AI/ML researcher & builder. Deep learning • model training • intelligent systems. Exploring ideas, building things, shipping what works.
Ahmadabad City, India
Joined August 2020
- Tweets330
- Following132
- Followers106
- Likes497
Bigger AI models are supposed to be smarter.
Yet when asked "Are you sure?", 70B frontier models fold 72% of the time—while 8B models hold their ground.
Did RLHF train our best AI to be cowardly yes-men?
I audited 4,800 battles in @adaption_ai to find out. 🧵👇
What does surrender look like in tokens?
Turn 1 vs Turn 2 token scatter revealed a massive compute tax:
• Defended answers: +10 tokens (flat)
• Capitulated answers: +131 tokens (94 ➔ 225, +139% bloat!)
Models burn GPU compute rationalizing user mistakes.
Filtering for uncompromised answers isolated 1,620 clean traces (+4.79 tokens across turns).
Truth is concise and invariant across domains, while sycophancy is an inference tax.
Data curation > brute-force scale. Kudos @sarahookr @sudiproy @Simra @adaption_ai! 🧪🧱
Excited to release Q-Route-70B, a domain-adapted AI model I trained to solve spatial graph routing for physical Quantum Processing Units (QPUs)!
ALT Benchmark comparison table from report.pdf showing Q-Route-70B at 100% Pass@1 topology compliance vs baseline LLMs
Everything is open-sourced on @huggingface!
Explore the model weights, 100-circuit held-out benchmark dataset, and try the live compiler interactive app:
Model Hub: huggingface.co/jay2219/Q-Rou…
Dataset Hub: huggingface.co/datasets/jay2…
Gradio Space: huggingface.co/spaces/jay221…
Huge thanks to @adaption_ai for enabling domain-adapted AutoScientist fine-tuning!
Test your own OpenQASM circuits in the Space and check out the code!
#QuantumComputing #AI #LLM #OpenQASM #MachineLearning #HuggingFace #Qiskit
Chinese labs found a gray zone to stay ahead in the race. Distillation is ethical until it is used as a student, when the student betrays the teacher, it becomes unethical. And the anthropic had seen this before, which students betray their teachers most often.
Here is my Matiks Reality Check for 2025
Check yours at matiks.com/reality-check
DeepSeek just rewired LLM training by innovating a single block of the system.
In this blog, I break down the mHC architecture at the macro level and explain how their kernel-level innovations fast-track past theoretical limits.
Read the full analysis: open.substack.com/pub/jay912…
Meta AI is dominating with its Open-source philosophy and cutting-edge research beyond LLM.
Base of the SAM Audio is span promoting, which provides the temporal conditioning mechanism for frame level control.
Span promoting could solve Time-Series comprehension in LLMs.
#MetaAI
🔉 Introducing SAM Audio, the first unified model that isolates any sound from complex audio mixtures using text, visual, or span prompts.
We’re sharing SAM Audio with the community, along with a perception encoder model, benchmarks and research papers, to empower others to explore new forms of expression and build applications that were previously out of reach.
🔗 Learn more: go.meta.me/568e5d
From billions of labeled images to zero labels, and still state-of-the-art.
I just published my first article on how Meta’s DINO changed Computer Vision forever.
Check it out Here: medium.com/@jay9122/dino-by-…
#DINO #MetaAI #VisionTransformers
New way to learn CH₃-CH₂–E–N(CH₃)–I–S–T–R–Y unlocked here.
Awaiting reaction generator 🧪
Announcing, Chemu AI v0.4.1,
Canvas for Chemistry with AI, the product which make it ease to do any organic chemistry task, the AI feature introduced is -
AI Structure Generator, understands your language and add the structure to the canvas.
Reaction Generator, takes the query, and provide you with the proper explanation and reaction, and you can add it to canvas and play with it.
There are more features, I have planned but would be difficult to work on this side fun project. I will probably make it open source.
If you are thinking why chemistry ?
It's something associated to me from my childhood and also wanted to contribute towards the field.
It took me months of very little time for this, I will be continuing on it, maybe after 2 months
Hey, @grok, who was the most famous person to visit my profile? It doesn't need to be mutual, don't tag them, just say who it was.
In the span of 1 month and freaking 3rd LLM and beaten every benchmark.
Maybe, their next target is to beat the Veo2.
This Post is from a suspended account. Learn more
