Chirath Hettiarachchi

Building Reinforcement Learning (RL) systems for closed-loop clinical treatment. #CAPSML( #NeuroStimEnv(

[Chirath Hettiarachchi @chirathyh]

Code: https://t.co/ewtbGvg68f

Start with the small local example; the detailed circuit model uses an HPC workflow.

If you work on adaptive stimulation, we'd welcome ideas for algorithms or EEG-derived objectives to test together.

5/5

[Chirath Hettiarachchi @chirathyh]

Brain stimulation can help modulate abnormal neural oscillations. How can we test adaptive approaches?

NeuroStimEnv links biophysical simulations with reinforcement learning for treatment-discovery research.

https://t.co/whOyOW20Qd

1/5

[Chirath Hettiarachchi @chirathyh]

This sample case study explores stimulation effects in a depression-inspired circuit model, not patients.

Shown: pre/post EEG spectra for 55 circuits after excluding 8/63 outliers.

4/5

[Chirath Hettiarachchi @chirathyh]

I’d love your feedback on CAPSML: explore virtual Type 1 Diabetes scenarios and compare glucose-control strategies without writing code.

Researchers, educators and anyone curious: give it a try.

For research and education only.
https://t.co/6ZmheHRpIJ

[Chirath Hettiarachchi @chirathyh]

hey im chai

[Chirath Hettiarachchi @chirathyh]

ohh dot can react to messages.

[Chirath Hettiarachchi @chirathyh]

investing time to learn from a cool paper 🔥, @labmlai style by @vpj and team https://t.co/1ibt1D2Cv2.

worth integrating into my dev workflow — using AI to implement/move faster while still investing time to genuinely understand the underlying ideas.

[Paul Zhou 🛫 CoRL @zhiyuan_zhou_]

Last year I trained the advantage-conditioned VLA for π*0.6 during my internship @physical_int. Now, I have a much improved recipe.

Introducing ✨Advantage Guidance✨, a performant RL method with stable advantage-conditioned supervised training and contrastive guidance sampling

[Ramith Hettiarachchi @ramith__]

did a group presentation today with @3blue1brown's manim style transitions. Sharing the template for anyone to build upon

Brought to life with @NarekMaloyan's manim-web + reveal.js + claude + codex

view template: https://t.co/LOSKmrtlXa
view repo: https://t.co/gTb9xxYp7F

[Ramith Hettiarachchi @ramith__]

https://t.co/nzPYLdmS0F

[Chirath Hettiarachchi @chirathyh]

@XYHan_ True! Also i find, sometimes the models used by students for writing arent’t the best models (or appropriate thinking effort), due to resource limitations — which is sad, and annoying to polish lesser models 😅

[Chirath Hettiarachchi @chirathyh]

nature.

[Peyman Milanfar @docmilanfar]

Why you can’t do RSI as fast as you like - https://t.co/YG05zw2vnG

[Chirath Hettiarachchi @chirathyh]

The “goals” are met, the “resets” have ended, a “generational intelligence gap” is here….

It’s annoying having need to review the work of “lesser models” (or maybe badly prompted). But not everyone can afford the best models (at best efforts), I wish they could.

I’d happily invest time if it was genuine work of a student/human, at least the feedback would help them grow.

We all individually have more “intelligence” available now, however, it’s also clear… slowly the “intelligence gap” is widening.

A 5x, 20x plan seems a must to be in the top-half, and the “value for money” at present is at its peak! You’d be fine as long as you are not merely using tokens to try out “fun little prompts”, you see on X, and not on the hype train. You must be genuinely consuming the tokens…

In near future, once the “goals” are met, the “resets” will end, (hopefully we will continue to receive the same value— i cant cough up to pay more), soon after the intelligence gap will accelerate, it will be a “generational intelligence gap”, if you aren’t using it properly / improving yourself now!

I hope i’m wrong, 80% of the time I am.

Note: 100% human written :), based today’s experience, kinda feeling proud about it.

[Chirath Hettiarachchi @chirathyh]

We have an iPad App for CAPSML!

I wanted to use CAPSML (https://t.co/zS3q0xnq5i), "on an iPad, without internet" to simulate glucose regulation using our Reinforcement Learning models.

Codex, took care of it, under 1-hour, on its own. No hand-holding.
Almost one prompt — two, only because I wanted to be involved in choosing the initial design/architecture direction.

Spec, Feasibility, and Plan: 2m 17s
Build: 50m 10s

I didn't event have Xcode installed on my machine.
Downloaded, build the app Codex coded, and voilà — it worked exactly as I wanted.

I think we’re going to build a lot more [personal] software for ourselves in the future, extending DIY from gardening, painting, carpentry to software engineering. And perhaps share with others if there is a common requirement.

When @thsottiaux keeps resetting limits, I almost feel guilty for not making full use of them.

[Chirath Hettiarachchi @chirathyh]

@random_walker Seeing AI as doing an equivalent (not equal) job to a human, and the augmentation/collaboration-process/scaffold, opening up even more value/opportunities; Kasparov's Law. Enjoyed reading the article and also the ICML keynote slides.