We accidentally solved robotics by watching 1M hours of YouTube
Copyright, scraping, and legality
- Several comments question whether scraping YouTube at this scale violates YouTube’s ToS and copyright; many think it probably does.
- Others argue ToS may be unenforceable if the scraper never agreed to them (citing hiQ v. LinkedIn and the fact that videos are accessible without login in some regions).
- There’s debate over whether YouTube, as a non‑exclusive licensee of user content, can legally restrict downstream scraping at all.
- A minority take the view that, regardless of legality, mass “pirated” training data is now de facto tolerated for big AI labs, effectively eroding copyright in practice.
Aaron Swartz, double standards, and the justice system
- A long subthread compares aggressive enforcement against Aaron Swartz for bulk academic scraping to the apparent impunity of large AI companies doing similar or worse at scale.
- One side characterizes Swartz as “hounded to death” by disproportionate prosecution for an arguably public‑interest act; the contrasting lack of criminal action against AI firms is cited as evidence of plutocratic double standards.
- Others push back: Swartz’s case involved physical network intrusion, not merely scraping; many defendants survive harsh prosecutions; his suicide is attributed primarily to mental illness and stress, not solely government action.
- There is broad agreement that the system is harsher on individuals than corporations, but disagreement on how much to change prosecution norms versus mental‑health support.
Hype vs. reality: “accidentally solved robotics”
- Multiple commenters reject the title as “extremely oversold”: current success rates on simple tasks and strong camera‑pose sensitivity are seen as very far from “solved robotics.”
- Critics note similar ideas and datasets (web video, affordances, world models) have existed for nearly a decade; this work is viewed as solid incremental progress, not a breakthrough.
- Vision-only models are widely viewed as insufficient for robust real-world manipulation without touch/force sensing, handling failure modes, and explicit causal or physics modeling.
General-purpose humanoids vs. specialized robotics
- Some argue humanoid or fully general robots will always be slower, more expensive, and less efficient than specialized machines for most industrial tasks.
- Others counter that the value of generalization is high: factories routinely plug humans into ad-hoc roles; a human-like, easily instructed robot could replace such flexible labor.
- There is tension between “job-shop” style, human-centric manufacturing and highly optimized, human-minimized, specialized automation.
Writing style, credibility, and prior work
- Many readers find the blog post nearly unreadable: meme-laden, lowercase, semi-ironic “Twitter/Discord” style, with vague claims, loose numbers, and shifting “we.”
- Several suspect LLM-assisted or LLM-authored text and note factual sloppiness and “history rewrites”; they recommend reading the underlying FAIR paper instead.
- Some point out the paper itself is more modest: web-scale video pretraining plus limited robot data yields decent zero-shot planning in constrained settings—interesting but not revolutionary.