Tanishq Mathew Abraham

Has anyone successfully done RL post-training of GPT-oss with meaningful performance gains?

What libraries even support it? I guess technically TRL/axolotl, maybe Unsloth... but there are no good examples of doing it...

Check out our website: sophontai.com

Read our manifesto/announcement: tanishq.ai/blog/sophont

If you're interested in building & collaborating in this space, whether you're in genAI or medicine/pharma/life sciences, feel free to reach out at: contact@sophontai.com

researcher at Princeton University, and served as Head of NeuroAI at Stability AI. He led the team behind the MindEye publications, which achieved state-of-the-art reconstructions of seen images from brain activity.

With over five years of experience applying generative AI to medicine, I bring a wealth of expertise, having previously served as the Research Director at Stability AI and CEO of MedARC. My co-founder, Paul Scotti, has a decade of experience in computational neuroscience, was a postdoctoral ...

We are strong believers in open, transparent research, based on our proven track records in medical AI research and building open science online communities. We hope to continue this science-in-the-open approach to research at Sophont to build towards the future of healthcare.

Our vision is to merge highly performant, specialized medical foundation models into a single holistic, highly flexible medical-specific multimodal foundation model. We believe that this should be done open-source for maximum transparency and flexibility

These approaches succumb to the parable of the blind men and the elephant: The blind men are unimodal medical models and the patient is the elephant.

AI is clearly needed to enhance doctors’ ability to provide the best care.

However, currently deployed medical AI models are inflexible, rigid, suited for narrow tasks focused on individual data modalities.

I have EXCITING news:

I've started a company!

Introducing Sophont

We’re building open multimodal foundation models for the future of healthcare. We need a DeepSeek for medical AI, and @sophontai.bsky.social will be that company!

Check out our website & blog post for more info (link below)

Btw, I posted this 2 weeks ago on Twitter but forgot to post here, so doing it now. Twitter is probably going to always be the fastest place to get updates from me unfortunately 😅

NEW BLOG POST: LLMs in medicine: evaluations, advances, and the future

www.tanishq.ai/blog/posts/l...

A short blog post discussing how LLMs are evaluated for medical capabilities and what's the future for LLMs in medicine (spoiler: it's reasoning!)

Yeah, but compute scaling can mean lots of things including synthetic data for example

Artificial superintelligence

New blog post coming tomorrow on medical LLMs...

I restarted my blog a few weeks ago. The 1st post was:

Debunking DeepSeek Delusions

I discussed 5 main myths that I saw spreading online back during the DeepSeek hype.

It may be a little less relevant now, but hopefully still interesting to folks.

Check it out → www.tanishq.ai/blog/posts/d...

Are folks still here? 😅

github.com/deepseek-ai/...

Okay so this is so far the most important paper in AI of the year

Anthropic, please add a higher tier plan for unlimited messages 😭🙏

Decentralized Diffusion Models

UC Berkeley and Luma AI introduce Decentralized Diffusion Models, a way to train diffusion models on decentralized compute with no communication between nodes.

abs: arxiv.org/abs/2501.05450
project page: decentralizeddiffusion.github.io

The GAN is dead; long live the GAN! A Modern Baseline GAN

This is a very interesting paper, exploring making GANs simpler and more performant.

abs: arxiv.org/abs/2501.05441
code: github.com/brownvc/R3GAN

Happy birthday to my incredible and awesome Mamma! 🥳🎉🎂

To many more years of health and happiness.

Tiara (my sister) and I love you very much ❤️❤️❤️

Happy 19th birthday to my amazing sister Tiara Abraham! 🥳🎉 🎂

Proud of you graduating with your Master's degree at 18 and starting your doctorate in music degree this past year!

Excited to see what this final teen year holds for you!

Flow matching is closely related to diffusion and rectified flows and Gaussian flow matching is equivalent to denoising diffusion.

GitHub - facebookresearch/flow_matching: A PyTorch library for implementing flow matching algorithms, featuring state-of-the-art continuous and discrete flow matching implementations. It includes prac... A PyTorch library for implementing flow matching algorithms, featuring state-of-the-art continuous and discrete flow matching implementations. It includes practical examples for both text and image...

PyTorch library: github.com/facebookrese...

Inventors of flow matching have released a comprehensive guide going over the math & code of flow matching!

Also covers variants like non-Euclidean & discrete flow matching.

A PyTorch library is also released with this guide!

This looks like a very good read! 🔥

arxiv: arxiv.org/abs/2412.06264

Normalizing Flows are Capable Generative Models Normalizing Flows (NFs) are likelihood-based models for continuous inputs. They have demonstrated promising results on both density estimation and generative modeling tasks, but have received relati...

Normalizing Flows are Capable Generative Models

Apple introduces TarFlow, a new Transformer-based variant of Masked Autoregressive Flows.

SOTA on likelihood estimation for images, quality and diversity comparable to diffusion models.

arxiv.org/abs/2412.06329

Refusal Tokens: A Simple Way to Calibrate Refusals in Large Language Models

"We introduce a simple strategy that makes refusal behavior controllable at test-time without retraining: the refusal token."

arxiv.org/abs/2412.06748

Can foundation models actively gather information in interactive environments to test hypotheses?

"Our experiments with Gemini 1.5 reveal significant exploratory capabilities"

arxiv.org/abs/2412.06438

Training Large Language Models to Reason in a Continuous Latent Space

Introduces a new paradigm for LLM reasoning called Chain of Continuous Thought (COCONUT)

Directly feed the last hidden state (a continuous thought) as the input embedding for the next token.

arxiv.org/abs/2412.06769