Beginning around March 2026, I started using large language models (LLMs) for scientific research. I think there's no going back: it's a transformative technology to which our scientific community and practices urgently need to adapt. Here I collect some tips and links for scientists who want to learn what current AI models can do for scientific research and how to explore them, with practical tips on how to start, and some links to current discussions and concerns about their impact on theoretical fields.
Other colleagues at UFF and I applied for and obtained 1 year of free Anthropic Pro plans, for up to 25 user seats per Principal Investigator, through this Claude for scientists programme. The PI can convert some (or all) of the seats into a discounted Max plan (R$82.50/month), which includes 5x more token use as well as access to top-tier models such as Claude Fable 5.2 (as of Nov. 2026). It's been working great for me, and I'm also providing access to students in my group.
I also applied for and obtained access to top-tier Google AI models through the Google Antigravity Gemini for Science programme (featuring Gemini Carbon as of Nov. 2026); this will be free for at least 6 months. Unfortunately it's a single licence for myself, and it doesn't cover students or collaborators. OpenAI has set up a similar programme for researchers, now closed but with new applications opening come Nov. 23rd, 2026.
Free AI models available online are great at giving instructions to help you set up the system – try Google Gemini, DeepSeek or ChatGPT. Here are a few tips based on my experience.
1- Pay for access to a frontier model. Even a 20-dollar-a-month subscription will give you access to models that are much better than the free tier, and that can work more autonomously as coding assistants. As I mentioned above, I've been using both Claude and Gemini models through programmes offering cheap or free access for scientists. Be sure to check the privacy settings, if you don't want to have your data used for model training.
2- Instead of using web-based interfaces, use a coding environment. I use the Claude plug-in for Microsoft Visual Studio Code, and Google's Antigravity editor (itself based on VS Code) for Gemini. Besides letting you interact with the agents inside the editor, this gives the agents practical autonomy to download, install and run programs, such as compilers, computer algebra systems, SDP and other convex geometry code, etc. This corresponds to the "Auto" setting in Claude, with other models having equivalent settings.
3- Learn to direct the agent so it can keep working on a problem or project across multiple sessions. The models are already prepared for this; it's a matter of specifying what you need. In my case, I start by uploading reference papers and telling the agent to keep a log of progress in a human-readable .tex file; sometimes I ask it to interview me to learn what I want from the project, and to write some reference text for its own project memory. When the context window is filling up, I ask it to make sure the project memory is up to date (simple .md text files, readable by both the bot and humans), and to write a prompt for resuming in a fresh context window. Copying and pasting that prompt into a fresh session (or pointing the fresh session to the handoff prompt stored as a file) lets you continue work on the same project over multiple sessions. For reference, some of my projects have taken close to 100 sessions, and some have needed to be split into sub-projects.
4- Managing token use and context windows. A token is the unit of model use; it corresponds roughly to a fragment of a word, counted both in what is sent to the agent and in what it returns. If you interact for many rounds with the same agent instance, and especially if the agent's outputs are large (e.g. programs, draft text), your context window will fill up – these days a typical frontier model has a context window of one million tokens. As this limit is approached, the model becomes less efficient and has to compact the chat history, losing some of the earlier discussion. That's why, for longer and more ambitious work, it's important to set up a memory system and to open fresh sessions that import only the key findings of previous ones.
5- Error correction is key: have new, independent sessions criticise and find errors in previous outputs, also using different models. I haven't automated this multi-agent approach, though some settings (like Claude Code's "ultracode" mode) automatically spawn several agents to work on tasks in parallel, e.g. drafting something while doing a literature search. Of course, having multiple agents working at the same time uses up tokens much faster and, if they are not well coordinated, can result in document version collisions, etc.
6- The quality of the scientific text the models write tends to be a bit disappointing, falling short of their performance in numerical work and code (clearly superhuman), mathematical reasoning (excellent) and broad knowledge of the literature. Expect to have to revise the text. In my current experience (Nov. 2026), Claude Fable 5.2 is OK and Gemini Carbon is the best (although I've done little work with Carbon so far). Many colleagues swear by ChatGPT, but I haven't tried its best tier.
7- Setting up your AI coding agent is easier than it sounds. There are many tutorials online, and the models themselves can walk you through it in easy steps via their web interface, in plain human language and with infinite patience. For example, Claude helped me find a cloud computing provider offering free access to a small 2-core machine, set it up for me, and now runs compute jobs there autonomously, like a software engineering intern under my supervision.
