DeepSeek call argues continual learning is still the main unsolved AI problem
hrishioa · x · 2026-07-24
A discussion from a DeepSeek call argues that continual learning remains the core unsolved problem: pretraining solved learning at scale, but humans still win because they can keep learning from tiny amounts of new data.
The thread separates learning new information from recursive self-improvement (RSI), treating RSI as a weak form of persistent memory — improving the agent’s weights or harness to preserve gains. It also suggests there may already be a 10–20x application-layer gain available by moving work off the main agentic thread and letting a separate agent improve the system, with potentially even more upside.
The quoted line pushes a more deflationary timeline: first solve “learning to learn,” then reach a self-iterating singularity, and only later embodied intelligence that can operate in the physical world.
More from AGI Musings
- AMP founder says poor scheduling and chip stalls waste most AI compute — Kyrannio · 2026-07-24
- Biosecurity debate says bioweapon risk stays urgent as technology keeps improving — sebkrier · 2026-07-24
- ‘Web search plus prompting is AGI,’ the post jokes about model progress — willccbb · 2026-07-24
- The Temptation of AI: Developer Finds Manual Documentation Becoming a Struggle — sull · 2026-07-24
- Elon Musk Predicts AI Will Exceed Total Human Intelligence Within Five Years — Polymarket · 2026-07-24
- Dev Rant: If AI Coding Agents Mess Up, Why Are We Paying for the Tokens? — larumis · 2026-07-24