The conversation builds on Joanne and Leo’s recent essay on the great reorg, which argues that AI’s full impact will only arrive when companies redesign how work gets done from first principles. 

Today, AI is already helping individuals work faster, but many organizations are struggling to translate that into gains at the team or company level. Azeem calls this problem “congestion.” When one part of the company speeds up, the next downstream process becomes the bottleneck. As AI takes on more work inside organizations, it reveals where they are too slow, too rigid, or too dependent on processes built around human constraints.

Joanne and Azeem explore what it will take to build AI-native companies, from new human roles like system architects, validators, and accountability owners to new tools designed for agents rather than humans.

For founders, Azeem argues that the most important metric may be cycle time: how quickly a company can learn from customers, ship, adapt, and repeat. The org chart is being redrawn, and the companies that thrive will be the ones that learn how to collaborate with agents instead of simply bolting AI to old workflows.

What we covered:

• 00:00 Cold open: Why this moment favors startups
• 00:28 The great reorg, explained
• 02:16 AI’s productivity paradox
• 03:36 Congestion as the new bottleneck
• 06:30 What congestion looks like in practice
• 09:16 Bottlenecks across compute, healthcare, and drug discovery
• 13:44 The new human roles in AI-native organizations
• 15:33 The growing importance of accountability as agents do more work
• 22:09 The future of expertise
• 27:27 Are you managing the agents, or are the agents managing you?
• 31:28 What remains human in venture capital
• 33:08 Throwing out old assumptions about process
• 37:12 Designing for agents, not humans
• 39:19 Cycle time as a key AI-native metric
• 44:15 The mental model that stalls AI adoption in large orgs
• 48:27 Azeem’s advice for early-stage founders
• 51:19 New startup opportunities in the great reorg
• 55:08 What changes in the next three years

Links: 

Exponential View - https://www.exponentialview.co

The great org: A human’s guide - https://foundationcapital.com/ideas/the-great-reorg

Listeners of AI in the Real World can receive complimentary access to Exponential View Premium for three months, visit: https://www.exponentialview.co/airealworld. The offer is valid through Wednesday, June 24.

Tennr focuses on patient flow: the process of getting patients from one provider to the next with the right records, insurance information, clinical context, and scheduling in place. It's a problem that sits in the messy middle of healthcare, where delays, denials, missing documentation, and manual workflows are the norm. Today, its platform processes referrals and intake documentation for roughly 10% of Americans each year.

Healthcare has long been considered too hard for startups to crack. But advances in LLMs have made it possible to build products that can handle the unstructured, edge-case-heavy nature of healthcare administration. As Trey explains, those advances have created an opening for founders willing to stay close to customers and learn the market directly.

Trey also gets into his first-time founder lessons and Tennr's path to scale. In the early days, he and his co-founders spent too much time building in isolation, then swung too far the other way by saying yes to nearly everything customers asked for. At one point, they were approaching $1M in revenue, only to realize they were spreading themselves across too many use cases. The decision to refocus resulted in a short-term hit to their revenue, but became the moment the team locked in and found product-market fit.

Along the way, Trey reflects on the discipline of questioning your own assumptions, how Tennr uses AI internally, and what the patient experience could look like if Tennr realizes its mission of making healthcare move faster.

Chapters:

• 00:00 Cold open
• 00:35 What Tennr does
• 03:20 Tennr's road to PMF
• 07:40 Startup advice that is true but hard to follow
• 11:30 Going against conventional wisdom: why selling to healthcare was worth it
• 13:20 Why Tennr is not an AI company
• 16:21 What it takes to get AI to work in healthcare
• 17:44 How Tennr thinks about building an industry-level context graph
• 19:30 Why focused product companies beat general-purpose labs
• 21:10 The importance of fresh perspectives in regulated industries
• 22:40 Founders Trey looks up to
• 24:27 How Tennr uses AI internally
• 27:20 AI's impact on productivity and the future of org charts
• 28:34 What Tennr could look like in five years

Ken has spent more than four decades working at the intersection of robotics and AI, focusing on one of the most persistent challenges in the field: how machines perceive and manipulate the physical world.

