1. The Reverse Information Paradox — Satya Nadella

  • Why read: Explains why companies should protect their learning loops when adopting AI instead of handing their data exhaust to model providers.
  • Summary: The "Reverse Information Paradox" happens when buyers give away their proprietary knowledge (corrections, context, and prompts) to use the models they pay for. Without a hard trust boundary, economic value shifts to the owners of the learning infrastructure instead of the creators of the knowledge. To stop this, enterprises should build private evaluation systems and keep full ownership of their institutional memory. Every firm needs to control its learning environment so its knowledge compounds internally.
  • Read more

2. Where evals are going: agent-as-a-judge — Aparna Dhinakaran

  • Why read: Looks at the shift in evaluating AI systems, moving from grading static text to judging multi-step agent trajectories.
  • Summary: Traditional LLM-as-a-judge methods were built for prompt-response interactions and fall short for modern autonomous agents. Agents fail in sequences, where stuck loops, forgotten context, or broken tool calls hide behind a reasonable final answer. Assessing these systems requires "agent-as-a-judge" setups equipped to trace requirements and investigate intermediate steps. Recent research shows agent judges are faster, cheaper, and match or beat human experts in evaluating complex AI tasks.
  • Read more

3. The Great Divergence in Software Engineering — Paul Stack

  • Why read: Examines the widening gap between engineering teams adopting AI and those resisting it.
  • Summary: Using AI in software engineering represents a different way of working where automated pieces compound development speed. Some teams are shipping large platforms with minimal headcount, while others ban AI tools or stall in evaluation committees over unreliability fears. Successful teams treat AI unreliability as an engineering problem, building guardrails and conventions instead of abandoning the technology. Organizations built for the throughput of two years ago are structurally unequipped for this leap in productivity.
  • Read more

4. How to Think About Build vs. Buy in the AI Era — Brendan J Short

  • Why read: A framework for deciding which parts of your AI stack to outsource and which to own internally.
  • Summary: Advanced models have changed the "build vs. buy" math, making custom internal tools cheaper and faster to develop. Companies should buy core infrastructure like dialers, CRM plumbing, and data warehousing, where reliability outweighs differentiation. However, they should build and own their "intelligence layer," which includes ICP definitions, lead scoring logic, and deal strategy. Outsourcing this layer risks losing the workflows and organizational judgment that provide an actual competitive advantage.
  • Read more

5. AI Makes It Easier to Build the Wrong Product — Grey

  • Why read: A reminder that while AI lowers the cost of writing code, the cost of building the wrong thing remains high.
  • Summary: Because AI lets teams generate code quickly, many organizations revert to large-batch implementations and produce entire milestones before validating them with users. Faster code generation answers whether something can be built, but not if it should be. To avoid "decision debt," teams should use AI to deliver end-to-end vertical slices that test uncertainty. Keeping development batches small and functional offers cheap opportunities to adjust assumptions before committing to a product direction.
  • Read more

6. Self-Improving Agents, But Improving at What? — Josh Rosen

  • Why read: Looks at how AI agents are evolving from static tools into systems that refine their own operations and understanding.
  • Summary: Agent research is increasingly focusing on systems that learn and optimize post-deployment by adjusting their harnesses, prompts, and tool policies. The conventional loop records task trajectories, identifies failures, proposes component changes, and promotes updates that improve benchmark scores. But optimizing strictly for task completion limits the system's ability to transfer knowledge to new problems. The real frontier for self-improvement involves agents refining their core understanding of a domain, finding simpler decompositions, and extracting reusable first principles from their work.
  • Read more

7. AI’s Biggest Winners Have the Lowest Margins — Daniel Kornum

  • Why read: Shows why traditionally low-margin businesses stand to gain the highest financial upside from AI automation.
  • Summary: Labor-intensive companies like manufacturers and field-service operators have historically been stuck in low-margin environments with massive manual coordination costs. AI offers a way to reduce administrative burdens like scheduling, dispatching, and exception handling. In a business with a 3% profit margin, easing coordination costs by 10% can lead to a 20% increase in overall earnings. The companies that move first to implement AI behind the scenes will reset their cost positions and capture value before commoditization lowers prices.
  • Read more

8. I was wrong about MCPs — Adam

  • Why read: Argues that the Model Context Protocol (MCP) is better understood as a framework for generalized agent capabilities rather than a basic integration tool.
  • Summary: MCP is known for translating API endpoints into tools for LLMs, but its power extends beyond connecting agents to single services like GitHub or Resend. Generalized MCP servers grant agents the ability to search for capabilities and execute custom code environments, which increases reliability. Instead of risking compounded errors across a chain of non-deterministic LLM calls, an agent can write a deterministic script once and reuse it. This shifts MCP from a tool provider to a capability engine that changes how agents execute complex workflows.
  • Read more

