Aman Sanger

I'm Aman. I write code and am obsessed with AI.
I'm building cursor, a new way of programming using AI.

I've worked at Google, Bridgewater, you.com, and graduated from MIT.
While there, I did NLP & comp bio research. I also ran a small AI consultancy.

Thoughts

1. Llama-2 is expensive
2. Latent Space Podcast

Shorter thoughts

1. 4-bit weight-quantization is more expensive than 16-bit
2. flash attention isn't helpful when generating tokens
3. llama-1 needs multi-query attention
4. instruction finetuning is underrated

Contact

• Github
  
• Email
  
• Twitter
  
• LinkedIn