Sign In

Shayan Mohanty

Head of AI Research @ Thoughtworks. Prev. Watchful (Acq: TWKS), Meta

Shayan Mohanty is an accomplished professional with a strong educational background in Economics from the University of Texas at Austin. With a diverse work history, Shayan has held several key positions in reputable organizations across various industries, showcasing his expertise and adaptability in the tech and AI fields.

Highlights

Apr 17 · linkedin

Today, I’m both thrilled and nostalgic to share that our talented team and tech @ Watchful has been acquired by Thoughtworks. This milestone marks a new chapter not only for our company, but for me personally.

Reflecting on the journey, I’m filled with gratitude. Starting Watchful was a leap into the unknown, which would not have been possible without the early support from our dedicated investors and partners. Your belief in us and our vision has been instrumental to getting us here today. To our incredible team and every customer who joined us on this ride — thank you for turning this dream into reality.

The decision to join Thoughtworks was actually fairly straightforward; our goal as a company has always been to help shape the AI industry towards a data-driven and sustainable future. It turns out that Thoughtworks has done exactly that several times before (see: Agile, Microservices, Data Mesh, etc.). We’re excited to be able to tackle this challenge together.

As I take on my new role as Head of AI Research, I’m eager to address some of the core challenges that prevent AI-enabled technologies from achieving true ubiquity. We’ll be bringing our scrappy startup energy to bear on these problems, and look forward to engaging closely with the community and our partners.

To everyone who’s been in our corner: thank you for taking a bet on us.

(This is certainly not an exhaustive list, but a smattering of folks who helped us along the way!) Xiao Guo Rachel Laycock Prasanna Pendse Mike Mason Mark Boland Ashu Garg Joanne Chen Foundation Capital Astasia Myers Morgan Livermore David Greenbaum Quiet Capital Maarten Hooft Quest Venture Partners John Frankel Conor Webb ff Venture Capital Semil Shah Haystack Simon Crosby Pete Soderling Itamar Novick

Sep 11 · linkedin

Just published a deep dive into estimating token importances in #LLM prompts using text embeddings! You can read about it here: https://lnkd.in/gqBQSM5x

and play with the demo here: https://lnkd.in/gXRKkTYZ

As we interact more with LLMs like ChatGPT, how we prompt these models becomes crucial. But crafting the 'perfect' prompt remains an art. How do we make it more of a science? How can we interpret how our input affected the output of increasingly opaque models?

Our hypothesis: There's a relationship between a model's embeddings and its interpretations. If we can tap into this, we might uncover token importances without diving deep into the model itself.

Initial results were intriguing! Especially when comparing GPT-2 vs. GPT-3 embeddings against GPT-2 Integrated Gradients. There were surprises and plenty of learnings.

One key takeaway: The quality of embeddings seems to hold significant predictive power over model output. It seems that the better the embedding, the more insight we gain, irrespective of the target model's architecture. 🧠

This is just the beginning! There's so much more to refine, test, and explore. But the initial signs are promising. More to come - we think this is the first step towards more automation in prompt engineering.

What are your thoughts on prompt engineering and interpretability in closed-sourced LLMs? Do you have your own techniques that help you with LLM explainability? Let's discuss! 🚀

#NLP #ChatGPT #ML #AI

Shayan Mohanty
Add to my network

Location

United States