The Signal — Enterprise builders are shipping with Claude Code and Anthropic agents at scale (Salesforce betting $300M annually). Simultaneously, the foundation model war is over: winner determined by cost-per-task, not raw capability. Open-weight models now competitive with proprietary. LeCun is right: continuous, noisy data remains unsolved.
Consensus: Bullish (agentic layer) / Bearish (moat erosion) | Conviction: High
What's Moving
- Agentic protocols (AG-UI, MCP) — Real-time tool orchestration with UI sync now table stakes; threading support emerging. Practical layer where enterprises make decisions. (via @svpino)
- Open-weight catching proprietary — MiniMax-M2.7 (230B, open) benchmarks at Opus 4.6/GPT-5.4 parity while costing 5% to run. SambaNova inference at 440 tokens/sec. Game-changing cost structure. (via @svpino)
- Flash models + commodity inference — Google 3.2 + rumored GPT 5.6 launch signals race bottoms-out at "cheap enough to automate everything." Efficiency, not intelligence, determines winner. (via @bindureddy)
- Enterprise coding agent adoption velocity — Non-chronically-online senior engineers impressed by Claude Code state-of-art. Salesforce deploying agents into service, support, distribution simultaneously. Adoption ahead of narrative. (via @svpino, All-In)
- LLM ceiling on continuous data — YLeCun unambiguous: LLMs "totally suck" at high-dimensional noisy data (robotics, real-time control, sensors). Humanoids / embodied AI require different substrate. (via @ylecun)
Blind Spot — The agentic win is masking a harder truth: foundation models may have hit a capability plateau. Consensus celebrates "GPT 5.6 incoming" but the real competition is efficiency ratios, not new capabilities. Open-weight models eroding proprietary pricing power faster than new features can justify it. India's 1B+ ChatGPT images signals geographic arbitrage—scale in emerging markets, not capability edge. And nobody is solving the LLM failure mode on continuous, messy, real-world data. That's where robotics and embodied AI live.
One Actionable Idea — Track which enterprise teams are mixing agentic layers (Claude Code + routing to cheaper open-weight models for sub-tasks). That's the emerging architecture. Proprietary-only bets are losing.
Sources: @svpino (agentic adoption, open-weight benchmarks), @bindureddy (efficiency wars, flash models), @ylecun (continuous data weakness), All-In (enterprise Anthropic deployment), @sama (India scale signal)