Audio and video are AGI's blind spot: models can't verify what they hear or see
coldrolledpotmetal · reddit · 2026-10-22
A Reddit essay pushes back on "AGI has arrived" claims: frontier AI excels at static tasks (text, images, code) but fails at anything unfolding over time.
- Audio: lab effort has gone almost entirely into conversational speech, not general audio understanding. Following one voice in a noisy room or identifying when instruments enter breaks models in ways human hearing never would; speaker separation gets brittle with overlapping talk, and tools like demucs don't help because the bottleneck is the underlying models.
- Missing verification loop: coding agents improved fast because models can run and check output. With audio, a model can't listen to its own result and tell whether it sounds right — the feedback loop that drove agentic coding doesn't exist.
- Video: Gemini samples at 1 fps by default, effectively watching a slideshow; Google's agentic video mode with zoom-in is a workaround for not perceiving time continuously.
Author's conclusion: a system that can't tell if a song sounds right or follow a crowded-room conversation isn't general yet, however solvable these gaps may be.
More from AGI Musings
- ML pioneer Dietterich: OpenAI's agents ran for months with essentially no supervision — tdietterich · 2026-10-22
- Altman hints OpenAI's internal models do math public models can't, flags RSI speed limits — VraserX · 2026-10-22
- New symmetric Venn diagrams with 17, 19 and 23 curves found via AI-assisted search, breaking 2014 record — pickover · 2026-10-22
- Ex-OpenAI researcher's hot take: NYC trails SF by 4-7 months in AI, London 12+ — jefrankle · 2026-10-22
- Debate on AI proof formalization: mass-formalize textbooks first, then let AI build on the repos — jacobaustin132 · 2026-10-22
- Formalization ≠ understanding: the core split in the AI proof debate — eigenstate · 2026-10-22