1. Agentic AI adoption is on fire at @Uber, and it's... — X (formerly Twitter)

  • Why read: How Uber used AI agents to overhaul company workflows beyond engineering.
  • Summary: Uber paired engineers with domain experts for two-week sprints to automate operations, finance, and legal workflows. By observing experts in action, the teams realized the biggest gains come from rebuilding entire processes instead of automating isolated tasks. They reduced capital allocation planning from 15 hours to 30 minutes and financial pacing from two days to 10 minutes. This method removes unnecessary approvals and legacy tools. It proves that finding the best AI use cases requires watching people work and building solutions with them.
  • Read more

2. Harness Engineering for Self-Improvement — lilianweng.github.io

  • Why read: A framework for how AI systems self-improve by optimizing their deployment environment instead of changing their weights.
  • Summary: A "harness" is the system around a base AI model that manages its execution, tools, memory, and evaluation. Lilian Weng argues that engineering this harness is necessary for AI to self-improve. This means moving past basic agent frameworks to persistent memory and reliable workflows. Good harnesses use loops where models plan, execute, test, and fix their actions, similar to software engineering. Standardizing these harnesses like operating systems is key to getting the most out of models on complex tasks. The deployment layer is becoming as important as the model itself for real-world performance.
  • Read more

3. After inference — Angular Ventures

  • Why read: An argument that AI's next hurdle is turning expensive real-world experience into learning instead of relying on synthetic data.
  • Summary: Recent AI breakthroughs relied on scaling cheap resources like internet data and compute. But physical experience is expensive to get. As companies deploy robots and agents, they face an open-world problem: routine successes teach little, while edge-case failures are valuable but hard to simulate. Simulation falls short because it cannot predict unknown failures or unmodeled physics. Founders need to build systems that feed deployment data straight back into model training. The edge will go to systems that can learn continuously from messy, expensive real-world data.
  • Read more

4. The Task Economy - Data will be the next $1 Trillion Category — X (formerly Twitter)

  • Why read: Introduces the "Task Economy" as the next phase for scaling AI, shifting focus from compute tokens to high-quality domain data.
  • Summary: Inference tokens are the popular metric for AI growth right now, but the real bottleneck is getting high-quality data. As internet data runs out, models need complex, domain-specific tasks for reinforcement learning. This "Task Economy" goes beyond basic data labeling. It requires expert feedback and specialized workflows to generate training signals. Owning or creating proprietary datasets and task environments will become highly valuable. Companies that help create, score, and group complex tasks will drive the next wave of model improvements.
  • Read more

5. How tech workers are feeling in 2026: a workforce splitting in two — lennysnewsletter.com

  • Why read: A survey showing how AI is splitting tech workers into two camps: those who feel empowered and those who feel overwhelmed.
  • Summary: A large survey shows a divide in the tech industry: half the workforce feels highly productive using AI, while the other half reports anxiety and declining optimism. Burnout is up 11 points from last year, and 40% worry about job stability. Employees are more worried about being forced to work at an unsustainable pace for the same pay than they are about losing their jobs to AI. Designers and researchers report the highest anxiety and are least likely to recommend their fields. Managers rolling out AI tools need to watch employee well-being and set realistic productivity expectations.
  • Read more

6. The Second Derivative: Why No One Understands the AI Boom — Groundbreaker

  • Why read: A financial warning comparing the AI boom to the 2008 mortgage crisis, highlighting the risk of slowing growth rates.
  • Summary: The AI boom relies on continuously accelerating growth, not steady revenue. Like the subprime mortgage crisis, the weak point is the "second derivative", the rate of growth acceleration. If growth stays positive but slows down, the financial models and debt obligations of AI companies could fall apart. The market often misses this, viewing defaults as external shocks rather than built-in flaws. Investors and operators need to watch growth rates and avoid heavily indebted models that require endless exponential expansion.
  • Read more

7. AI just filled it in. Here's what actually protects a business now. — beehiiv.com

  • Why read: A warning for product leaders that software features are no longer a moat in the age of rapid AI development.
  • Summary: Building a product with complex features used to be a strong moat. Now, AI lets competitors copy sophisticated engineering in weeks instead of years. Companies can no longer rely on product differentiation. The real moats today are defining a category, engaging a community, and running a tight operation. Product leaders need to focus on customer loyalty and owning the industry conversation to survive fast-moving AI clones.
  • Read more

8. How to Think About Build vs. Buy in the AI Era — Substack

  • Why read: A framework for deciding when to build AI software in-house versus buying off-the-shelf.
  • Summary: Better AI models have changed the "build vs. buy" math, making custom development faster and cheaper. You should still buy commodity infrastructure like dialers, CRM plumbing, and data warehouses. But companies are now building their own intelligence layers: business-specific logic for lead scoring, deal strategy, and ideal customer profiles. Building custom AI systems lets teams double their output by using their specific company context. Operators should build workflows where their proprietary data and judgment give them an edge.
  • Read more

