This image attempts to show our progress to AGI (artificial general intelligence). It comes from a new paper from the Centre for AI Safety defines AGI using human psychometric tests, revealing a “jagged” cognitive profile in current models. GPT-5 scores 58% against a well-educated adult baseline, up from GPT-4’s 27%. But notice the zeros, both models completely fail at long-term memory storage despite excelling at maths and knowledge.
This framework gives us fixed goalposts rather than constantly moving targets. Progress is real but uneven. The next breakthroughs won’t come from scaling alone, they’ll require solving memory, speed, and sensory processing. Current memory solutions are hacks, injecting fragments into conversations rather than true dynamic learning. As OpenAI co-founder and visionary Andrej Karpathy notes in his latest podcast interview, we’re building “ghosts” that mimic text rather than embodied intelligences. These models lack a “cognitive core”, they can’t consolidate experiences like humans do during sleep. The path forward may demand architectural reimagining, not just bigger models.