Claude Science

Running a local server and using it in the browser suits us well, since our data can't be sent out freely. A good fit. But people used to RStudio and Jupyter will have to readjust.

On Ubuntu 20.04 the first run failed right away: bwrap version too old. I finally got it running by setting up a separate Conda environment with a newer version. Teachers who rarely touch the command line will likely get stuck here.

I asked for a quick review and it went off writing search scripts, stuck for minutes, and some references it finally gave couldn't be found. The databases lean biology, with almost no way into cross-disciplinary searches. Not smooth.

Two prompts burned the five-hour allowance, then I sank about fifty more dollars in. It really can run for hours on its own doing literature and data analysis, but can't finish the whole job in one go, and money goes faster than you'd think.

In a meeting I casually had it explore one direction. Its half-hour output would take me an afternoon, or two or three people a week or two via the team process. Branching to chase several lines at once saves the most effort.