Room 4 of 5
Scale
Not better. Not faster. Bigger. More parameters. More connections.
The answer should be simple: it gets a little better. Like making an engine bigger makes a car a little faster. Linear. Predictable.
Grab the scrubber below. Drag it left: that’s a tiny model, one million parameters. It can complete a sentence. That’s about it. Now drag it right. Slowly.
This demonstrates the general pattern of capability emergence. Exact thresholds vary by model architecture and evaluation method.
Did you see it? “Reason about a problem” didn’t fade in gradually. It popped. One moment the model couldn’t reason. The next moment it could. Nobody added a “reasoning module.” They just made it bigger, and reasoning appeared.
This is a staircase, not a ramp. Some researchers argue the steps are an illusion — that we’re measuring with a coarse ruler. But the practical effect, for the people using these models, is the same: capabilities appear that weren’t there before.
At the far right, the last line lights up in amber: “Do something unexpected.” At sufficient scale, the model does things its creators did not anticipate. This is called an “emergent capability.” It is why the people building these systems are simultaneously excited and terrified.
The people spending fifty billion dollars on the next generation of models are making a bet: that the staircase has more steps. So far, it always has. But “so far” is doing a lot of work in that sentence.
What this means when you use AI
Don’t assume the AI you tested last year is the same product today. Capabilities appear with scale. If you dismissed AI after a bad experience with an earlier model, you may not have seen what the current generation can do.
At the same time, “more capable” does not mean “automatically correct.” A model that can reason about a problem can also reason its way to a confidently wrong conclusion. Capability and reliability are different things. Choose the right model for the task, not always the biggest one.