On the topic of capabilities, agentic research assistants have come a long way in 2025. Elicit & Edison Scientific are ones to watch, but they still struggle to adequately cite the sources for the claims in the reports they generate our conclusions they come to. Contrast this with Deep Research models, which include nonsense from press releases and blog posts, even when explicitly asked to exclude these, though they have improved significantly too. Progress, but we’re still a long way from what a average PhD could put together. They sure do it more quickly, but moving more quickly in the wrong direction isn’t that helpful in research! One encouraging direction I see some tools (notably Moara.io) moving is towards automating the drudgery of systematic data extraction, freeing up experts to do the high-context analysis required for synthesizing the literature.
If you’re not an expert, you’ll no doubt be impressed, but be careful. Gell-Mann amnesia is still as much a thing with LLM-assisted research as it ever was with science journalism.
On the topic of capabilities, agentic research assistants have come a long way in 2025. Elicit & Edison Scientific are ones to watch, but they still struggle to adequately cite the sources for the claims in the reports they generate our conclusions they come to. Contrast this with Deep Research models, which include nonsense from press releases and blog posts, even when explicitly asked to exclude these, though they have improved significantly too. Progress, but we’re still a long way from what a average PhD could put together. They sure do it more quickly, but moving more quickly in the wrong direction isn’t that helpful in research! One encouraging direction I see some tools (notably Moara.io) moving is towards automating the drudgery of systematic data extraction, freeing up experts to do the high-context analysis required for synthesizing the literature.
If you’re not an expert, you’ll no doubt be impressed, but be careful. Gell-Mann amnesia is still as much a thing with LLM-assisted research as it ever was with science journalism.