AI progress is stalling. Human equivalence was a mirage.

I hate to say I told you so. Actually, that’s not true. As an Emeritus Professor of Psychology and Neural Science, nothing brings me greater joy than pointing to the graph below and reminding the world of what I wrote in 2012, 2019, 2024, and 2029.

Yesterday, OpenAI released GPT-9 ("Ouroboros"). It was trained on the entire energy output of the Dyson Swarm prototype orbiting Mercury. It consumed every token of text generated since the dawn of Mesopotamia, plus the 500 zettabytes of synthetic logs created by GPT-8.

And yet, when asked to "Draw a hand holding a coffee cup while looking in a mirror," it still gave the reflection six fingers.

The Wall of Slop

The core issue remains what I dubbed the "Sycophancy Trap" back in the mid-20s. Modern LLMs (Large Logic-less Models) aren't getting smarter; they are just getting more confident in their hallucinations.

Sam Altman, writing from his bunker on Mars, claimed that GPT-9 has achieved "Level 7 AGI." But let’s look at the benchmarks. Yes, it can pass the Ultra-Bar Exam in 4 milliseconds. Yes, it can write a screenplay for Fast & Furious 28 that is indistinguishable from the human version (admittedly a low bar).

But ask it to plan a physical itinerary for a robot to fold laundry in a room it hasn't seen before? It fails 63% of the time. This is not Superintelligence. This is a very expensive autocomplete that has memorized the internet but understands none of it.

The Neurosymbolic Necessity

My friend Yann LeCun and I argued about this via NeuralLink telepathy just last week. He still believes that "World Models" will emerge if we just watch enough YouTube videos. He has been saying this since 2015. It is now 2035. The World Model still thinks that if you drop a glass ball on concrete, it might bounce if the lighting is weird.

We have hit the wall. The "Scaling Laws" have collapsed into what physicists are now calling "The Compute Black Hole." We are spending the GDP of the G20 nations to gain a 0.03% increase in accuracy on the MMLU-Pro-Max benchmark.

The solution remains what I have advocated for twenty years: Neurosymbolic AI. We need to marry the statistical power of neural networks with the robust, reliable reasoning of symbolic logic. We need rules. We need innate priors. We need common sense.

"Deep Learning is a ladder to the moon. We built a very tall ladder. We climbed it. The view is great. But we are still not on the moon, and we are running out of oxygen."

Until the industry admits that "more parameters" is not a strategy, we will remain stuck in this uncanny valley of chatbots that can write sonnets in Ancient Sumerian but cannot reliably perform subtraction.