9. The end of the working writer — substack.com

  • Why read: The economic reality for commercial writers as AI replaces routine content generation.
  • Summary: People view human expression as irreplaceable, but most commercial writing, like product copy, white papers, and SEO posts, is valued because it exists, not because it is art. In 2026, AI drafts this content rapidly, turning writers into editors. This threatens the entry-level freelance jobs that usually support early-career writers. If your writing serves purely informational or commercial ends, it will be automated. To survive, writers need to focus on deep creativity, criticism, and strategic editing.
  • Read more

10. Stop fixing your trial. Fix who's in it. — Substack

  • Why read: A reminder for product teams that bad conversion rates are usually an audience problem, not an onboarding problem.
  • Summary: When trial conversions drop, founders usually try to fix the onboarding flow. But the real issue is often that the signups were never buyers to begin with. They were students, researchers, or people in the wrong roles. Product polish will not convert someone without a budget. Product leaders should manually review signups and talk to users to separate real prospects from casual browsers. Fix your targeting before you spend time optimizing the trial experience.
  • Read more

11. RSI for AI: Truth, Fiction and Economics — X (formerly Twitter)

  • Why read: An economic look at Recursive Self-Improvement (RSI) in AI and its market impact.
  • Summary: Recursive self-improvement, where AI designs, builds, and deploys its own upgrades, could become reality by 2028. This shifts the competitive landscape, as successful RSI could create winner-take-most dynamics among top AI labs. Adversarial distillation and open-source models will determine who captures this economic value. Pessimists worry about job displacement and bottlenecks, while optimists think RSI productivity will overcome diminishing returns and create new sectors. Operators need to prepare for a world where AI accelerates its own research without waiting for humans.
  • Read more

12. [AINews] The Field Guide to Fable — Substack

  • Why read: Techniques for unhobbling advanced AI models to unlock their reasoning and coding potential.
  • Summary: Users often restrict highly capable models like Fable 5 with outdated prompts. This guide explains how to unhobble these models so they can handle complex tasks like HTML generation and design without micromanagement. Tactics include asking the model for a "blindspot pass," brainstorming different directions, and having the model interview you to clarify unknowns. Builders need to drop old assumptions and demand fast, cheap, high-quality results from new models.
  • Read more

13. Children of the Magenta Line — substack.com

  • Why read: An aviation analogy showing the danger of relying on AI tools instead of foundational learning.
  • Summary: Pilots who rely too much on autopilot lose their manual flying skills and understanding of the aircraft. The same applies to AI in education and knowledge work. Giving beginners free access to chatbots lets them skip the foundational learning required for critical thinking. Background knowledge is the scaffolding needed to evaluate claims and understand complex subjects. The danger is not automation, but dependency that leaves users helpless when AI fails. We need workflows and educational systems that use AI to support thinking, not replace it entirely.
  • Read more

14. Build the Ark Before the Storm — X (formerly Twitter)

  • Why read: A look at navigating the financial AI bubble and a slowly eroding cultural foundation.
  • Summary: The tech industry is in an AI financial bubble of high valuations and massive data-center spending. But the bigger danger is a society losing its shared purpose and relying entirely on capital. History shows financial crashes are devastating when societal cohesion is low, but strong cultures can rebuild. Leaders need to build resilient institutions rooted in values like craft and duty before the financial correction hits. Preparation means recognizing the social trends that will determine the severity of upcoming disruptions.
  • Read more

15. Anthropic Will Out-Earn Every Public Software Company Except Microsoft — SaaStr

  • Why read: Anthropic's financial trajectory shows how much enterprise value AI captures.
  • Summary: Anthropic is growing so fast it could pass the revenues of software giants like Salesforce and Adobe by year-end. This shows the massive enterprise demand for AI infrastructure and models. Anthropic's scale redefines expectations for B2B startup growth and highlights the value in the AI stack's foundational layer. Enterprise budgets are moving heavily toward core AI providers. The enterprise software market is being restructured by companies that deliver top-tier model intelligence.
  • Read more

Themes from yesterday

  • Agentic Workflows are the New Standard: Companies are moving from using AI for isolated tasks to deploying agents that rebuild entire workflows.
  • The Economic and Defensive Realities of AI: Product features are no longer moats, and the financial models driving the AI boom have hidden vulnerabilities.
  • The Polarization of the Tech Workforce: AI is dividing the workforce into those who feel empowered and those who are anxious about their pace of work and job security.
  • The Rise of the Task Economy: With internet data running out, the market is shifting toward complex, domain-specific tasks to train the next generation of models.