Why are researchers turning to AI models that run locally on their own computers rather than on remote servers? What can these models already do, for example when searching the literature? And where do they reach their limits?
David Brzostowicki took a closer look at these questions for the scientific journal Nature. Privacy, reproducibility, and cost were the reasons he kept hearing. While researching his article, David also talked to Eric, our president and the developer of DEVONthink To Go. After all, many of our customers use DEVONthink to work with important or sensitive documents: medical records, legal files, or their own research results. Data like this often shouldn’t, or legally can’t, end up on someone else’s servers, let alone be used to train LLMs. That’s why DEVONthink supports not only server-based AI models but also local ones, for example through LM Studio or Ollama.
David’s article, “How to Search the Literature with Local AI Systems,” appears in the current issue of Nature (vol. 658, pp. 830–831) and is also available online. To read the full text, you’ll need a Nature account or access through your institution.
And yes, we’re a little proud to have made it into Nature ![]()
