science.
Chat assistants and coding agents are excellent general tools. Lemma adds what scientific work needs: the software, the compute, the checks and the sources, ready in your browser.
General AI chatChatGPT, Gemini, Claude
Coding agentClaude Code, Codex, Cursor
Lemmafor science and engineering
Scientific software
AI chatA general sandbox
Coding agentYou install and maintain it
LemmaReady: 92 Python packages, plus Julia, LaTeX, Lean, MEEP and ngspice; Tidy3D, Optiland, QuTiP, Qiskit, gdsfactory, pymatgen, PyTorch…
Compute
AI chatLimited sandbox
Coding agentYour laptop or your own cluster
LemmaFrom 1 CPU · 4 GB up to 20 CPUs · 64 GB, or an NVIDIA H200 GPU with 192 GB. Long jobs keep running in the background
Checks its own work
AI chatWhen you ask
Coding agentWhen you write the tests
LemmaBuilt in: Plan & verify mode, and a verification & validation report for every task
References
AI chatCan cite papers that do not exist
Coding agentNot its job
LemmaEvery reference checked for existence, correctness and relevance
Literature
AI chatWeb search
Coding agentWeb search, if set up
LemmaA curated knowledge graph of scientific papers, plus the web
Formal proof
AI chatNo
Coding agentIf you set up Lean
LemmaComing soon Lean 4 checks of key results
Open-weight models
AI chatNo: the provider's own models
Coding agentSome tools, if you set them up
LemmaAvailable: run Lemma on open-weight models for extra data protection
What you get
AI chatText in a chat
Coding agentCode files in your project
LemmaInteractive notebooks with sliders and 3D views, annotated PDFs, your manuscript written and compiled in LaTeX, BibTeX
Who can use it
AI chatAnyone
Coding agentPeople at home in a terminal and git
LemmaAnyone, in the browser: no setup, no installs
The evidence
Lemma finds more of the errors in real papers.
93%of author-confirmed errors found in 26 maths and CS papers. General AI models: 48–79%.
84%in 50 corrected or withdrawn physics papers. Gemini 3.1 Pro: 52%.
103/104references with a known answer judged correctly.