Ep.07 – Investing in AI, with Amir Shevat, Early stage startup investor, author

In this episode, Amir Schvatz takes us on a journey from his days building products at tech giants like Microsoft and Google to his current role as an investor in early-stage AI startups. 

He discusses the emerging trend of "PM Coding," where humans act as product managers and AI becomes the developer, and shares his thoughts on the future of software creation through AI collaboration.

"As AI becomes a commodity, the real value will come from the layers above it, where humans collaborate with AI to create software." Amir dives into his vision for the future of AI and platform-based investments, offering strategic insights and practical advice for startups navigating the evolving tech landscape.”

🎧 ​​Listen on Spotify and Apple Podcasts

About our guest

Amir Schavtz is the author of Designing Bots by O'Reilly. Currently, he is an investor at Dark Mode, a venture fund focusing on early-stage startups, particularly those in the fields of enterprise AI, modern data stacks, and developer platforms. 

Before transitioning into venture investing, Amir built a notable career at major tech companies. He started his journey at Microsoft, then moved on to Google, where he worked on Chrome, Cloud, and Android. Amir’s entrepreneurial spirit led him to join Slack just two weeks before its platform launch, where he contributed to scaling the product to 250,000 weekly active developers. He also worked at Twitch, managing the company’s API, before founding his own startup, Reshuffle, which was acquired by Twitter.

Following the acquisition, Amir worked at Twitter, where he helped open the API under Jack Dorsey’s leadership. However, his time at Twitter came to an end when Elon Musk’s acquisition led to the closure of the API and a restructuring of the platform.

As an investor, he focuses on supporting innovative companies that build infrastructure for the next wave of digital tools, particularly in AI and developer tools.

0:00:00 - Intro

0:02:09 - Vibe coding

0:05:33 - Running AI code in a sandbox to test your vibe coded applications

0:07:29 - Genetic algorithms and organic coding

0:09:06 - Incentives not guardrails to guide both AI and people

0:11:15 - Social media with the user as the product, and the rise of Bluesky

0:18:32 - What Amir looks for when investing in a company

0:23:37 - Companies still need someone who understands the domain to drive growth

0:26:46 - Ideas that Amir wouldn’t invest in yet

0:31:20 - The cost of AI, and how specialised models would help reduce cost

Insights and quotes from the episode

Amir Schvatz: "My name is Amir Schvatz. I turned to the dark side about two and a half years ago and became an investor in a fund called Dark Mode. We invest in early-stage startups, focusing on developer platforms, enterprise AI, and the modern data stack. Before this, I worked at Microsoft, Google, Slack, and Twitch. I later started my own company, Reshuffle, which was acquired by Twitter. Afterward, Jack Dorsey hired me to open the API at Twitter, and then Elon closed it down."

Jim: "Awesome, it's great to have you on the show! You've done so much, it's fascinating to see you transition from building things to now supporting emerging companies. At Dark Mode, you're investing in the early-stage phase. What’s the hottest trend you’re excited about right now?"

Amir Schvatz: "The hottest current trend I’m seeing with AI is what I call 'PM Coding.' I prefer that term over 'Vibe Coding,' which is the idea that humans collaborate with AI to create software. I think humans should be the product managers, while AI acts as the developer. It’s a much deeper way to look at how we’ll work with AI in the future."

Jim: "Do you think vibe coding is a viable model for building software, or do you think it’s still too early, like creating MVPs that need a lot of fixing?"

Amir Schvatz: "For the 80% of boring, repetitive tasks in software development, vibe coding will be amazing. These tasks often involve CRUD applications, which are essentially just wrappers around databases or APIs. But for the 20% of truly complex, cutting-edge development done at places like Google or Facebook, that’s still going to require experienced developers."

Amir Schvatz: "Right now, vibe coding is in its infancy, and it’s easy to dismiss it as 'bad.' But just like the early days of mobile phones, which were bulky and had short battery life, vibe coding will improve rapidly. I’m already building something with vibe coding, and what I’ve accomplished in just a few nights is better than what most engineers could do with the same time."

Jim: "It sounds like you’re optimistic about the evolution of vibe coding. Do you see it working well in a sandboxed environment, like testing applications without breaking production?"

Amir Schvatz: "Yes, absolutely. There’s a huge opportunity in AI infrastructure, and platforms like Daytona, which allow for testing AI in sandboxed environments, are key. Integrating tools like Heel, an AI black box tester, with vibe coding could create an ecosystem that makes building software both easy and reliable. It’s an exciting area for innovation."

Jim: "I like the metaphor of organic software evolution you mentioned earlier. Could AI continuously tweak software to improve it over time?"

Amir Schvatz: "Yes, exactly! Imagine AI looking at how users interact with a page and automatically evolving the design—like making a button more appealing or changing copy to boost conversions. In the future, software could evolve organically based on user interactions, continuously optimizing itself."

Jim: "What about the incentives in AI? Some worry that AI could evolve in unhealthy ways, like promoting anger-driven engagement for clicks. How do we balance this?"

Amir Schvatz: "Incentives matter more than guardrails. If the AI’s incentive is to sell, it won’t optimize for rage-inducing content because that doesn’t lead to sales. Guardrails are important, but incentives are the more powerful driver of behavior. If the AI learns that positive engagement drives better results, it will naturally evolve toward that."

Jim: "So AI will eventually learn to optimize for positive outcomes based on its incentives?"

Amir Schvatz: "Exactly. If AI is incentivized correctly, it will understand what works best and align itself to those goals. It’s about creating the right environment for AI to learn and grow, much like a good manager gives challenges to their team instead of telling them exactly what to do."

Jim: "Switching gears, you’ve mentioned platforms as a key area of investment. What is it about platforms that excite you?"

Amir Schvatz: "Platforms are a personal mission for me. I believe that platforms enable people to build amazing things. When I worked at Android, I saw how people became successful on the platform. The same happened with Twitch and other big companies. Platforms allow for immense scalability, and that’s what I look for when investing—something that can power many innovative tools, not just a single product."

Jim: "So you’re looking for companies that provide the infrastructure upon which other tools can be built?"

Amir Schvatz: "Exactly. Platforms create the foundation for a whole ecosystem of tools. Companies like Daytona are building platforms that provide a much-needed infrastructure for running AI. When investing, I look at the technology, the team, and whether it addresses a real pain point in a large market."

Jim: "What’s your view on AI’s cost and its future?"

Amir Schvatz: "AI is currently expensive, but it’s only going to get cheaper. As competition grows, especially from places like China, we’ll see the cost of running AI drop significantly. We’re also going to see more specialized AI models for specific tasks, which will be cheaper to run than general-purpose models. Just like how sugar started expensive but became cheap over time, AI will follow that same trajectory."

Jim: "Do you think the future lies in specialized AI models rather than general-purpose ones?"

Amir Schvatz: "Yes, absolutely. We’re already seeing specialized models like the coding models from Alibaba, which are small but powerful because they’re focused on one task—coding. The more targeted the AI, the cheaper and more efficient it will be. In the future, we’ll have specific models for specific tasks, and that will help with predictability and reduce errors."

Jim: "This is a fascinating discussion. Thank you so much for your time, Amir."

Amir Schvatz: "Thank you for having me. It's been great to share my thoughts on AI and the future of software creation."