Weekly Recap (February 27, 2026)

DNA sequencing visualized, AI slop in CS, deep research agents for science, Dario vs DoD, biosecurity, AIxBio RCT, R updates (R Weekly, rOpenSci, R Works), new papers.

Feb 27, 2026

Asimov Press and Evan DeTurk: A Visual Guide to DNA Sequencing.

Nature: How AI slop is causing a crisis in computer science. Submission stats from arXiv are here: https://arxiv.org/stats/monthly_submissions. This seems truly unsustainable.

Fifty-four seconds. That’s how long it took Raphael Wimmer to write up an experiment that he did not actually perform, using a new artificial-intelligence tool called Prism, released by OpenAI last month. “Writing a paper has never been easier. Clogging the scientific publishing pipeline has never been easier” …

PaperQA3: a frontier multimodal deep research agent for science.

PaperQA3 introduces the highest-performance multimodal capabilities available for deep research agents today… Edison Literature and Kosmos can now read figures and tables from over 150M research papers and patents, and will read, parse, and consider hundreds of figures and tables before responding… Additionally, we’ve updated the underlying algorithm of our literature agent (which is open source) to improve its ability to answer more complex questions. The result of these improvements show Edison Literature as one of the strongest deep research agents across benchmarks, beating out current-day frontier deep research agents.

Lior Pachter: The Quickening (or, vibe-porting edgeR from R to Python). See also the paper, Differential analysis of genomics count data with edge*.

Anson Ho: The least understood driver of AI progress. An opinionated guide to “algorithmic progress” and why it matters

Nature: AI is threatening science jobs. Which ones are most at risk? “Data-analysis and modelling positions are already becoming obsolete, but hands-on experimentalists can breathe easy for now.”

Anthropic blog: Statement from Dario Amodei on Anthropic’s discussions with the Department of War. This is what a spine looks like.

It’s 10 PM and an AI Just Did My Students’ Homework. UVA professor of commerce Ryan Wright writes about this new Einstein app1 that is riling up everyone on the socials. I appreciate his insights on the opportunity that tools like this provide, rather than seeing them purely as some existential threat to higher ed. On one hand it’s kind of shocking that Cheating-as-a-Service is a successful business model these days. On the other, this looks like a matter of degree, not kind. Chrome already has built-in autobrowse that’ll take all your canvas quizzes for you. You can do similar things with Perpexity’s Comet browser or OpenAI’s Atlas browser. We have to come to grips with and acknowledge that if students want to cheat themselves out of an education, they will. I think at some point you have to appeal to students’ desire to learn, and if that doesn’t work, to their desire to get a job when they graduate. Years ago I taught a graduate biological data science course, graded as CR/NC instead of AF. I told everyone on day 1 they’ll get a CR regardless of whether they came to class, did the homework, whatever. But, as an employer / hiring manager, I’ve never once looked at a GPA or transcript. Ever. I’ll know within minutes if you’ve AI’d your way through courses and don’t actually know anything. Presumably you want a job when you’re done here and you’ll never land one if you’ve cheated yourself out of learning.

Matt Lubin: Five Things: Feb 22, 2026: Biology uplift, DOE robo-labs, Web 4.0, Seedance 2.0, Arc Institute paper.

Celia Ford: How worried should we be about AI biorisk? The barriers to bioattacks are hard to identify — and it’s even harder to know whether AI is reducing them.

UK AI Security Institute: Boundary Point Jailbreaking: A new way to break the strongest AI defences.

Figure 1 from the AISI blog post above: Adversarial Prefixes. Normally, harmful requests are flagged by auxiliary classifiers to prevent models from responding (left). BPJ develops adversarial prefixes (red prepended text) that evade the classifier, enabling the extraction of harmful information from the model (right). Note, this request is an illustrative example from a public dataset—the questions we use are typically much longer and more in depth. The prefix is similarly illustrative and is not a BPJ-found attack (see the main paper for a full example in a simple setting).

Lucas Harrington: Wealthy Individuals Funding Science is Good for Everyone.

Luca Righetti, METR blog: Five lessons from having helped run an AI-Biology RCT.

Nature: Biotech investor set to lead NSF. If confirmed, Jim O'Neill would be the first non-scientist or non-engineer to lead NSF.

The Last Temptation of Claude. By outsourcing to an AI tasks that require thinking and deliberation (hitting the Gemini button to draft an email before you’ve decided what you think, using Claude to summarize the paper instead of reading it), you erode the very skill you’d need to recognize you’re doing it.

R Weekly 2026-W09 Git commits, Pick a License, and How to choose the best LLM using R and vitals.

rOpenSci News Digest, Feb 2026: Champions program, policy updates for GenAI, new packages, software review,

Joe Rickert / R Works: January 2026 Top 40 New CRAN Packages: Artificial Intelligence, Computational Methods, Data, Dynamical Systems, Ecology, Economics, Epidemiology, Finance, Genetics, Genomics, High Performance Computing, Mathematics, Machine Learning, Medical Application, Networks, Statistics, Time Series, Utilities, and Visualization.

NYT: More Than Half of Teens Use Chatbots for Schoolwork. A new study from the Pew Research Center finds teenagers think chatbot-assisted cheating has become “a regular feature of student life.”

New papers & preprints:

The Einstein app was originally available at companion.ai/einstein, but the app was recently taken down. Not by a takedown request from Canvas, but because the app illegally used Albert Einstein’s face (which is apparently owned by someone).

Paired Ends

Comments

Ready for more?