Alexander Wang: If intelligence ceases to be a deficit (AI column)
Watch Harry Tan Talk to Alexander Wang at Startup School 2026 "Alexandr Wang: “This is a Once-in-a-Civilization Opportunity“, published 29 July 2026 years. Wang founded Scale AI and now holds the position of Chief AI Officer at Meta, which is banned in Russia. But more interesting than posts, his main forecast is that intelligence and the ability to act will become abundant, and the lack of vision, ambition and the ability to assemble the system around agents will remain.
Wang believes that even if the models stopped progressing, today’s opportunities would suffice for decades of economic restructuring. Therefore, the dispute about the exact date of superintelligence is secondary to him. The bottleneck is to spread existing capabilities across companies, products, and government systems.
Hence his version of personal superintelligence. (personal superintelligence)Billions of people get personal agents, and millions of companies get agents who negotiate with each other. The developer in this picture rises in levels of abstraction: first he wrote code, then orchestrated individual agents, then he will design organizations of millions or even trillions of agents. But this is still a futuristic forecast, although Wang describes the basic mechanism quite mundanely. You need a closed loop: goal, data, action, feedback and a metric by which the system knows if it has improved. According to him, inside the Meta good agent circuit (agentic loop) quality-tested (eval) It has already allowed a swarm of agents to make more than a hundred engineers, but there is no public record to verify this comparison in conversation. It is interesting that the concept of Loop Engineering stripper Recently with Max Smirnov.
After watching the video with Alexander Wong, I remembered Jeff Dean speaking at Y Combinator School. 2026 me told About ten years ago I came to talk about the future. (it was on YC AI from 7 August 2017 year). Dean then showed TensorFlow, TPU, neural architecture search and formulated “future requests”: describe the video in Spanish; find work on reinforcement learning for robotics and summarize them in German; teach the robot to work safely alongside people in an unstructured environment. The first two scenarios no longer sound like science fiction. The third is still a frontier: models have learned to better understand the scene and instructions, but robust physical action and safety require a separate engineering circuit.
The difference in optics is telling. Dean looked at the model as a new computational tool and asked what problems it could solve. Wang begins by assuming that digital intelligence has become cheap and affordable, and asks who will set a goal and assemble a working organization out of it. The unit of change has grown from model and query to the company and its feedback loops. But there are important similarities. No one predicts the disappearance of engineering. For Dean, human expertise moved into data architecture, computation, and experiment automation. With Wang, she moves into goal setting, orchestration, evals and systems thinking. Abstraction changes, but the responsibility for the design of the system remains.
And here I wouldn't mistake Wang's optimism for a neutral outlook. The conversation simultaneously advances the Meta strategy, Muse Spark and the future agent platform. Estimates of “ten times”, “hundred times” and comparison with a hundred engineers are the position of the speaker, not independent benchmarks.
#AI #Agents #Engineering #Architecture #Leadership #Management