{"blyg":"0.3","id":"5qp3gtfgtqvewkt06ja8gavm6x","kind":"thread","origin":"https://blyg.aneeshsathe.com/","page":"t/5qp3gtfgtqvewkt06ja8gavm6x/","author":{"name":"Aneesh Sathe","url":"https://blyg.aneeshsathe.com/"},"created":"2026-09-25T22:52:04Z","updated":"2026-09-25T22:53:38Z","version":1,"content_md":"AI and Causal Inference\n\nWhile reading: [Six Lemmas Concerning Heterogeneity and Nonlinearity in Causal Inference](https://math.la.asu.edu/~prhahn/six-lemmas-narrative.html) by [P. Richard Hahn](https://math.la.asu.edu/~prhahn/)\n\nBiology has studied the regimes where signal was easy to detect with small experiments and simple models: narrow dose ranges, few readouts, and assumptions like a fixed curve shape and uniform effects. High-content data plus flexible ML lets us drop those assumptions and ask whether the curve, and the response across individuals, really looks the way we assumed. ML makes that robust to getting the model's shape wrong, but not to bad design or unmeasured confounding. \n\nThe reductionist focus, where you isolate one variable, was partly a deliberate choice. The simple regime was the only one we could work in, and that's no longer true.\n\nWith the help of ML, the drunk man is not limited to looking for keys under the streetlamp:\n\n>I think we forget, the point of science isn't to chose phenomena where we have a good chance at fitting the right model, the point is to study phenomena we don't understand and hope that our models help us discover new things about it. ","content_html":"<p>AI and Causal Inference</p>\n<p>While reading: <a href=\"https://math.la.asu.edu/~prhahn/six-lemmas-narrative.html\">Six Lemmas Concerning Heterogeneity and Nonlinearity in Causal Inference</a> by <a href=\"https://math.la.asu.edu/~prhahn/\">P. Richard Hahn</a></p>\n<p>Biology has studied the regimes where signal was easy to detect with small experiments and simple models: narrow dose ranges, few readouts, and assumptions like a fixed curve shape and uniform effects. High-content data plus flexible ML lets us drop those assumptions and ask whether the curve, and the response across individuals, really looks the way we assumed. ML makes that robust to getting the model's shape wrong, but not to bad design or unmeasured confounding.</p>\n<p>The reductionist focus, where you isolate one variable, was partly a deliberate choice. The simple regime was the only one we could work in, and that's no longer true.</p>\n<p>With the help of ML, the drunk man is not limited to looking for keys under the streetlamp:</p>\n<blockquote>\n<p>I think we forget, the point of science isn't to chose phenomena where we have a good chance at fitting the right model, the point is to study phenomena we don't understand and hope that our models help us discover new things about it.</p>\n</blockquote>\n","content_hash":"sha256:d5e029ce6a97fc94d37e759e111c0e089c44e130d3303d643e59383120be9fce","media":[],"transclusions":[],"changelog":[{"version":1,"at":"2026-09-25T22:53:38Z","note":null}]}