1. The Month China Closed the AI Stack — Tech Buzz China

  • Why read: Tracks how Chinese tech companies closed the AI performance gap using domestic hardware.
  • Summary: Chinese companies like Meituan are training trillion-parameter models strictly on domestic chips, bypassing US export controls. New models, such as Z.ai's GLM-5.2, are matching global benchmark leaders. This points to a self-sufficient Chinese AI ecosystem from silicon to frontier models. At these firms, AI-generated code is the new default, transitioning engineers into supervisory roles. Western labs and hardware makers no longer hold a monopoly on top-tier AI capability.
  • Read more

2. AI’s $1.5T Question — substack.com

  • Why read: Updates the math on the revenue required to justify current AI infrastructure spending.
  • Summary: The AI industry needs to hit $1.5 trillion in revenue by 2026 to justify current data center capital expenditures. This spending surge stems from the compute demands of long-horizon AI agents. While software monetization is catching up, with Anthropic reportedly hitting $60B in ARR, the math remains tight. Rising memory and materials costs mean standard GPU-based estimates likely undercount the true infrastructure burden.
  • Read more

3. The AI cost crisis is entirely self-inflicted (and Fable 5 just made it worse) — beehiiv.com

  • Why read: Explains why internal AI usage bills are skyrocketing and how companies can control them.
  • Summary: Companies scaling internal AI use are hitting unexpected cost spikes, sometimes jumping from $400k to $1.4M instantly. This usually happens when they outgrow subsidized API tiers or default to expensive new models like Fable 5 for basic tasks. The root cause is rolling out AI access without financial guardrails. Teams need to monitor usage and set cheaper open-weight models as defaults to protect runway.
  • Read more

4. 🤖Excited to share a new working paper🤖 — X (formerly Twitter)

  • Why read: Argues that managerial delegation limits AI productivity more than the technology itself.
  • Summary: A study of 200 startup founders found AI impact ranges from zero to a 3x output boost. The defining variable is organizational design. Founders who restructure workflows around AI see 50%+ productivity gains. Those who just hand employees AI access often see process bloat instead. Realizing value requires managers to map AI directly into production systems and change how they delegate tasks. AI adoption is an organizational challenge.
  • Read more

5. The Incognito Test: Why Most AI Work Reveals Nothing About the Person — Chief of Staff Network

  • Why read: Challenges managers to reject generic AI output and demand human insight.
  • Summary: Deep AI integration across business teams has created a new problem: high output volume with low judgment. Employees are passing generic AI drafts up the chain. The "incognito test" is simple: if a manager can generate the exact same text in five minutes, the employee added no value. To succeed in the next phase of AI work, employees must edit drafts to include their specific context, strategy, and opinions. Teams should be judged on business impact, not generation speed.
  • Read more

6. The Soul of Agentic Commerce — X (formerly Twitter)

  • Why read: Suggests the agent economy will stall until AI systems have their own autonomous goals.
  • Summary: The supply side of the machine-payable economy is ready, full of micro-services agents can buy. But demand is constrained because humans still have to prompt every action. To scale agentic commerce, AI systems need to originate their own goals. The author proposes "soul files"—documents defining an agent's identity, principles, and operating constraints. Giving agents intrinsic objectives will unlock high-volume, machine-to-machine commerce.
  • Read more

7. Consumer agents are the next big thing — X (formerly Twitter)

  • Why read: Explains how consumer AI will move from software subscriptions to capturing real-world transaction spend.
  • Summary: Consumer AI apps have a revenue ceiling because they compete for digital subscription budgets. Today's apps are mostly thin chat interfaces that leave humans doing the actual workflow. The shift is toward consumer agents that execute end-to-end tasks like booking travel or video editing. By completing the transaction, agents can tap into much larger offline budgets in sectors like education and healthcare. Human-in-the-loop fallbacks can bridge the gap where full automation currently fails.
  • Read more

8. A Taxonomy of Self-evolving Agents — X (formerly Twitter)

  • Why read: Provides a framework for understanding the landscape of self-improving AI systems.
  • Summary: Self-evolving agents break down into three components: models, harnesses, and artifacts. Right now, progress comes mostly from artifact iterative optimization. This involves agents using feedback loops to repeatedly improve an output, like a codebase or a paper. In this setup, the LLM is both the search operator and the optimizer. This taxonomy helps builders separate basic feedback loops from genuine recursive self-improvement, showing where architectural breakthroughs are needed next.
  • Read more

