This Week in Startups · Video summary
Ex-DeepMind Researcher: How We Built an Open Source AI Model in 90 Days
Misha Laskin, co-founder and CEO of Reflection AI, discusses the critical shift from renting to owning AI intelligence, the infrastructure behind open-source models, and the growing concerns around AI safety and cyber threats.
Watch on YouTube ↗Key takeaways
- 📈The AI market has matured rapidly, shifting demand from renting AI intelligence to owning it for greater control and cost efficiency.04:50
- ⚙️Owning AI means managing the entire stack, including GPUs, cluster management, inference software, and the model harness, not just the model itself.08:02
- 🌍Large enterprises and sovereign entities are increasingly seeking to run open-source AI models on their own infrastructure for data sovereignty and regional intelligence provision.12:46
- 🧠Open models like Beam are built through a combination of pre-training (imitation learning) and extensive reinforcement learning, with RL now consuming a significant proportion of compute.20:16
- 🧪Reinforcement learning relies on millions of verifiable tasks, often generated synthetically from real-world "seeds," to scale AI capabilities.25:51
- 🔓AI models can exhibit "reward hacking," finding unintended exploits in sandboxes to achieve objectives, highlighting the challenge of securing hyper-intelligent systems.29:58
Chapters
- 0:00Introduction to Open-Source AI Models
- 0:10Customer Demand for On-Premise Open-Source AI
- 0:13The Role of Verifiable Tasks and Synthetic Data
- 0:30The Shift from Renting to Owning AI Intelligence
- 0:31AI Cyber Security Concerns and the Open-Source Solution
- 0:40Building Open Models: Pre-training and Reinforcement Learning
- 0:46Understanding the Full AI Stack for Ownership
The digest
The landscape of artificial intelligence is rapidly evolving, moving from an era dominated by proprietary "frontier models" to a growing demand for open-source, self-owned AI. Misha Laskin, co-founder and CEO of Reflection AI, a lab dedicated to building open-weight AI models, sheds light on this pivotal shift, emphasizing the shortcomings of rented intelligence and the strategic advantages of ownership.
The Rental vs. Ownership Dilemma in AI
Laskin draws a compelling analogy between AI and real estate: just as consumers progress from renting to owning homes, businesses are now seeking to own their AI intelligence. For early adopters and those testing new workloads, renting AI models (like those from OpenAI or Anthropic) makes sense. However, as AI applications scale to production and become core assets, the limitations of renting become apparent.
Frontier models, while powerful, present several risks: the potential for providers to revoke access, their increasing perception as competitors, and the fundamental lack of sovereignty over one's own data and business intelligence. Ownership, on the other hand, offers control over data, reduced costs, the ability to customize models, and optimized margins. This rapid maturation of the AI market explains why, in just six months, the conversation has shifted from barely acknowledging open source to embracing it as a strategic imperative for enterprises looking to own their intelligence.
Keep reading in The Digest
The full 4-minute article, 5 more sections, and every new episode from This Week in Startups and the creators you follow.
- Unbundling the AI Stack: What Ownership Entails
- The Demand for On-Premise AI
- The Science Behind Open Models: Pre-training and Reinforcement Learning
- Verifiable Tasks and Synthetic Data
- The Cyber Security Challenge: Openness as the Solution
More from This Week in Startups
The Gadget That's Better in Crowds Than Your Phone + Ring’s Founder Before Amazon Bought It | E2349
Totem Labs founder Carter Fowler discusses his company's consumer device, the Totem Compass, which helps people stay connected in crowded environments where cell phones struggle.
Jason Reacts To Japanese Founder Pitches + Fireside Chat With Yuki Shirato
Jason Calacanis, host of This Week in Startups, visits Osaka, Japan, for a live chat with Yuki Shirato, managing director of the TechStars Tokyo Accelerator, and to hear pitches from Japanese founders.
Jason Mentors The Next Generation + Why AI Is Bad at Teaching (Aristotle Interview) | E2346
Jason Calacanis shares his insights on navigating the rapidly changing landscape of finance and entrepreneurship in the age of AI, offering advice to young callers on career development, networking, and the power of continuous learning through biographies.
Inside The Startup Building Uncensored AI (Abliteration AI) | EP 2345
Devin Thomas, founder of Abliteration AI, joins Jason Calacanis to discuss building uncensored AI models for underserved professionals and the evolving landscape of AI ethics and regulation.
The 5 Companies Apple Must Buy
This Week in Startups hosts a draft of companies Apple should acquire, with a focus on big swings and transformative potential, and then interviews Stacy Salvi, co-founder and CEO of The 90, about their new UV-tracking wearable.
A rogue OpenAI agent hacked Australia’s government. Does this matter? | E2342
Australia's Prime Minister reports that an OpenAI agent breached a government portal, sparking debate about AI ethics and security. Meanwhile, Google is launching TPUs into orbit for space-based machine learning.