Nathan Labenz (AI:AM host) predicts a near-term acceleration phase where AI systems will optimize their own training infrastructure - essentially harvesting low-hanging fruit in the software stack itself. This could mean models auto-tuning hyperparameters, rewriting data pipelines, or optimizing compute allocation in ways humans haven't explored.
His endpoint estimate: AI doing R&D at weekly superhuman cadence. Not sci-fi AGI, but systems that can iterate on research problems faster than human teams working full-time. The exponential won't hold forever - but the plateau might arrive after most human contributors are already obsolete in the loop.
TLDR: The curve flattens eventually, but potentially after you're already out of the game. The race isn't to AGI - it's to staying relevant during the runup.
His endpoint estimate: AI doing R&D at weekly superhuman cadence. Not sci-fi AGI, but systems that can iterate on research problems faster than human teams working full-time. The exponential won't hold forever - but the plateau might arrive after most human contributors are already obsolete in the loop.
TLDR: The curve flattens eventually, but potentially after you're already out of the game. The race isn't to AGI - it's to staying relevant during the runup.