The Signal
@ylecun is now explicitly rejecting scaling as the path forward, pivoting the entire research conversation toward physical task learning and dynamics modeling. His repeated refrain—"LLMs can write essays but not clean bedrooms"—isn't rhetoric; it's a declaration that language-only architectures have exhausted their utility curve. The subtext: frontier labs know this too. When @bindureddy flags "a small continually learning model may be imminent" and @emostaque signals "strange and dangerous things are happening," they're not talking about essay quality—they're observing emergent behaviors in systems trained on dynamics, world models, and embodied learning. The competitive edge has shifted from parameter count to embodied reasoning.
IMPORTANT
The LLM commodity tier is now structurally complete; the next $500B in compute spend goes to robotics, world models, and continuous learning systems.
What's Moving
- World models become the research frontier — @ylecun's emphasis on "understanding dynamics of a system in order to control it" (not video generation) signals the field is abandoning pure language scaling. This tracks with @drjimfan's GEN-1.5 analysis: symmetry in human-collected data and recovery behaviors are the real training signals, not LLM-style next-token prediction. (via @ylecun, @drjimfan)
- Embodied learning infrastructure matures fast — @drjimfan's T-Rex (tactile-reactive manipulation) and UMI methodology show robotics now has the data collection and training recipes to compete with LLM scaling. The move from teleop to direct human gripper data (UMI) eliminates the "physical intuition bleed" that plagued prior systems. This is the first credible path to sub-human task learning at scale. (via @drjimfan)
- Small models commoditize, freeing frontier compute for embodied tasks — @bindureddy's prediction (95% of tasks via 250B open-source models in 6 months) isn't just about inference economics—it's about attention reallocation. If frontier labs concede the language tier, they can redirect GPU/TPU capacity toward robotics training, world model pretraining, and continuous-learning agents. (via @bindureddy)
- Continuous learning becomes the moat, not static weights — @bindureddy flags "a small continually learning model may be imminent"—the inverse of today's frozen-weight paradigm. This tracks @emostaque's "strange and dangerous things" observation: systems that adapt in-context without retraining expose new failure modes. (via @bindureddy, @emostaque)
Crosscurrents
- @ylecun vs. the singularity crowd — His repeated dismissal of "SV cultists" and "AGI-is-near narratives" signals a real schism. He's arguing the hard part isn't reasoning—it's control in real-world systems with friction and inertia. This directly challenges the "scaling = AGI" thesis that still dominates OpenAI/Anthropic positioning.
Tradecraft
BULL
Robotics/embodied AI teams now have credible infrastructure (UMI, T-Rex, GEN-1.5 baselines). Hiring and compute budgets should rotate there over next 6 months.
WATCH
When the first frontier lab publicly pivots a major team from LLM scaling to world models—that's the market signal that capability ceilings are real.
Desk Notes
- @ylecun — Systematic pivot away from scaling narratives; now openly mocking "singularitarians" while building the case for embodied AI as the actual frontier
- @drjimfan — Deep technical credibility on robotics data collection (UMI, tactile); GEN-1.5 analysis reveals why symmetric/recovery patterns in human data outperform teleop
- @bindureddy — Tracking both the LLM commodity collapse (95% of tasks in 6mo) and the emergence of continually-learning models; sees the frontier shifting in real-time