How I Define AGI
A follow-up essay: a working definition of AGI, how it will come about, and the precautions it needs.
Studying the path to AGI.
I'm a student working toward original research in neuroscience, recording how I think we actually get there.
Academics
School and activities. Full resume →
- Class of 2029Bala Cynwyd Middle School
- 2nd yearMath Team
- 2nd yearTSA (Technology Student Association)
- 1st yearScience Olympiad
- Yellow ballTennis
Writing
Notes and essays about intelligence, AI, and the questions I keep coming back to.
My Perception of AGI
Why the brain's efficiency may be an important clue for building more general AI.
Research
Investigations I'm actually running.
RNN Flip-Flop Study ResearchComplete
Comparing a regularized and an unregularized RNN trained on the classic 3-bit flip-flop task, using the fixed-point analysis methodology from Sussillo & Barak to see how regularization reshapes the dynamics the network actually learns.
"Does regularization change the underlying dynamics an RNN learns to solve the flip-flop task, or just how cleanly those dynamics can be read out?"
Projects
What I'm building to test the ideas I write about.
Ponder EngineeringWorking
A resident assistant for my laptop, built in an evening and then rather a lot longer. It triages my inbox, keeps my calendar and task list, tells me who I've forgotten to reply to, drafts replies I still have to send myself, and clears out the caches quietly eating my disk. It is called Ponder, after the thinking it leaves me time for. It is unavailable for now while I make something new.
"An assistant is only useful if it remembers — and only safe if every irreversible thing it can do has a human in front of it."
AGI Development Plan PositionConcept
A plan for how I think AGI actually gets built: stop feeding models the internet, and start feeding them the way a baby learns — small amounts of real, embodied, sensory experience gathered over time, instead of scraped text by the billions. Laid out in full, including the hardware gap it still has to solve and the safety precautions it needs from day one, in "How I Define AGI."
"If sample-efficiency is part of what makes intelligence general, simply scaling today's architectures may never cross the threshold into AGI — no matter how good the benchmark scores get."
Now
What I'm working on and reading at the moment.
- BuildingPlanning the next test for the RNN flip-flop study: retraining both networks from different random starting points. The final paper draft is on its page.
- ReadingAoPS Calculus — David Patrick; Gödel, Escher, Bach: An Eternal Golden Braid — Douglas Hofstadter; Theoretical Neuroscience — Peter Dayan & L. F. Abbott
Research interests
Areas I want to understand more deeply and contribute to over time.
- 01Neuroscience
How the brain learns so much from so little — and what that can teach us about building artificial intelligence. - 02AI Ethics
What it takes to make a system this powerful safe — precautions built in from the first design choice, not bolted on after. - 03AGI
What it will take to close the gap between narrow models and broadly capable intelligence.
Current question
Does regularization make small RNNs more robust to noise, or did the regularized network just start from a lower baseline loss? Retraining both from different random starting points, with matched training, is the next test.
About
I'm Ethan Ganz, a student working out how we actually get to artificial general intelligence — and what it'll mean once we do. My bet is that the answer runs through the brain: how it learns so much from so little, on about the power of a dim light bulb. So I study neuroscience, record what I think, and build small things to test it. I take the scaling camp seriously too — I just don't think anyone's settled it yet. Off the page, I'm on Math Team, TSA, and Science Olympiad, and I play tennis.