Ken shares why tasks that seem trivial to humans, like picking up a glass or folding laundry, remain profoundly difficult for robots, and why the physical world introduces a level of uncertainty that can't be fully simulated.

Our conversation also covers:

• What it would take to reach a ChatGPT moment in robotics
• Why simulation data is not enough without real-world grounding
• And why the next decade of robotics depends on combining cutting-edge models with good old-fashioned engineering

Chapters:

• 00:00 Cold open: Why robotics still needs good old-fashioned engineering
• 03:46 Hype cycles and winters in robotics
• 05:08 Why folding laundry is still hard for robots
• 10:38 What robots are good at today
• 15:00 Automation and the rise of warehouse robotics
• 19:39 Can LLMs and generative AI work for robotics?
• 26:52 The limits of simulation data and the sim-to-real gap
• 29:44 Why humanoids are still far from practical
• 36:34 What founders need to know about robotics timelines
• 37:08 Why robots need grounding and exploration
• 39:00 Combining the power of LLMs with traditional engineering
• 40:42 Why Ken is optimistic about the future of robotics

Karan shared his journey from pursuing a PhD at Stanford to co-founding Cartesia with Albert Gu - taking their academic research on SSMs and turning it into a product that is initially focused on voice applications.

Karan explained how SSMs differ fundamentally from transformers, offering a more efficient, memory-based architecture that processes information sequentially rather than in batches. This approach enables more human-like AI systems that can remember and adapt to new information in real-time.

The conversation also covers:

• How SSMs offer significant efficiency advantages for applications like voice agents and on-device AI
• Why architecture innovation remains crucial alongside advances in data and training techniques
• The challenges and opportunities in the voice AI space
• His perspective on the open-weights vs. closed-model debate in AI
• Advice for researchers who want to become founders

Chapters:

• 00:05 Cold open
• 01:47 Karan's founder journey
• 03:45 Transitioning from academia to entrepreneurship
• 05:17 The art of recruiting
• 09:22 Karan's vision for Cartesia
• 13:36 Innovation in AI architectures
• 16:33 Advantages of State Space Models (SSMs)
• 19:34 The false dichotomy: models vs. applications
• 21:17 Best practices for voice AI
• 25:10 Thoughts on the broader AI landscape
• 31:37 Open source vs. closed models
• 35:00 Karan's thoughts on fundraising
• 41:27 Advice for founders building AI startups
• 46:03 Upcoming trends that get Karan excited

Within the AI community, Rohan is best known for co-creating Alpaca, a project that demonstrated how fine-tuning Meta's LLaMA model could achieve ChatGPT-level performance for under $600.

Rohan shares his journey from early work in computer vision at UC Berkeley to his Ph.D. at Stanford, where he explored methods for making AI more accessible.

He explains the technical breakthroughs behind Alpaca, including self-instruct, a method that uses a stronger language model (OpenAI's text-davinci-003) to generate synthetic data that is then used to fine-tune a weaker model (Llama). This approach, which underpins Alpaca and its follow-up projects AlpacaFarm and AlpacaEval, illustrates how small-scale post-training can significantly enhance model performance.

The conversation also covers:

• The promise and challenges of synthetic data for training AI models
• What it will take to build foundation models that are 100x better
• The future of multimodal AI and why it matters
• Why better evals are critical to the next wave of AI advances

Chapters:

• 00:00 Cold open
• 1:40 Rohan's journey into AI research
• 04:50 Transitioning from vision research to LLMs
• 06:18 The story behind Alpaca
• 08:55 How Alpaca works
• 10:45 The AI community's reception of Alpaca
• 12:26 The evolution of Alpaca related projects
• 14:22 The role of synthetic data
• 19:38 Challenges in multimodal AI
• 24:31 Future of foundation models
• 30:00 Importance of data in AI
• 34:48 Staying up to date with the latest AI research
• 36:12 Advice for founders

HiddenLayer is a cybersecurity platform purpose-built for AI systems. Tito's journey into AI security began at Cylance, where he led the response to one of the first real-world AI attacks. This experience led him to start HiddenLayer, which recently raised a $50M Series A.

