How AI21 Labs Builds Frontier Models For The Enterprise, with Co-Founder and Co-CEO, Ori Goshen

What if deep learning isn’t the future of AI—but just part of it?

In this episode, Ori Goshen, Co-founder and Co-CEO at AI21 Labs, shares why his team set out to build reliable, deterministic AI systems—long before ChatGPT made language models mainstream.

We explore the launch of Wordtune, the development of Jamba, and the release of Maestro—AI21’s orchestration engine for enterprise agentic workflows. Ori opens up about what it takes to move beyond probabilistic systems, build trust with global enterprises, and balance research and product in one of the most competitive AI markets in the world.

If you want a masterclass in enterprise AI, model training, architecture tradeoffs, and scaling innovation out of Israel—this is it.

🔔 Subscribe for deep dives with the people shaping the future of AI.

This episode is broken down into the following chapters:

00:00 – Intro

00:47 – Why AI21 started with “deep learning is necessary but not sufficient”

02:34 – Building reliable AI systems from day one

03:46 – The risk of neural-symbolic hybrids and early bets on NLP

05:40 – Why Wordtune became the first product

08:14 – From B2C success to a pivot back into enterprise

09:43 – What AI21 learned from Wordtune for enterprise AI

11:15 – Defining “product algo fit”

12:27 – Training models before it was cool: Jurassic, Jamba, and beyond

13:38 – How to hire model-training engineers with no playbook

14:53 – Recruiting systems talent: what to look for

16:29 – How to orient your models around real enterprise needs

17:10 – Why Jamba was designed for long-context enterprise use cases

19:52 – What’s special about the Mamba + Transformer hybrid architecture

22:46 – Experimentation, ablations, and finding the right architecture

25:27 – Bringing Jamba to market: what enterprises actually care about

29:26 – The state of enterprise AI readiness in 2023 → 2025

31:41 – The biggest challenge: evaluation systems

32:10 – What most teams get wrong about evals

33:45 – Architecting reliable, non-deterministic systems

34:53 – What is Maestro and why build it now?

36:02 – Replacing “prompt and pray” with AI for AI systems

38:43 – Building interpretable and explicit agentic systems

41:09 – Balancing control and flexibility in orchestration

43:36 – What enterprise AI might actually look like in 5 years

47:03 – Why Israel is a global powerhouse for AI

49:44 – How Ori has evolved as a leader under extreme volatility

52:26 – Staying true to your mission through chaos

Subscribe to the Barrchives newsletter: https://www.barrchives.com/

Spotify: https://open.spotify.com/show/37O8Pb0LgqpqTXo2GZiPXf

Apple: https://podcasts.apple.com/us/podcast/barrchives/id1774292613

Twitter: https://x.com/barrnanas

LinkedIn: https://www.linkedin.com/in/barryaron/