AI and scientific research, open challenge between OpenAI and Anthropic: models versus narrative

The battle between OpenAI and Anthropic now also shifts to AI-enhanced scientific research. If, in fact, a few weeks ago Sam Altman's company launched Prism, a version of ChatGPT designed to make it easier for researchers to search for sources, now Dario Amodei's company is responding with the weapon of dissemination.

Claude's developers have in fact decided to launch a blog to tell about use cases of the model in support of scientific research, as well as practical guides for researchers interested in introducing artificial intelligence into their laboratory work.

Editorial initiative that opens with two first examples. The first is a contribution by Matthew Schwartz, a physics professor at Harvard, who has developed a more accurate way of calculating the C parameter, an indicator used in particle physics to calculate the shape of debris produced by the collision of electrons and protons. All this took 36 million tokens and 40 hours of work, reducing a process that would otherwise have taken a year.

The second is instead a text by Siddharth Mishra-Sharma, a lecturer in the Computing & Data science faculty at Boston University, which takes its cue from the process by which Claude wrote a compiler, i.e. a programme capable of translating machine language into human language, capable of translating the kernel, i.e. the core, of the Linux operating system. To write the required 100,000 lines of code required 2,000 hours of work sessions and $20,000 in API costs. The next step was to employ the same methodology to develop a differentiable version of a solver of the cosmological Boltzmann equations, which model how matter and energy were distributed after the Big Bang.

Not a specific model, as rival OpenAI did, but a storytelling to demonstrate what the role of LLMs can be in supporting scientific research.