In this conversation, Tito shares his insights into the evolving AI threat landscape, from shadow AI to the growing arsenal of strategies that bad actors are using to target generative models. He explains why data poisoning is one of the most urgent threats facing humanity, and why safeguarding AI will define the future of cybersecurity.

Learn more about HiddenLayer here: https://hiddenlayer.com

Chapters:

• 00:00 Cold open
• 04:37 The inflection point of AI
• 10:57 Evolving AI threat landscape
• 11:19 Understanding AI vulnerabilities
• 17:20 The dangers of data poisoning
• 23:47 Generative AI attacks
• 29:50 Shadow AI risks
• 33:01 How HiddenLayer works
• 41:16 Lessons from AI security engagements
• 44:17 Tito's thoughts on the future of AI security
• 50:14 Optimism for AI innovation

Melody has been a pioneering technical leader and co-founder of two startups. She spent 15 years running engineering teams at Google, where she helped shape Google Cloud and improve search. Here, she shares her journey from the early days of search to today's cutting-edge work in generative AI for marketers at Jasper.

• 00:00 Intro: What does it mean to build a sticky product?
• 01:08 Melody's journey in tech
• 03:42 Building trust in AI Systems
• 07:56 The evolution of AI and current market dynamics
• 12:34 Melody's thoughts on AI wrappers vs. AI native products
• 14:53 Data strategies for AI in production
• 17:43 Developer experiences with AI
• 21:16 Melody weighs in on Multimodal AI
• 22:24 The role of agents
• 25:48 Challenges in AI Hiring
• 31:54 The future of contextual AI
• 35:14 Enhancing developer productivity

I'm excited to chat with Darius Lam, founder and CEO of NEX, a recent addition to our portfolio. Darius was previously head of product for computer vision at Cerebras, another Foundation Capital company that's building wafer-scale chips for AI workloads.

Today, Darius develops state-of-the-art multimodal foundation models at NEX. Our conversation explores the ins and outs of creating these models, including training efficiency, output controllability, and the last mile of AI deployment. I also get Darius's take on the evolution of open source AI and what every AI founder should know before they write their first line of code.

Chapters:

• (00:00:00) Intro
• (00:03:36) Darius's journey into computer vision
• (00:06:50) Founding NEX
• (00:09:33) Training a foundation model from scratch
• (00:11:14) Key innovations at NEX
• (00:12:26) Challenges of competing as a startup
• (00:15:00) What NEX hopes to achieve
• (00:18:58) Why vertical integration matters
• (00:21:18) The development stack for AI-native products
• (00:28:49) User growth and education
• (00:31:54) Where AI is heading next
• (00:39:29) Lessons learned from building NEX

Joining me is Sean Lie, co-founder and CTO of Cerebras Systems. I've known Sean since early 2016 before leading the Series A investment in Cerebras. At the time, Sean was just wrapping up a three-year stint as Chief Architect of data center solutions at AMD, which he joined via AMD's acquisition of his last startup, SeaMicro.

Sean is one of the biggest brains I know around hardware systems, bringing a rare expertise and passion for advanced architectures, many of which break from tradition and traditional approaches to computing.

In this episode, Sean guides us through the world of AI hardware innovation. He breaks down the concept of wafer-scale processing, the challenges of building monstrous AI chips, and the massive potential of sparsity. He offers valuable lessons from his journey as a deep tech entrepreneur, including the importance of thinking end-to-end and planning for scale from day one.

We close by peering into AI's future and sharing advice for founders taking on similarly complex, system-level problems.

• (00:00:00) Introduction
• (00:02:28) Sean's background in hardware systems
• (00:05:26) Cerebras' origin story: Recognizing the need for AI-specific hardware
• (00:10:13) Wafer-scale processing and its advantages
• (00:15:50) Technical challenges and innovations in wafer-scale integration
• (00:19:18) The importance of end-to-end system design in AI hardware
• (00:24:54) Cerebras' approach to scaling AI models
• (00:28:27) Introduction of the Wafer-Scale Engine 3 and its capabilities
• (00:32:00) Sparsity in neural networks and Cerebras' hardware acceleration
• (00:36:49) What's next for AI hardware?
• (00:37:38) Closing thoughts: Advice for deep tech entrepreneurs

Joining me today is Eugene Cheah, the founder of Recursal AI. Recursal is working to commercialize RWKV, a new AI architecture that Eugene played a key role in developing.

