Tiago Leite Ferreira

We are currently witnessing a phase shift in the discourse around AGI. We’ve spent the last decade obsessed with the vertical Y-axis of intelligence, asking if models can score higher than a human on the LSAT or reach an IQ of 300. But looking at the recent saturation of benchmarks, I’m starting to think we are focusing on the wrong variable. The high end of machine intelligence is already very high. Even if we assume there is a hard mathematical constraint on intelligence—a diminishing returns curve where being smarter stops yielding better results—it doesn't matter. The real revolution isn't about reaching God-mode IQ. It's about velocity and volume. We are moving from a world where high-level cognition operates at biological clock speeds to a world of silicon speeds. Even if an AI never exceeds the reasoning capability of a smart human graduate student, the ability to spin up 100 trillion instances of that student and run them at 100x real-time is a force multiplier that human intuition struggles to comprehend. However, if intelligence is about to become infinite and instant, why won't we solve physics, biology, and energy overnight? This is where the naive Singularity argument breaks down. It assumes that intelligence is the only bottleneck to progress. It ignores the second, much harder constraint: Information. You can think of Useful Intelligence as a function of two inputs: Compute and Data. We are solving the Compute side, but the Data side is governed by the laws of physics, specifically entropy. If you lock a super-intelligence in a Faraday cage and ask it to cure Alzheimer's, it will fail. It doesn't matter if it has an IQ of 50,000. It will fail because the solution to Alzheimer's is not a logic puzzle hidden inside its training weights. It is a biological reality that exists outside the box. The model lacks the bits of information describing how specific proteins fold in a chaotic, noisy cellular environment. Intelligence is, at its core, a search space optimization engine. In closed systems like Chess or Go, the search space is massive, but the information is perfect. The rules are rigid. The board is the world. An AI can explore this space purely through self-play. Reality is not Chess. Reality is sparse, noisy, and high-entropy. You cannot simulate your way to ground truth if you don't have the priors. This brings us to the concept of the Useful Ceiling. The value of machine intelligence follows a sigmoid curve that is about to slam into the wall of physical verification. Initially, utility is low. Then, as we see now, the AI operates as a master compression algorithm, reading the entire internet to reason and code, causing utility to explode. Finally, we hit the ceiling where the AI generates hypotheses faster than we can verify them. This is the merging of the curves. The curve of Problem Difficulty is being matched by Machine Intelligence, but the bottleneck effectively shifts from Reasoning to Experimentation. A super-intelligent agent can look at a material science problem and identify the three molecular structures most likely to result in a room-temperature superconductor. That is an incredible reduction of the search space. It turns a million-year search into a 5-minute inference. But you still have to build it. You still have to synthesize the molecule. You still have to run the clinical trial. You still have to stress-test the bridge. The AI can think at the speed of light, but it can only act at the speed of atoms. We are entering an era where the cost of generating a brilliant idea drops to zero. The new scarcity is not smarts—it is the bandwidth of the physical world. The Useful Ceiling of machine intelligence is the point where the cost of computing the answer becomes negligible compared to the cost of verifying the answer against the entropy of the real world. We aren't waiting on the AI to get smarter. The AI is waiting on us to build better robots, sensors, and labs to feed it the data it’s hungry for.