Felipe Herrera

Deepfakes are everywhere, but digital forensics investigators are fighting back. Learn more: scim.ag/4deiMOT @NewsfromScience

These vague posts about RSI are insufferable, but there really has been a step change with frontier agents. Here are some observations I've seen in 2026: - One of my former PhD students gave an agent an old ML paper we wrote and the agent trained a better model without intervention. - I was able to port my chemistry textbook from tensorflow to pytorch with minimal effort using claude code - I asked codex to convert a reveal.js presentation from HTML to PPTX and it did it with one correction. Not really research, but a task that I didn't even think was possible - Our internal roadmap at @EdisonSci no longer slips and we're sometimes completing projects early. Maybe we just got better at planning. - we've internally done things like autoresearch and found similarly that the proposed research ideas and execution are now compelling enough to just run without human intervention. - both opus 4.6 and 5.4 are able to browse the web much better and have the tenacity to do tasks like "visit every biotech with a TPD, download their pipeline, and add the indication area to a spreadsheet." - One of my PhD students was able to set-up GRPO training of a multi-turn reasoning model for relaxing crystal structures with a simulation as the reward function. This would normally require a research team and a lot of time, but he did this while managing a newborn and writing his thesis. - I also notice that the kind of code coming out of LLMs now is code I could never have written. Like its use of heredoc/variable expansion are beyond my CLI brain. Personally, these observations add up to a big change happening right now and we're on the precipice of some big changes in how research is done.