Startup Field Guide by Unusual Ventures: The Product Market Fit Podcast

How open source AI will find product market fit: A conversation with Databricks, and AI startup Together

Sep 25, 2023

Founders of Databricks and Together discuss the rise of open source AI, building with open source, standardization of LLMs, the role of academia in AI research, innovations in training data, growing accessibility of machine learning with LLMs, the future of the open source ecosystem, and best practices for parameterizing LLMs.

46:33

Creator website

Episode guests

Vipul Ved Prakash

Reynold Xin

AI Summary

AI Chapters

Episode notes

Podcast summary created with Snipd AI

Quick takeaways

When using open source models, start with smaller models and gradually try different ones that align with your application goals and performance requirements.

Continuously monitor and improve models by evaluating and experimenting with different models, and leveraging established practices from applied machine learning.

Deep dives

Choosing and Experimenting with Open Source Models

When starting with open source models, it is important to consider the size of the model and how it aligns with your application goals and performance requirements. You can start with smaller models and gradually try different models that rank well on the parameter count. There are resources available, such as Databricks' webpage that recommends best-in-class open source generated AI models for different use cases.

Introduction

2min

The Rise of Open Source AI and LLMs

10min

Limitations of Existing Benchmarks and the Need for Rigorous Evaluation

3min

Potential Consolidation in the Open Source AI Industry

3min

Collaboration between academia and industry in AI innovation

5min

Innovations in Training Data for Open Source AI Models

19min

The Importance of Developing Guidelines for AI Models

4min

Open source AI models have become key drivers of innovation and collaboration. An increasing number of developers and end users are leveraging open source technologies. There is immense potential in the long-term impact of open source AI.

In this episode, we are releasing a conversation on the future of open source AI between Wei Lien Dang (Unusual Ventures), and Reynold Xin (Databricks) and Vipul Ved Prakash (Together).

Join us as we discuss:
3:16: The rise of open source LLMs and foundation models

7:23 Building open source AI platforms to serve customers

10:35 Why Together and Databricks decided to build with open source

13:33 LLMs and the need for standardization

21:09 The role of academia in AI research and innovation

26:57 Innovations in training data

30:55 Making the decision to choose open source models

36:52 Growing Accessibility of Machine Learning with LLMs

40:31 How the open source ecosystem will evolve in the future

47:18 Best practices for parameterizing LLMs over time
Wei Lien Dang is a General Partner at Unusual Ventures and leads investments in infrastructure software, security, and developer tool. Wei was a co-founder of StackRox, a cloud-native security company prior to its acquisition by Red Hat. He can be reached at wei@unusual.vc and

Twitter

Vipul Ved Prakash is the co-founder and CEO of Together. He was also the founder of Topsy and Cloudmark.

Reynold Xin is the co-founder of Databricks. Last valued at $43B, Databricks has been a juggernaut data infrastructure business built on Apache Spark analytics engine. They recently launched multiple AI products including Lakehouse AI and their own open source LLM — Dolly.

Unusual Ventures is a seed-stage venture capital firm designed from the ground up to give a distinct advantage to founders building the next generation of software companies. Unusual has invested in category-defining companies like Webflow, Arctic Wolf Networks, Carta, Robinhood, and Harness. Learn more about us at https://www.unusual.vc/.
Further reading from Unusual Ventures:

Why the future of AI-native infrastructure will be open

How good is your LLM? Nobody know yet

What AI builders should know about data protection and privacy