note1n0jamfsr9ln09njjqs06ltc36nrwaa4qmcsyr43chdrvpa55szyqn5kqhv
This made me think about the other side, how much more intelligence will there be from today (at the frontier), independent of compute required (within the same fundamental autoregressive transformer and pretraining + RL paradigm)? My gut says at least another OG GPT-4 to today's frontier jump (Grok-4/Gemini 2.5 Pro/Claude Sonnet 4/GPT-5)
But I think a qualitative jump to really feeling like human-level intelligence will require something like a module within the network that pays attention to the coherence of a new "thought direction" WRT its own model of reality, not just next-token-prediction on steroids.
(that qualitative jump would manifest as a huge hallucination reduction, the ability to zoom out and self-correct after mistakenly honing in on the wrong path, in short, the model feeling like it's responding from a place of operating from a real mental model of the world, as opposed to today's models, responding from a place of most-likely-next-token)
And a neurology foray I made a few days ago leads to say
"wow neurology and ML are starting to converge, and some of the computational models that neurologists are making strike me as believable substrates for this type of 'check thoughts against a causal world model' module I think is required for the next step in LLMs". That didn't really feel true to me a few years ago
Examples of neurologist's recent models I looked at: TEM machine (entorihnal+hippocampus analogous), Spaun (spiking NN, surprisingly holistic macro brain model)"
And recently HRM or Heirarchical Reasoning Models (did well on the ARC-AGI-1 benchmark, caused a bit of a stir) strike me as the latest thing closing the neurology/AI gap, and accomodating insights from both.
Excited and kinda concerned for the next 5 years.
This is the time to build AI tech that gives power to individuals! Otherwise we might be SOL, at the mercy of oligarchical whims...