Harsha Bommana

Compositional SSL Research @ Lossfunk

[Harsha Bommana @harshablast]

@ziikarmali Gaming!!

[Harsha Bommana @harshablast]

@27upon2 Can you build an agent to do this tho?

[Harsha Bommana @harshablast]

We're still figuring out how to build a great research lab. This is a huge step in the right direction!

Thinking about AI research with these mental models has completely changed the way I approach and value research contributions.

It's so different from engineering!

[Lossfunk @lossfunk]

Attributes of a great research question

[new post] 👇

[Harsha Bommana @harshablast]

I recently sat down with Front Page by the @Analyticsindiam to unpack @ylecun's new LeWorldModel

We discussed:
• Representation collapse
• The exact difference between the LLMs and World Models
• The implications of LeWorldModel

https://t.co/vV6sJi5dDX

@AmitRajaNaik

[Lossfunk @lossfunk]

We are very glad to announce @AnthropicAI as the first sponsor for our Conference for AI Scientists (CAISc) 2026, co-organised by Lossfunk along with @bitspilaniindia.

Top three spotlight papers at the conference will be awarded $2,000 each worth of model credits.

This is very exciting for us because scientists across the world have been doing with amazing things with Claude and we believe these awards are a great chance for motivated teams to meaningfully multiply their research output.

Check out how senior scientists have been using Claude 🧵

[Harsha Bommana @harshablast]

Exciting times ahead! 🚀

[Lossfunk @lossfunk]

🦠🐵🙋🤖

[Harsha Bommana @harshablast]

@navbenny @RuntimeBRT Yooo LFG!!! 🚀

[Naveen Benny @navbenny]

Happy to be covered by @RuntimeBRT!

This is the era of the Intelligence Arbitrage. No longer will the markets be dominated by speed and information asymmetries, in the future, the markets will reward the most intelligent models.

But unlike general purpose LLMs, this will need specialized models that understand the language of the markets: price. That is what we are building.

Over the last year, our breakthrough foundational models have been managing our fully autonomous hedge fund delivering industry leading alpha. There is no human in the loop and we don't know why it does what it does, but it is just incredible to watch!

It's been a long journey but I'm excited to keep pushing the frontier and see what lies ahead :)

[Lossfunk @lossfunk]

📢 [New Preprint] VLMs often struggle to play video games.

Is the bottleneck reasoning or perception?

We gave three strong VLMs object coordinates while they played Atari, VizDoom, and AI2-THOR. Perfect coordinates helped every model. Self-extracted coordinates only helped one VLM, indicating perception to be the bottleneck.

Accepted at LM Reasoning Workshop, AAAI 2026

[Harsha Bommana @harshablast]

@raj_chandra3 @Sameeran_B Bro that’s Sourav

[Harsha Bommana @harshablast]

We made these posters for a friend on his birthday for party decor

Top: Last year, we had to finetune a flux model with LoRA for a few hours on his face images
Bottom: This year, we just prompted Gemini

AI progress is crazy

[Harsha Bommana @harshablast]

Brilliant way to probe frontier LLMs' actual reasoning abilities. Even with esoteric languages, given the documentation & examples, models should be able to answer questions, right? But they can't!

Check out this work by @inceptmyth for more details!

[Lossfunk @lossfunk]

🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%.

Presenting EsoLang-Bench.

Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵

[Harsha Bommana @harshablast]

Day by day, I'm losing my faith in the concept of backpropagation

[Harsha Bommana @harshablast]

@inceptmyth I tried using claude code to analyze my experiment logs (around 50 plots/images for each run). It does a great job at giving me an overview of the data, but fails to give me fine-grained insights that I would have missed as a human.
Might be better with a custom agentic workflow

[Paras Chopra @paraschopra]

Kickstarted Batch 7 of @lossfunk residency!

🧵 with their profiles and goals for the next 6 weeks

[Harsha Bommana @harshablast]

Giving a talk at @lossfunk tmr with @vitransformer! This will be about hierarchies in visual representation learning. We’ll be doing a deep dive on two pivotal papers in this field.

Do register at https://t.co/SSznaLoFCl if you’re interested in attending!

[Harsha Bommana @harshablast]

@SarahLevinger You've nailed the disconnect! At PerzAI (https://t.co/8oaq74wrPX), we understand that the customer journey is far from linear. Our AI-driven tools analyze real buyer behaviors to create content that resonates at every stage of their actual journey.

[Harsha Bommana @harshablast]

@loshminft Definitely! The key to impactful content is uniqueness and personal touch, which is often lost in today's template-driven strategies. We're building Perz AI (https://t.co/8oaq74wrPX) to ensure that every piece of content is tailored and engaging, turning generic into genuine.

[Harsha Bommana @harshablast]

@jakesnowfall Couldn't agree more on the pitfalls of copying content. That's why PerzAI (https://t.co/8oaq74wrPX) focuses on crafting personalized and unique content that resonates with each brand’s audience, ensuring authenticity at scale.

[Harsha Bommana @harshablast]

@Chris_Martine2 At Perz AI (https://t.co/8oaq74wZFv) creating personalized content using data-driven insights is our mission. Please check it out and let us know what you think! We would love to hear your feedback and insights.