The AI-at-work story has two chapters right now, and they're in tension. Chapter one: models are genuinely getting good enough to handle the complex, high-stakes work that was supposed to be human-only territory. Chapter two: the companies deploying them are discovering that "AI does the work" and "humans stop talking to each other" are the same sentence.
01
The AI productivity dream has a loneliness problem
Zara Zhang, a VC who covers AI, relayed an observation from a founder that's worth sitting with: this founder rolled out OpenAI's Codex Max across his entire team, and it worked. People got their work done. Meetings disappeared. Collaboration dropped to near zero. The team culture quietly fell apart. Zhang's framing is sharp: most enterprise AI usage is single-player, and nobody designed for what happens when every employee has their own private AI collaborator they prefer to their actual colleagues. ---
Why it matters: This is the canary in the coal mine for any company doing a broad AI rollout. The productivity numbers will look great in Q3. The attrition and culture surveys will look bad by Q1. If your team is measuring AI success by output per person and not by anything relational, you're optimizing for half the picture.
The open secret about which models power your AI agents
Shawn Wang, who runs Latent Space, made a pointed observation about the AI industry's quiet hypocrisy: most agent companies are quietly running Chinese models under the hood but won't say so publicly because they're pitching government and defense clients. He highlighted one lab, Cog, that did the unglamorous work of actually productionizing a Chinese model responsibly: they built multilingual propaganda and censorship evaluations, corrected for those tendencies in post-training, and got inference speed to 1,000 tokens per second at low cost. ---
Why it matters: If you're a company that bought an "American AI" agent product and assumed that meant American models all the way down, you may want to ask your vendor to be more specific. The supply chain opacity in AI is a real procurement risk, and most buyers aren't asking the right questions yet.
Box CEO sees the model quality leap clearly, calls out Grok 4.5
Box CEO Aaron Levie posted his read on the current model generation: the latest releases are crossing a meaningful threshold on complex knowledge work, specifically legal, healthcare, and professional services domains. He called out Grok 4.5 by name for its cost-to-performance ratio and predicted that better reasoning plus vertical training will unlock significantly more value from enterprise documents. ---
Why it matters: Levie has a direct financial stake in whether enterprise AI matures, so this isn't neutral commentary, but the underlying point is real. If you're a law firm or hospital system that evaluated AI assistants 18 months ago and found them too unreliable for serious work, the evaluation is worth redoing.
Guillermo Rauch posted a brief but enthusiastic note about his personal agent stack coming together, referencing a tool shared by developer Nishimiya that he plans to use for personal productivity agents. Light on specifics, but notable given yesterday's coverage of Vercel's Better Auth acquisition and Rauch's broader push toward a developer platform built for agentic apps. ---
Together AI launches reserved capacity for open models
Together AI is now offering Provisioned Throughput: reserved inference capacity for frontier open models including MiniMax M3 and GLM-5.2, with token-based pricing, a 99% uptime SLA, and pricing they claim is up to 90% cheaper than proprietary APIs. No GPU management required.
Why it matters: For developers already using Chinese open models (and per Shawn Wang's post above, there are more of them than admit it publicly), this removes the last friction point. Predictable capacity and cost on models that were previously a reliability gamble is a meaningful step toward legitimate enterprise use.