20VC

20VC: Jev Creator on Why Data Is More Important Than Compute | Will US Open-Source Models Catch Chinese | The Cost of Intelligence and Tokens: What No One Sees But Needs to Know with Diogo Almeida

16 Nov 2026 · 74 min · Episode page ↗

The true bottleneck to AI's transformative power in software is not its intelligence or speed, but its unreliable decision-making, which prevents full automation.

Summary

This conversation challenges conventional wisdom in the AI landscape, asserting that while AI models possess remarkable intelligence, their inconsistency in critical decision-making renders them unsuitable for full software automation. The guest argues that the current focus on scaling compute and optimizing for human-in-the-loop applications (like chatbots) misses the core problem, advocating for a shift towards 'machine-native AI' that prioritizes reliable, automated decision-making to unlock a 'Cambrian explosion' in software, rather than merely making existing human tasks slightly more efficient.

Interesting ideas

Reliability of Intelligence is AI's Biggest Bottleneck. AI models are highly capable but inconsistently reliable; they might perform complex tasks like solving Millennium Prize problems, yet fail at basic accounting. This 'Einstein 95% of the time, Mr. Bean 5% of the time' phenomenon makes them untrustworthy for critical automation, which is why software hasn't been truly transformed. The implication is that without addressing this intelligence reliability, AI will remain primarily a human-in-the-loop assistant, rather than a fully autonomous agent.
Data Trumps Compute, and the 'Right Task' Trumps Data. The belief that scaling compute alone leads to AGI has plateaued; instead, data is far more critical for model improvement, a lesson learned from applied machine learning. More profoundly, defining the 'right task' for AI is paramount because it dictates what data is needed and how it should be structured. Optimizing for human feedback or benchmark scores creates models reliable for those specific tasks, but not necessarily for background software automation, highlighting a misdirection in current AI development.
AI is a Bubble Driven by Miscalibrated Expectations. The guest believes the AI sector is in a bubble, characterized by overinvestment and a 'FOMO' mentality, rather than a clear understanding of AI's actual capabilities. Demand for AI could burst when companies confront the reality of its limitations, especially its unreliability for high-stakes decisions. This contrarian view suggests that many organizations are miscalibrating AI's potential, leading to wasteful spending and eventual disillusionment.
The Future of Software: Machine-Native AI and a 'Fourth Logic Gate'. Instead of AI as a 'co-worker' or a thin wrapper around existing software, the real revolution lies in treating AI as a 'fourth fundamental logic gate' alongside existing computational primitives. This would allow for a 'Cambrian explosion' where software can automate tasks previously deemed impossible, moving beyond mere efficiency gains to entirely new capabilities. This vision requires a shift in how developers approach problem-solving, integrating AI directly into core logic rather than as an add-on.

Questions worth thinking about

If 'reliability of intelligence' is the core problem, what new metrics or evaluation frameworks are needed to measure it effectively for automation?
How might the 'right task' for AI in sectors like healthcare or finance differ from current human-in-the-loop applications, and what specific data would be needed to train models for these new tasks?
If AI drives a 'mega Jevons paradox' of demand, what are the societal and economic implications for energy consumption and resource allocation?

Written by AI from the episode's full transcript. It can get things wrong: if something here looks off, tell us at hello@podpings.com.

More from 20VC

All 28 20VC summaries →