This post, published on 23 March 2026 under Anthropic's Science label, announces a new blog about AI and science. It is an announcement with no data, so this note is short. It matters for this library mainly as the frame for the two pieces published alongside it, Vibe physics: The AI grad student and Long-running Claude for scientific computing, and for the questions it raises about research practice.
What the post says
The motivation
The post ties the blog to Anthropic's mission of speeding up science, citing Dario Amodei's essay Machines of Loving Grace and its idea of a "compressed 21st century," in which decades of progress happen in a few years. It points to early examples, each linked to an outside source: AI helping mathematicians find new proofs, individual researchers running computational analyses that once needed a team, and biologists identifying gene relationships across datasets of millions of cells.
It then states a change in what AI does: "Just as computers took on the task of computation, AI is now taking on parts of cognition." One side effect, it says, is that work that used to need years of specialized training can be done more quickly and cheaply.
Open questions about research practice
The post lists questions it calls sociological, about the practice of science and its institutions:
- what research apprenticeship should look like;
- how to keep trust in the literature when AI becomes central to producing it;
- what it means to be a scientist when the bottleneck "shifts from execution to management."
It also qualifies its own optimism. It describes AI's scientific capabilities as "still in beta": models can seem superhuman at some parts of the scientific workflow but can hallucinate results, be overly sycophantic, and get stuck on problems a practitioner would find trivial. It quotes the mathematician Timothy Gowers, writing that "our research is greatly sped up by AI but AI still needs us."
What the blog will publish
Three types of post are planned (section "What we'll cover"):
- Features: articles on a specific result or line of work, explaining both the science and AI's role in it, by Anthropic researchers and guest contributors.
- Workflows: practical guides for researchers using AI in the natural and formal sciences.
- Field notes: roundups of results, tools and open questions across the field.
The two launch pieces are Matthew Schwartz's account of supervising Claude through a theoretical physics calculation, and a tutorial on running long autonomous tasks for scientific computing.
Science programs at Anthropic
The section "Science at Anthropic" lists the AI for Science program, which gives API credits to researchers on high-impact projects; Claude for Life Sciences, with partnerships across research institutions, pharmaceutical and biotechnology companies; and Anthropic's role as a core partner in the US Genesis Mission, described as a multi-billion-dollar initiative across industry, academia and government. It adds that many Anthropic researchers working on the models' scientific capabilities come from biophysics, chemistry and neuroscience.
Limitations and open questions
- An announcement, not a study. The examples of AI-assisted discovery are linked to outside sources and are not evaluated in the post; none was read for this note.
- Questions without answers. The post poses its questions about apprenticeship, trust and scientific identity but does not say how the blog or Anthropic will address them.
Research questions
- How do AI-assisted papers disclose the AI's role? The post asks how trust in the literature can be kept. Surveying acknowledgment and methods sections of papers that report substantial AI assistance, and comparing them with journal and preprint-server policies, would show what readers are currently told.
- What do supervisors check when AI does the calculation? The post's "management" framing suggests a shift in the scientist's role. Logging the checks that researchers run on AI-produced work across several projects would describe that role concretely; the Vibe physics post gives one account.
- Which failure modes named in the post dominate in practice? It lists hallucinated results, sycophancy and getting stuck on trivial problems. Classifying the errors found in a set of AI-assisted research projects into these categories would show their relative frequency.
- How does the blog's mix of post types develop? Counting features, workflows and field notes published over the blog's first year, and the share written by outside contributors, would show which of its stated aims it pursues.
How it connects
- Vibe physics: The AI grad student and Long-running Claude for scientific computing were published with this post and are named in it as its first two pieces.
References
- Anthropic. (2026, March 23). Introducing our Science Blog. https://www.anthropic.com/research/introducing-anthropic-science
- Amodei, D. Machines of Loving Grace. https://www.darioamodei.com/essay/machines-of-loving-grace. Linked from the post and cited as the post describes it; not read for this note.