Gretel

That was one for the books. Huge thanks to everyone who stopped by Gretel's booth, caught a talk, or joined us for happy hour. An unforgettable week at #GTC25 #syntheticdata

What an action-packed week at #NVIDIA #GTC25! I'm still digesting all of the incredible technical sessions I had the chance to attend and will post about them separately. For now, a huge shoutout to our entire Gretel team who absolutely crushed it. It's a privilege to work with you all! The level of interest at our booth was off the charts, and the quality and depth of conversations with partners, customers and prospects was nothing short of amazing. I'm incredibly proud of the work this team has accomplished so far and the groundbreaking capabilities we continue to ship every day. 🙇 The world has seen nothing yet. 💪 Kirit Thadaka, Maarten Van Segbroeck, Ph.D., Alexander Watson, Ali Golshan, Peter S., Cameron Howard, Alex Maestretti, Brian Law, Ph.D., Carlyn Aarish, Maygol Kananizadeh, Eric Pham-Hung, Himanshu Kamra, Mark Weatherford, Paul Smith, Rebecca Kao

What an amazing week at NVIDIA GTC this week. Couldn't ask for a better team and experience with Gretel. I am incredibly proud of everything we have accomplished.

The ABCs of Synthetic Data session starts in 10 minutes. You can still register and catch it here: https://bit.ly/4ih6JkW

Amazing time at the AI after Dark happy hour last night with Gretel, Weights & Biases, @lambda labs and NVIDIA at #GTC

Incredibly proud of Ali Golshan and the Gretel team for this recognition as one of the top 10 most innovative AI companies for 2025. A great start to the year with lots more to come!

Attending #GTC25? Here are two events you shouldn't miss tonight: 💬 Demo: Synthesizing Dialogues at 3 PM Stop by Microsoft’s Booth (#514) to catch Gretel's Product Lead, Kirit Thadaka, demoing how to create synthetic dialogue datasets. These can be used to develop and enhance conversational AI applications, including customer support chatbots, virtual assistants, and interactive learning systems. 🍻 Happy Hour Tonight at 7:30 PM After a full day of compute and GPUs, join Gretel, Lambda, NVIDIA, VAST Data, and Weights & Biases at the “AI After Dark” party for free cocktails and IBUs. Reserve your spot here: https://lnkd.in/e_9mjCzr See you there! #SyntheticData #AI #NVIDIAGTC #SuperBowlAI

We are thrilled that Gretel has been named of The Most Innovative Companies of 2025 by Fast Company. It is great to see the emergence of synthetic data and its impact on AI get the recognition it deserves. 🚀

Honored to see Gretel recognized by Fast Company as one of the most innovative AI companies in the world! This award is a testament to the incredible work our team has put in over the last five years—and a sign that the world is waking up to privacy-preserving synthetic data as the sustainable path to responsible AI. Check out the full Fast Company Most Innovative Companies list here: https://lnkd.in/ekxmwYzp #AI #SyntheticData #Privacy #MostInnovativeCompanies

Synthetic data with embedded reasoning traces is a powerful, accessible way to train AI. Here’s why it makes a difference: Enhanced Transparency and Trust 🛡️ By providing a clear, step-by-step explanation with every synthetic example, errors are easier to spot and correct—ultimately boosting the reliability of AI systems. Systematic Generalization with Human-Like Logic 🥄 When models learn why certain answers are correct, they’re better prepared to handle new problems. A key technique is “spoonfeeding” via synthetic reasoning traces—like the approach used in Phi‑4—where a model builds its reasoning step by step to arrive at robust conclusions. Alignment with Advanced Reinforcement Learning Techniques ⚖️ DeepSeek-R1’s GRPO algorithm rewards correct final answers rather than the quality of the chain-of-thought. Yet, the released model benefits greatly from a supervised fine-tuning warm start on high-quality synthetic traces. There’s room to push this further by including these traces in the reward signal—helping models interpret user cues and make more transparent decisions. Cold-Start Training for Reinforcement Learning 🚀 Even a small, high-quality collection of problem-solution pairs can fill the “empty blackboard” at the start of training. Though these examples may be modest in volume, they can jump-start learning without relying on costly, human-provided data. If you want to generate your own synthetic dataset with everyday scenarios enriched by detailed reasoning traces, check out this blog (https://bit.ly/41yhM3o) or follow along in this notebook (https://bit.ly/4iBKvKx). #syntheticdata #reasoning #LLMs #AI