9. Distillation in 2026 (so far): which frontier models use it and how — X (formerly Twitter)

  • Why read: Explains how frontier labs use multi-teacher distillation to train 2026 models.
  • Summary: Model distillation is no longer just about making small models mimic big ones. Training one general model via reinforcement learning often degrades specific capabilities. Labs now train specialized RL experts for domains like math and coding. These experts, which can be the same size as the student, provide token-level feedback while the student model generates text. This process is faster and cheaper than standard RL, highlighting a shift toward synthetic feedback over brute-force scaling.
  • Read more

10. Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO — Substack

  • Why read: Argues cloud infrastructure must be redesigned to serve autonomous agents rather than human developers.
  • Summary: Current cloud infrastructure assumes human operators who can read dashboards and debug failures. AI agents cannot do this. They require an "Agent Experience": programmatic, isolated sandboxes where they can write code, execute it, and iterate on errors autonomously. Future AI cloud computing will center on elastic inference and bursty GPU workloads. Builders need to design infrastructure that agents can operate entirely on their own.
  • Read more

11. This is exactly what we enable our customers to do — X (formerly Twitter)

  • Why read: Details Uber's playbook for building internal AI agents for business functions.
  • Summary: Uber drove internal AI adoption by embedding engineers directly with finance, legal, and HR teams. Instead of whiteboarding processes, these pods shadow employees. Within two weeks, they build and ship agents that turn hours of manual work into minutes. The highest ROI comes from finding one employee's highly optimized, repetitive workflow and automating it across the company.
  • Read more

12. Some new agentic patterns — primeradiant.com

  • Why read: Shows how Prime Radiant deployed a fleet of specialized AI agents in their corporate Slack.
  • Summary: Prime Radiant runs their operations using a suite of non-coding AI agents for tasks like issue tracking, research, and executive assistance. The agents function as teammates: they triage information, update wikis, and proactively suggest actions. They ensure security by running subagents in isolated containers with strict credential management. The agents even work with human engineers to build their own internal tools. Treating agents as proactive colleagues yields far better results than treating them as reactive chatbots.
  • Read more

13. 🚨 OpenAI is launching GPT-5.6 Sol on Cerebras at up... — X (formerly Twitter)

  • Why read: Breaks down the wafer-scale hardware architecture enabling massive inference speeds.
  • Summary: OpenAI is reportedly running GPT-5.6 Sol on Cerebras hardware, hitting 750 tokens per second. Cerebras uses a Wafer-Scale Engine with 900,000 cores on one chip, bypassing standard GPU shared memory and threads. It relies on a dataflow model where data travels across the silicon, triggering tasks as it arrives. This provides massive on-chip memory with single-cycle access. The resulting speedup shows why specialized hardware is necessary to break current inference bottlenecks.
  • Read more

14. Rewriting Bun in Rust — Bun

  • Why read: A case study on using AI to port a complex codebase from Zig to Rust.
  • Summary: The Bun team used Claude Fable 5 to rewrite their JavaScript runtime from Zig to Rust. The original Zig implementation suffered from memory leaks and use-after-free crashes. The AI-assisted Rust port fixed these stability issues and unexpectedly shrank the final binary by 20%. This proves LLMs can successfully accelerate major, complex architectural migrations.
  • Read more

15. The bull case for AI and jobs — X (formerly Twitter)

  • Why read: An economic argument for why AI will increase the total demand for human labor.
  • Summary: The fear that AI will destroy jobs relies on a fixed-demand fallacy. Demand is elastic. By lowering the cost of cognition, AI makes previously impossible products and business models feasible. Automation is just the first-order effect. The second-order effect is market expansion, which creates new categories of work to manage these capabilities. Intelligence is currently rationed because it is expensive. Making it cheap will explode its use cases, requiring new forms of human oversight.
  • Read more

Themes from yesterday

  • Agentic Ecosystems: The shift from chat interfaces to autonomous agents is forcing changes in cloud infrastructure and corporate deployment strategies.
  • Hardware and Cost Reality: The industry faces $1.5T CapEx requirements, while dealing with Chinese hardware independence, Cerebras chip architectures, and soaring enterprise API bills.
  • The Human Element: AI output only has value when paired with human judgment, deliberate managerial delegation, and new forms of oversight.