Could LLM psychiatrists be better than meat ones?
[Not medical advice, obviously]
I had a call with my psychiatrist – the first one in about a year; I hadn’t really felt the need to see him until then – and we spoke for a scant ten minutes. I gave a one-minute description of how I was feeling, and he fiddled with the dose of a mood stabiliser I take and that was it.
No attempt to dig deeper into how I was feeling, or give me a standard questionnaire like the PHQ-9 or MADRS which would have given a numerical idea of how I was doing over time. The session cost £150 for ten minutes.
I wondered if an LLM couldn’t do a better job than this guy.
So I went and downloaded a copy of the Maudsley and shoved it into a Claude project. (It’s over 900 pages, so I trimmed the PDF to just the ~150 pages on depression first.) I gave it custom instructions to ensure it always checked and referenced the Maudsley guidelines.
Here is it dismantling some common myths (Opus 4.5):
User: how long does it take antidepressants to start working in depression?
The Maudsley is clear on this, and it’s actually good news: the idea that antidepressants take 2–4 weeks to work is a myth.
All antidepressants show a response pattern where the rate of improvement is highest during weeks 1–2 and lowest during weeks 4–6. Statistical separation from placebo is seen at 2–4 weeks in single trials (which is where the “lag effect” idea comes from), but after only 1–2 weeks in meta-analyses. When large numbers of patients are treated and detailed rating scales are used, an antidepressant effect is statistically evident at 1 week.
In clinical practice using simpler observations, an antidepressant effect in an individual is usually seen by 2 weeks.
The practical implication: in those who will ultimately respond, virtually all will have begun to show at least minor improvement by 4 weeks. Those showing no discernible improvement at this point will very probably never respond to that drug at that dose. …
User: what’s the evidence for pushing escitalopram to doses higher than 20mg? search online as well.
The Maudsley says there’s minimal evidence for SSRI dose escalation in general, but specifically notes that escitalopram (along with venlafaxine and tricyclics) may be an exception.
The Furukawa 2019 meta-analysis found SSRIs plateau around 20–40mg fluoxetine equivalents (~10–20mg escitalopram). But a Bayesian reanalysis suggested escitalopram specifically can be escalated higher than other SSRIs—up to ~30mg for ages 30–65.
A Korean RCT (2019) found 30mg vs 20mg in non-remitters showed improved depressive symptoms on MADRS, but not increased response/remission rates—a modest continuous effect, not a categorical one.
An open-label pilot escalating to 50mg found median remission dose was 30mg, with 38% of remitters needing 50mg. Tolerability declined above 40mg.
Why escitalopram might be different: It has a unique allosteric mechanism—binding to an allosteric site decreases its own dissociation rate from SERT, potentially potentiating effect at higher doses. Other SSRIs don’t do this.
Safety: QTc increases ~10.7ms at 30mg—modest and generally not clinically concerning.
Bottom line: 30mg is reasonable and better-supported than high doses of other SSRIs, though the evidence isn’t definitive.
Certainly it seems like, if not a replacement for a traditional psychiatrist, an LLM + a copy of the Maudsley is very good foil for classical care.
Is $15k a year typical for car insurance? In the UK it's a few hundred dollars a year at most unless you're a very young or very risky driver.
‘Reality has a surprising amount of detail’ must be one of the lindiest posts on here. Well done @jsalvatier!
Thank you - this is the comment I was going to write.
Bartender at my local place wearing a loose-fitting charcoal suit (think Father John Misty): painfully cool. Robert from Industry wearing a black suit on his first day at an investment bank: totally out of place (“get a new suit, you look like fucking Neo”).
Nit: scaling up RL by 100x and inference by 10,000x would be a 1:3 OOM ratio I think
Or nicotine spray (like Nicorette QuickMist).
I get "invite invalid" when I follow the link.
I have often wondered if having a “no complaining” experiment would be similarly useful.
This was a fun quiz. I'd not come across it before!
(I ended up with high scores on both indices, but higher on the second subscore (leadership and confidence) which surprised me a little. Overall I scored 249, which Claude reckons based on the chart they provide is about +1.5 SD among extraverts.)