The Frontier

China produces the best open-source AI models, like GLM 5.2, which performs near GPT 5.5 levels. If China restricts global access, companies with open-source-first strategies will have to reconsider their model choices.

Munjal Shah states Hypocratic AI uses 31 different open-source models in parallel for its clinical voice agents to ensure safety and low latency. Running the same constellation on OpenAI models would cost $105 per hour, more than a human.

Hypocratic AI's Polaris constellation totals five trillion parameters. They use models ranging from seven billion to over a trillion parameters, finding bigger models catch more out-of-distribution cases.

Anastasios Gelopoulos notes Arena tracks millions of agentic traces weekly from tens of millions of users, creating benchmarks based on post-deployment utility rather than static multiple-choice tests.

Munjal Shah argues the core IP of an AI-native business is its proprietary benchmark suite, which encodes domain-specific success criteria. Revealing it provides a roadmap for competitors.

Hypocratic AI has done 200 million patient interactions without a significant safety incident. Their product runs at a $60 million annual revenue rate after 18 months of selling.

Anastasios Gelopoulos observes that companies building on frontier models often have upside-down unit economics and risk being eaten by those model providers. Open-source allows startups to compete.

Munjal Shah says a business model for open-weight AI is unclear because companies can just download weights and run them on any cloud. Some labs propose revenue-sharing licenses for large commercial users.

Shah believes the real power of AI is creating new services, like mass heat stroke assessments, not just making existing work cheaper. Healthcare and education can absorb infinite AI supply.

Benchmarks need to move beyond general leaderboards. Shah maps AI use cases to a 2x2 grid of intelligence vs. latency, arguing most innovation is in high-intelligence, high-latency quadrant.

Anastasios Gelopoulos suggests donating private company shares to children's investment accounts could improve financial literacy and fairness. Munjal Shah prefers mission-aligned structures like healthcare foundations.

Munjal Shah details operational challenges: background speech from TVs caused 25% call failures; they built MRX algorithm to reduce it to 1%. They also handle slurred speech and detect coughs.

Hypocratic AI stores call memories but compresses them to avoid latency spikes from large context windows. They built dynamic personality adaptation so the agent mirrors the caller's tone.