ELIZA wasn't the only chatbot in the late 1960s... this JAMA+AI podcast explores how their creators took very different paths in beliefs about AI in medicine
Not to take anything away from the meta-analysis and the important points in the blog post, but there's a glaring error in the abstract: "the weighted average effect sizes were as follows: social isolation odds ratio (OR) = 1.29, loneliness OR = 1.26, and living alone OR = 1.32, corresponding to an average of 29%, 26%, and 32% increased likelihood of mortality, respectively."
The authors mean increased /odds/, not likelihood (probability). WHy does it matter? Well, when your whole paper is a statistical exercise, misusing basic statistical language in the abstract is not a great sign.
Before everyone goes bananas citing Goodhart's law: many universities and academic medical centers in the US don't care at all about impact factor - they care about grant $$, period full stop. (They appreciate the occasional high-impact paper that they can use in marketing materials, but it's really all about the $$.)
And for what it's worth, I've almost never heard impact factors discussed at NIH study sections, where investigator quality is explicitly on the agenda. Reviewers talk about relevant prior publications in the field, esp in marquee journals. [this latter feature is the reason we don't just put everything on biorxiv or equivalent and move on.]
And forgot to mention that there's all sorts of very cool new tools coming - one of my favorites is this one for targeted protein degradation:
https://www.ncbi.nlm.nih.gov/pubmed/28223226
Well, I think it's true that traditionally folks trying to do rational drug discovery are better at targeting certain proteins (cell surface receptors; some enzymes) with small molecules than others. Also true that first generation of most drugs were discovered serendipitously and only later was mechanism understood. BUT - notion that we're at the sharp downward slope would be a surprise to most pharmas and biotechs. While there's more of an emphasis on understanding entire pathways rather than single proteins, it's still generally the goal to find a single place in the pathway where a small molecule (or antibody, or whatever) will act. It's also /really/ hard to disrupt protein-protein interactions, though it's possible.