Suggestions
Pratik Bhavsar
└► NLP Scientist - pratik.ai - IIT Bombay
Pratik Bhavsar is an accomplished NLP (Natural Language Processing) scientist and founding engineer at Enterpret, where he played a crucial role in developing various AI systems, including semantic search and MLOps. He worked at Enterpret from February 2021 to May 2023, contributing to projects that categorized over 50,000 customer feedback categories and developed innovative search engines and chatbots for book recommendations.25
Background and Education
- Education: Pratik holds a Bachelor of Technology (B.Tech) from the National Institute of Technology Surat and a Master of Technology (M.Tech) from the Indian Institute of Technology, Bombay.2
- Professional Experience: Before his tenure at Enterpret, he was involved with other startups like Taskhuman. Currently, he is working as a Machine Learning Engineer at Galileo, focusing on data-centric ML applications.23
Skills and Contributions
Pratik is self-taught in data science and has extensive experience in:
- NLP and Deep Learning: He has developed systems for text classification, chatbots, and algorithmic trading using machine learning techniques.
- Community Engagement: He founded Maxpool, a community for data science professionals to discuss practical ML challenges. He actively participates in hackathons and has won several awards for his innovative solutions.123
Online Presence
He maintains an active online presence through various platforms:
- GitHub: Showcases his projects related to NLP and machine learning.1
- LinkedIn: Shares insights on AI developments and engages with the data science community.2
- Blogs: Writes about NLP topics on Substack and Medium, contributing to the broader understanding of AI applications.34
Pratik is recognized as a prominent figure in the AI community, having been listed among the top AI developers to watch in 2023.2
Highlights
I just confirmed sycophancy in LLMs by asking "Are you sure" on a super easy task.
LLMs tend to agree with the user even when the user is mistaken. They can change their correct answer to incorrect if they are asked, “Are you sure?”
Mind it, I am using the most powerful GPT-4o and it went instantly blind when I asked are you sure! What can be the reason for such a behaviour?
This behaviour was explored in the paper from Anthropic- Towards Understanding Sycophancy in Language Models
Seems this thing will shit its pants when a developer agent is asked "Are you sure" by the manager agent :P
Imagine the amount of stress generated for Mistral and Cohere in last 24 hours.
They are raising money to build models which are overthrown in a week.
PC: Yi Tay