8- While here I only discuss proprietary, subscription-based models, in the Brazilian context in particular it would be important to consider setting up local models, so as not to be tied to a foreign company's interests, and also because of sensitive data. There are many open-source and open-weight models available, of Chinese, American and European origin (see e.g. Europe's Mistral models). Small models of this kind can be run on a local server, while larger ones require beefed-up machines with expensive GPUs. Their performance trails that of frontier models by perhaps six months to a year, although this is changing rapidly. For an example of what a university can do, see the University of Illinois's Lumen service, which runs several open-weight models on university servers for faculty, staff and students, keeping all data under university control. In Brazil, there are initiatives such as LNCC's Manacá-1B, a small, fully open and reproducible model trained from scratch for Brazilian Portuguese, and UnB's AI and supercomputing lab (LmiSUP), which offers researchers access to its cluster through internal calls. Unfortunately I don't yet have any experience with this to share.
9- Check out some materials on the practical use of AI in math, computer science and theoretical physics:
9.1- Talk by theoretical physicist Tobias Osborne at UCL, describing his workflow using Claude Code (March 2026).
https://www.youtube.com/watch?v=UtJUWI4r-4k
9.2- "How to train your slop cannon", a short guide to using large language models for research in the natural sciences, by quantum information and quantum foundations scientist Jonathan Oppenheim with Dani Loader and Tobias Osborne (September 2026). Among other workflows, it shows how to have AI produce structured proofs that can be checked step by step, all the way to formal verification. There's also a GitHub repository with the paper and more materials.
https://superposer.substack.com/p/how-to-train-your-slop-cannon
https://github.com/Open-Science-Ledger/how-to-train-your-slop-cannon
(I haven't used this approach, but it looks useful and practical.)
9.3- Recent paper (October 2026) benchmarking AI agents on 50 unpublished, research-level problems in theoretical physics. The best agent, OpenAI's GPT-6-Astra (running in Codex), got the correct final answer for 49 of the 50. Interestingly, in some cases where the AI's answer at first seemed wrong, it turned out that the human experts had been mistaken.
https://arxiv.org/abs/2610.16733
9.4- A machine-checked proof, in the Lean proof assistant, that SICs exist in every dimension (Zauner's conjecture), by quantum information scientist Māris Ozols (October 2026). It assembles recent results by several authors, with AI agents (mostly Claude Fable, with GPT models handling the Lean proofs) doing the orchestration and much of the formalization, as described in the repository.
https://github.com/marozols/SIC-existence
The impact of AI is being felt in many theoretical research communities, most strongly in math and theoretical computer science. I've been following the discussion online, as it plays out in seminars, blog posts and other publications. The scientific community will have to adapt and change quickly, as will society more broadly. The motivations for and the process of doing science and math, the way credit is given for different contributions, the way our activities are funded: all of this will need to change, and the speed of change is far beyond what the community can comfortably adapt to. So it's urgent to reflect on and discuss this, and to act to soften the impact on early-career researchers.
Here are some thoughtful takes on this:
1- Fields medallist Terry Tao has been thinking a lot about this: giving talks, writing, coordinating activities and inviting guest posts on his blog. Lots of different takes, and useful links to grassroots action. https://terrytao.wordpress.com/tag/ai/
2- Grant Sanderson is a master of mathematical exposition, popular on the web as 3Blue1Brown. In this talk (October 2026) he reflects on how the whole mathematical enterprise can adapt to the impact of AI, a discussion that is also relevant to other areas.
https://www.youtube.com/watch?v=LKrwhjOI8Nc
3- "How to respond to the automation of research", a talk at the Simons Institute (Berkeley) by Caltech researcher Hsin-Yuan Huang (July 2026), on credit attribution and current AI capabilities for research, in the context of quantum information science. https://simons.berkeley.edu/talks/hsin-yuan-huang-caltech-2026-07-24
4- "AI and TCS: the next six months", a report by the Simons Institute's AI + TCS Working Group, which met in September 2026. It proposes twelve concrete near-term actions for theoretical computer science, such as a standardized AI-methodology section in every paper. https://simons.berkeley.edu/news-publications-videos/ai-tcs-working-group
5- To keep up to date: Transformation, Quanta Magazine's weekly bulletin on the latest news in math and AI, and what it means for mathematics as a whole. https://www.quantamagazine.org/updates/transformation/
I hope some of this is helpful! Feedback is welcome – please use my institutional email. I often post on these topics on Bluesky, feel free to follow me there.