9. Daniela Amodei: The Greatest Operator in AI — Arman Hezarkhani

  • Why read: A profile of Anthropic's President highlighting the role of operations and safety management in scaling AI companies.
  • Summary: Visionaries and researchers dominate the AI spotlight, but operators like Daniela Amodei build the structure required to scale these institutions. Drawing from a background in recruiting and risk at Stripe, Amodei runs most of Anthropic's workforce, allowing the organization to balance commercial scaling with safety protocols. Her operational strategy bets that safety and commerce are compatible, and the company that masters both will dominate the market. This highlights the importance of non-technical leadership in preventing a technological revolution from becoming an accident.
  • Read more

10. The Great Wave Has Arrived (from GLM CEO Jie Tang) — Bing Xu

  • Why read: A look into the operating philosophy of Zhipu, one of China's foundational AI model builders.
  • Summary: Zhipu's strategy relies on first principles and focus, avoiding fleeting industry trends in favor of long-term investments in artificial general intelligence. They view the current AI transformation as a technological shift that rewrites the ceiling of human and machine intelligence. The transition from perceptual intelligence to cognitive reasoning means any company that pushes this ceiling higher will redefine global industries. To succeed, AI organizations have to be willing to reset their successes to zero and commit to solving foundational challenges.
  • Read more

11. A single question to track progress from o3 to gpt-5.6 and beyond — Sebastien Bubeck

  • Why read: An inside look at how AI researchers use specific mathematical questions to measure real leaps in model reasoning.
  • Summary: To test the reasoning capabilities of frontier models, researchers use mathematically profound questions, such as calculating the path length of a gradient flow on a convex function. Early reasoning models like o3 were the first to understand the nuances of these questions, but models like GPT-5.6 now spend hours in deep thought to beat published state-of-the-art mathematical bounds. This tracking shows how AI is moving past pattern matching and discovering long-horizon solutions that previously required years of academic collaboration. It highlights AI's growing ability to untangle high-dimensional complexity.
  • Read more

12. AI Comes for the Fragrance Oligopoly: Alex Wiltschko (Osmo) x TWIML — Fireside Alpha

  • Why read: A case study on how AI-native startups use high data velocity to disrupt entrenched consumer oligopolies.
  • Summary: Osmo is challenging the flavor and fragrance oligopoly by relying on the speed of data generation rather than model architecture alone. While incumbents possess a hundred years of fragmented spreadsheet data, Osmo mathematically mapped the principal dimensions of odor and built infrastructure to continuously generate clean training data. This creates a moat where the attacker's data velocity outpaces the defenders' static accumulation. It serves as a blueprint for how AI can dismantle traditional industries by turning physical chemistry problems into digital mapping exercises.
  • Read more

13. Building Reliable Agentic AI Systems — Sarang Sanjay Kulkarni

  • Why read: A deep dive into how Bayer designed PRINCE, an Agentic RAG system for preclinical research.
  • Summary: Traditional keyword searches fail when navigating the fragmented data silos of preclinical drug discovery. To solve this, Bayer developed an agentic system that uses "context engineering" to control what information specialized models receive during research, reflection, and writing phases. They paired this with "harness engineering," implementing strict tool boundaries, reflection loops, observability, and human fallbacks to ensure accurate outputs. This approach shows how architecting the scaffolding around LLMs helps turn scattered enterprise data into trustworthy insights.
  • Read more

14. How To Actually Design With AI — Leon Lin

  • Why read: Practical advice on bridging the gap between AI execution and human product design.
  • Summary: Designing with AI still requires a human to provide the meaning, direction, and inspiration while the AI handles execution. AI understands design rules like typography and spacing, but lacks the taste needed to create something original and premium. Operators should define the emotion and purpose of the product, curate component inspirations from reference libraries, and then prompt the AI to build section by section. This modular workflow prevents generic AI output and ensures the final product reflects human judgment.
  • Read more

15. we just built an infinite context window for ai — dale

  • Why read: An engineering approach to solving the AI context limitation by mirroring how human peripheral attention works.
  • Summary: Increasing the size of an LLM's context window is inefficient, like trying to keep your entire peripheral vision in sharp focus. Instead of forcing a model to hold terabytes of records in its immediate attention, the Polygres team developed a dynamic context window that maintains a focused working set while querying a broader graph-based knowledge body. This allows the AI to notice relevant edge signals, pull them into focus, traverse connections, and move on without perfect recall of everything. The architecture shifts the bottleneck of AI capability from raw reasoning power to selective attention.
  • Read more

Themes from yesterday

  • Agent Evaluation & Self-Improvement: Since models now generate workflows instead of static text, evaluation is shifting toward "agent-as-a-judge" setups. Systems are also starting to refine their own core reasoning rather than just executing tasks.
  • Defensibility in the Intelligence Layer: The competitive moat is moving from generic infrastructure to owning private data exhaust, intelligence workflows, and the speed of generating proprietary data.
  • Rewriting Structural Economics: Low-margin sectors face upside potential as AI attacks coordination costs. Meanwhile, software teams adopting AI are widening the gap against those resisting it.