The Signal
Yann LeCun's 663-like post on July 4 landed a systematic critique that's gaining gravity: current LLMs solve discrete-symbol problems brilliantly but collapse on anything that requires real-world agency, continuous signals, or learning-by-doing. No level-5 autonomous vehicles. No domestic robots matching a ten-year-old's first-try competence. No agents smarter than a house cat at physical tasks. This isn't a technical quibble—it's a bet that the frontier isn't "bigger models" but "models that can predict consequences and plan." The signal: frontier capability theater (benchmark scores, token budgets, new model drops) obscures a hard constraint that no amount of scaling fixes.
IMPORTANT
The bottleneck isn't intelligence per token—it's the inability to learn from interaction with the physical world.
What's Moving
- Physical embodiment as the binding constraint — LeCun explicitly names the Moravec paradox (38 years old, still ignored). Discrete-symbol tasks max out; physical/continuous tasks plateau. This flips the entire "scale = capability" narrative. (via @ylecun)
- Agent-as-economic-actor model hits friction — @bindureddy's routing strategy (98 likes) shows Fable one-shotting complex tasks but sucking at chat; Gemini Flash strong on code, weak elsewhere. The implication: you can route around model weakness for text-native workflows. You can't route around missing embodied learning. (via @bindureddy)
- Vision gap stays a moat for closed-source — @bindureddy explicitly flagged GLM 5.2 has zero multimodal capability. Chinese open-source dominates on text/code; US closed models own perception. Agents can't be fully autonomous without both. (via @bindureddy)
- Meta's Watermelon drops as legacy on arrival — @bindureddy notes it's GPT 5.5 class when 5.6 launches next week (115 likes). A model generation behind before release. The velocity of frontier models makes anything not-leading-edge immediately obsolete. (via @bindureddy)
Crosscurrents
- Sam Altman's token-gifting narrative vs. LeCun's power-consolidation critique — Altman's July 5 post (4368 likes) equates two-word toddler speech to GPT 5.6 "discovering new math." LeCun's framing: both are symbol manipulation, neither approaches embodied cognition. The gap between these framings is where startup risk lives. (via @sama / @ylecun)
- Open-source velocity outpacing regulation, but vision remains the check — @emostaque's Meituan-LongCat MoE (10T tokens/month on OpenRouter) proves Chinese training efficiency scales without US GPUs. But it can't see. Regulation can't slow the march; physics constraints can. (via @emostaque)
Tradecraft
WATCH
July 8-14: GPT 5.6 Sol preview expansion. If it closes vision parity to Fable, the "US model unreliability" thesis collapses. If not, routing toward Opus/Fable for perception tasks hardens as structural.
WATCH
Q3 2026: Any Chinese model announcement claiming multimodal parity. The moment vision closes, embodied learning becomes the last defensible moat—and LeCun's critique becomes the business strategy.
Desk Notes
- @ylecun — The G in AGI is performative; real constraint is physical world prediction and planning, not token count.
- @bindureddy — Routing by intent now table stakes; single-model strategy is cost-destructive and capability-incomplete.
- @emostaque — Chinese training efficiency already won on text/code; vision gap is the remaining US leverage point.