Robert Y. Chen

🚨 The HHS led by Dr. Oz and RFK Jr. just dropped a major policy letter pushing for deprescribing, shared decision-making, and non-pharmacological treatments in mental health care. The key message: meds are essential, but shouldn't be the only tool in the toolbox 🧵

@PTrubetskov We agree

@therealRYC One PHQ-9 point sits inside measurement noise. Thin ground for a policy letter scaled to millions.

🚨 The HHS led by Dr. Oz and RFK Jr. just dropped a major policy letter pushing for deprescribing, shared decision-making, and non-pharmacological treatments in mental health care. The key message: meds are essential, but shouldn't be the only tool in the toolbox 🧵

@dignityhnrgrace Sigh. Agreed.

@therealRYC BBB decimates Medicaid over the next several years.

@dignityhnrgrace How should prescribers get better at this? Are there prescribers who you’ve felt have done a good job?

@therealRYC Prescribers as a whole lack any shred of competence in recognizing ADEs. And, who defines the length of an “adequate trial?” The current approach completely ignores ADEs, dismissing, and gaslighting patients who report them.

💡 Takeaways In my clinic and at the hospital I train at, I treat patients suffering from a range of mental health disorders spanning mild to severe impairment. I spend half my time optimizing meds and the other half trying to peel away unnecessary polypharmacy inherited from previous admissions or cowboy polypharmacy. It's encouraging to see federal backing for deprescribing and shared decision-making. We need to normalize asking: "Do you still need to be on this?" This is an overall net-good. The caution: the message must match the evidence. And the reality is that the best studies on deprescribing have not been impressive. Neither have lifestyle interventions. The key will be in tailoring the right treatment (or removal of treatment) to the right patient. Blanket statements that deprescribing, or exercise over meds, is always good or always bad, are not backed by evidence. The good news is that this letter is not making those types of statements. In fact, the letter states many times: "When clinically indicated". Where this goes next → watching how CMS aligns actual reimbursement with these deprescribing and non-pharmacologic treatment goals. Source: U.S. Department of Health and Human Services. Dear Colleague Letter. May 2026 Read the full letter here: samhsa.gov/sites/default/…

🔎 Caveats Further Continued And, in the largest randomized trial testing antidepressant deprescribing, there was a significant but clinically negligible difference of one point on the PHQ-9 favoring having a deprescribing consultant on-board. …work-com.offcampus.lib.washington.edu/journals/jaman…

@ericwu93 @james_y_zou @gxl_ai 🔥

@therealRYC @james_y_zou @gxl_ai Hi Robert, Paperclip updates automatically with the latest literature as its made available, depending on the source (e.g., biorxiv makes monthly releases).

Big Update🤩: #paperclip now includes full papers from all of arXiv, PubMed Central and 150 million abstracts!🖇️ You can give your LLM all that knowledge in one line—all optimally indexed for AI agents. Much more thorough and ~100x faster than web search, and free.

@akseljoonas Wow. I want to give this a try. What’s the best way to figure out cost structure for someone who isn’t AI native, but has a problem to work on? I imagine it could be expensive to test out 50 different training runs.

ml-intern is fully on mobile now you can launch 8 A100s from your phone. while on the couch. while commuting. wherever I just did this while biking. same sessions as your desktop too — start a run on your laptop, check on it from your phone, it's all there the research lab is now just wherever you are

@KexinHuang5 Does phylo save any of the data for internal use?

I’ve spent the past decade building bio AI models—until now, training them meant years, huge cost, and teams spanning AI, biology, and infra. Not anymore. Introducing a new capability at Biomni Lab: now any scientist can create, fine-tune, pre-train, and optimize bio foundation models on their own datasets, just by describing what they want. This is powered by a new feature called GPU-as-a-tool where we let AI agents launch and orchestrate GPU sandboxes. In the video, we show that you can use prompt to: - Fine-tune Borzoi for MPRA regulatory activity prediction - Fine-tune scGPT on a H1 hESC perturb-seq data - Fine-tune ESM2 for protein subcellular localization prediction - Pre-train a protein language model from scratch on UniRef - Build a novel multi-task ADMET model across 22 endpoints Another big challenge once you’ve trained or have access to a model is actually using it productively. As it is embedded within Biomni Lab, it bridges that gap, letting you “ask the model” to identify and prioritize relevant biological insights directly. Another exciting direction is lab-in-the-loop: scientists can design models, generate predictions, interpret results, and send them to the wet lab—all within one integrated biology environment. This is a preview capability and we’re looking for beta testers. Sign up here for early access: forms.gle/1yhCP6Vrc12DaS… Learn more about opportunities and limitations in our blog: phylo.bio/blog/ai-agents… @phylo_bio

Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code. But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along. So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions. TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.

Agreed - underestimating people is almost always a mistake. I do it even when I try not to. I can get better at it and own up to it when I make mistakes - it’s the algorithm for progress. Informed consent has changed a lot, thankfully. I’m sure contempt for patients persists in some areas of medicine. I would call that bad medicine. Hopefully you don’t encounter that on any sort of frequent basis, I certainly haven’t but I am biased because I work at a large academic hospital in the US which is not representative of care more broadly (in the states or abroad).

@therealRYC It's helped me to see the very low bar informed consent started from in the 1950s, and also the contempt for patients, which persists in some areas of medicine to this day. It's a big mistake to underestimate people imho.

@jill_d35 Great book

@therealRYC Was that supposed to be an apology for accusing me of rage baiting? Bad faith assumptions are unhelpful. I'm reading something very interesting about the early days of clinical trials and informed consent that you may or may not be interested in. x.com/i/status/20385…