Apoorva Mehta

Columbia MD, MBA & Resident Physician
RL, Data, Healthcare, Life Sciences

[Apoorva Mehta @apoorvamehtamd]

@Jack_Raines It was fun for a week

[Apoorva Mehta @apoorvamehtamd]

@deepseek_ai believes data is the bottleneck.

High-quality, diverse, & high volumes of healthcare, biology, & physical AI data are the blockers to models continuing to improve and making real impacts. If this is what you’re experiencing & are in demand of, reach out.

[Erik Reinertsen, MD, PhD @erikrtn]

@MarioATX_MD Content shared with ChatGPT for Clinicians is not used to train OpenAI’s models https://t.co/ip2wKLk8Z2

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

CancerBench: the frontier model cancer cure benchmark.

AI lab CEOs keep talking about curing cancer, so I made a benchmark.

One metric: how many types of cancer has your model cured?

All models are currently tied at zero.

It’s time to hillclimb!

https://t.co/SeEJz0FFYv

[Apoorva Mehta @apoorvamehtamd]

@kevinzhang @sahirjaggi is a man on a mission, I'm sure it was an easy yes

[Apoorva Mehta @apoorvamehtamd]

I think in the long-term, abstracting away all of the complexity with minimal human interaction will be the correct thesis.

I think what gets missed is a stepwise function of how you get there. Today, it’s hard to achieve that minimal of human interaction tied in with an extremely successful and sticky product. When I find myself trying to train an agent, I prefer a specific questionnaire that allows for it refining to my tastes, and to recreate the way I approach problems or tasks. I believe this would be what’s most sticky today, which is what will allow the agents to build personalized understandings that allow for full end-to-end management.

[Apoorva Mehta @apoorvamehtamd]

Employers are cutting insurance coverage of GLP-1s.

It's a perfect representation of misaligned incentives for insurance plans. GLP-1s are fantastic medications for many individuals that have diabetes and/or are facing issues with obesity, and has played a direct role in the obesity rates dropping for the first time in decades.

It would make sense that employers or insurance plans should want to take advantage of the savings that come from healthier covered lives, but instead, these plans want to cut coverage as there is no long-term incentive or benefit to think about an employee's lifespan when they're likely to leave the company before all of the benefits of the GLP-1s would be recognized.

There needs to be a complete overhaul of the insurance system mechanics.

[Apoorva Mehta @apoorvamehtamd]

What’s consensus on Boardy right now?

[Apoorva Mehta @apoorvamehtamd]

It's incredible to see the focus on the healthcare and life sciences from the frontier AI community. While some of the newly funded tech-bio companies may succumb to the realities of implementing in real workflows in poorly designed systems, I fundamentally believe the RL & model-training work here has just begun, and is going to have one of the largest impacts in the ecosystem.

[Apoorva Mehta @apoorvamehtamd]

There's a never-ending amount of possible discourse on why medical student research has incredibly perverse incentives.

No one is saying research in the way it is fundamentally meant to be explored to push the frontier of medicine and science is something students shouldn't be involved in. However, when it becomes an arms race for residency requirements, you end up with things you wouldn't expect. For example, medical students are often vying for labs and PIs that understand and optimize for this arms race, and certain students get the opportunity to be co-authors on the entire labs' research papers as a way to boost numbers. Everyone in these labs is incentivized to put out quantity over quality. The priority out of these publications isn't to optimize for medical or scientific impact, but personal performative output.

However, this is just one of many questionable parts of the physician training pathway.

[Apoorva Mehta @apoorvamehtamd]

Releases like these are just phenomenal for the life sciences community (and tech in general). The frontier moves so quick, and people like @ChrisHayduk are giving so many people the opportunity to follow along, get quickly acquainted, and contribute themselves sooner rather than later.

Really cool release, can't wait to work through myself.

[Chris Hayduk @ChrisHayduk]

Learning resources play a key role in shaping where talent goes. Today I’m launching my side project, BioTorch, in public beta to give researchers a clearer path into biological AI.

LLMs have a wealth of great tutorials, courses, and minimal implementations from great teachers. They make the path from curiosity to contribution easy to see and offer researchers from fields like math or physics a legible roadmap to transition into AI research.

In bio AI, that path is much harder to piece together. That friction can mean fewer researchers entering the field, fewer startups getting built, and less investment allocated to promising ideas. This ultimately leads to fewer life-saving cures being discovered than would otherwise be possible.

BioTorch is my attempt to help change that.

It starts with 116 PyTorch exercises and 17 model guides. You implement the building blocks, test your code, and see how the pieces fit into seminal models like AlphaFold2, ESM2, and RFdiffusion.

The goal is to make influential papers in bio AI something you can understand and build on, bringing researchers up to the frontier as rapidly as possible.

I want more people working on biology’s hardest problems. My mission is to help researchers find their way into the field, and BioTorch is the result of that ambition.