Eugene believes RWKV can address some of the main limitations of transformers - the technology behind models like GPT and Claude that have fueled the recent AI boom. By combining elements from both transformers and RNNs, RWKV aims to make major improvements in how efficient, scalable, and accessible AI can be.

In our discussion, Eugene takes us through his journey, from the early days of building RWKV's open-source community, to launching Recursal to commercialize it. He breaks down how RWKV works, what sets it apart from transformer models, its potential benefits, and its most promising use cases.

Read the notes from our conversation here: https://foundationcapital.com/whats-next-after-transformers/

• (00:00:00) Cold open
• (00:02:14) Eugene's early projects and foray into AI
• (00:05:26) Recursal's origin story
• (00:10:28) Differences between RWKV and Transformers
• (00:15:50) Use cases enabled by RWKV
• (00:19:18) Why AI efficiency matters
• (00:24:54) Recursal's roadmap: Scaling up and adding multimodal support
• (00:28:27) Current hurdles: Scaling and funding challenges
• (00:32:00) Future research areas in AI: Datasets and diffusion models
• (00:36:49) Resources for learning about alternative architectures
• (00:37:38) Closing thoughts on the future of AI

By enabling AI agents to collaborate, learn from each other, and contribute their unique skills, AutoGen is unlocking a new frontier of AI capabilities. It's quickly gained traction among both academics and enterprises and is currently powering a wide range of use cases, including synthetic data generation, code generation, and pharmaceutical data science.

In our conversation, Chi breaks down the core concepts behind multi-agent AI, the pros and cons of multi-agent architectures, and the real-world use cases enabled by AutoGen. He also shares some of the open research challenges he's tackling, his perspective on the future of AI, and what excites him most about where the field is headed.

Read takeaways from our conversation here: https://foundationcapital.com/the-promise-of-multi-agent-ai/

• (00:00) Intro
• (01:27) Chi's background and early interest in AI
• (05:42) Defining agents and their core capabilities
• (08:13) Pros and cons of multi-agent systems
• (11:23) Multi-agent architectures and the Society of Mind theory
• (14:36) Real-world use cases enabled by multi-agent systems
• (16:45) The backstory and genesis of AutoGen
• (19:43) How AutoGen's architecture leverages language models, tools, and human input
• (23:13) More examples of AutoGen's diverse applications
• (29:23) How AutoGen is being used in enterprises and production
• (32:42) Advice for AI builders focusing on the enterprise
• (40:24) What's next for AutoGen and open research challenges
• (47:17) Resource recommendations for AI builders
• (48:09) What excites Chi most about the future of AI

To kick things off, I'm excited to speak with Joon Sung Park, a Ph.D. student in computer science at Stanford. Joon works at the intersection of human-computer interaction, natural language processing, and machine learning. He's best known for his pioneering research on generative AI agents.

We break down how generative AI is transforming agent design, share advice for builders working with these models, and unpack why we haven't yet found the "killer app" for LLMs.

Read the top takeaways from our conversation here: https://foundationcapital.com/the-future-of-generative-agents/

• (00:00:00) Introduction
• (00:04:00) The impact of LLMs on agent capabilities
• (00:08:07) Tool use and simulations are the two main focuses of agent R&D
• (00:10:34) Multi-agent systems and their applications
• (00:14:15) LLMs represent a new paradigm for agent design
• (00:18:55) Multimodal models promise step-change improvements in agent performance
• (00:23:55) Unsolved challenges: grounding, representativeness, and scalability
• (00:26:40) Soft-edged problems are better bets for AI builders
• (00:33:00) The future of transformer architectures
• (00:38:49) Why agents need to solve real user needs
• (00:39:59) We haven't yet found the killer app for LLMs
• (00:44:47) Classic product principles still apply when building with LLMs