BioTorch is completely free while in beta.

Try your first problem now!
https://t.co/BHmN59S3bI

[Apoorva Mehta @apoorvamehtamd]

Believe you have an edge.

Thought I'd share some weekend inspiration. Everyone who is successful fundamentally believes they have a unique, undeniable edge that no one else can touch. And this is by no means a product of "success" that everyone deems valuable, it's a consistent belief in self that is your own unique combination of why you're going to leave an impact on the world.

Whether you're a $1B exited founder, a 6x unsuccessful start-up founder, or someone working as a product manager imagining and dreaming of your next move, you are unique, you have an edge, and you will be successful.

[Apoorva Mehta @apoorvamehtamd]

The real frontier of tech-enabled care delivery is the opportunity to spend more time with patients. In our medical training, everyone works through OSCEs (Objective Structured Clinical Examinations) with simulated patient scenarios and actors in which you learn & practice asking open-ended questions to be able to understand a patient's unique condition and progression of disease.

This was always a great starting point for trainees to understand the fundamentals of developing rapport with patients. However, there is so much more opportunity to have an impact when physicians have more time with their patients. Having an extra 10 minutes to sit down and talk about smoking cessation or substance use without the fear of falling too far behind on your clinic schedule could be the difference between decades of your patient's life expectancy. Getting the time and opportunity to explore cultural & socioeconomic factors more thoroughly, while having technology allow for better coordination and delivery of care in these more ignored factors of someone's health, is what will actually revolutionize care delivery.

[Apoorva Mehta @apoorvamehtamd]

Have thousands of hours of OTS ego data along with operational structure for bespoke collections with significant capacity scale in multiple locations with experienced operators, reach out if you're interested!

[Apoorva Mehta @apoorvamehtamd]

Wonder how big of a settlement we'll see in the future from possible negative impacts of AI tools (if at all).

[Apoorva Mehta @apoorvamehtamd]

@himanshustwts Lots of M&A coming soon

[Cameron R. Wolfe, Ph.D. @cwolferesearch]

I just published my complete guide to reinforcement learning for LLMs. It's a single, standalone resource for understanding RL from first principles to frontier research.

Read it here: https://t.co/iUMAWUrkQ1

The topics covered are as follows…

RL Fundamentals:
- A general framework for RL (policy, states, actions, transition functions, environment, rewards, trajectories, etc.)
- Core concepts of RL (returns, discounting, action / state / trajectory probabilities, value / advantage functions).
- Formulating and estimating the learning objective for RL.

Basic of RL for LLMs:
- Token-level (MDP) vs. completion-level (bandit) formulations.
- Outcome versus process rewards.
- Value estimation via value models / critics.
- KL divergence, RL setups (RLHF, RLVR, infra, etc.), importance sampling.

Basic Policy Gradients:
- Deriving the RL objective / vanilla policy gradient (VPG) from first principles.
- Different policy-gradient formulations (trajectory returns, reward-to-go, baselines, value functions, and advantages).
- Why policy gradient estimates have high variance (and how baselines help).
- Implementation of VPG.

REINFORCE and its variants:
- REINFORCE as a Monte Carlo implementation of the VPG.
- Completion-level vs. token-level implementations of REINFORCE.
- Reward baselines and KL-regularized rewards.
- RLOO, REINFORCE++ and other critic-free policy gradient algorithms.

Actor-Critic Methods:
- Training value models / critics alongside the policy to estimate advantages.
- Trust Region Policy Optimization (TRPO) and KL-constrained policy updates.
- How PPO simplifies TRPO with a clipped surrogate objective.
- Full PPO details (clipping, advantage estimations, value loss, KL, policy ratios, multi-epoch optimization, etc.) + implementation.
- Generalized Advantage Estimation (GAE) + implementation.
- The memory and computational overhead of actor-critic algorithms.

GRPO:
- Group Relative Policy Optimization (GRPO) as a simpler alternative to PPO.
- Group-relative advantage estimates + implementation.
- Why removing the critic is helpful.
- Limitations of vanilla GRPO algorithm.
- Dr. GRPO, DAPO, GSPO, and other recent improvements.

Recent research topics:
- Online versus offline RL.
- Rubric-based RL.
- RL scaling laws.
- Agentic RL + world modeling.

This post is a synthesis of my writing / learning on RL for over a year. I hope it’s helpful!

[snimu (Sebastian Müller) @omouamoua]

I recently wrote down a few rules of prompting for RL. There's exceptions for each of them, but they are good guidelines so I thought I'd make them public.

[Apoorva Mehta @apoorvamehtamd]

@AndrewYNg Great list, and at the end of the day, I think it's just so much about customization.

[Apoorva Mehta @apoorvamehtamd]

@siddarthv66 Got a big data collection operation, let me know if you’re interested