1. Agent swarms and the new model economics — cursor.com

  • Why read: Shows how Cursor uses multi-agent swarms to handle complex coding tasks (like building SQLite from scratch) by separating planning from execution.
  • Summary: Cursor grouped models into hierarchical swarms to handle larger, more complex tasks. They use smarter models for planning and cheaper, faster models for execution. This setup keeps agents on track: planners avoid getting stuck in the weeds, and workers only see the immediate task. To support hundreds of agents working at once, Cursor built a version control system that handles 1,000 commits per second. This approach cuts inference costs while keeping output quality high for long-running tasks.
  • Read more

2. How to Build a Self-Improving Outbound System on Codex — X (formerly Twitter)

  • Why read: A step-by-step guide to building a sales system that automatically improves its outreach based on real-world results.
  • Summary: Modeled on AI training loops, this system uses an agent to review past sales outreach and suggest tweaks to the playbook. The setup moves human judgment into configuration files and logs market outcomes as structured memory. An AI reads this memory, compares it to the current configuration, and opens a pull request with a specific change to improve messaging. Humans review the PR and check the evidence before merging. The result is a self-improving loop based on actual reply rates, without giving the AI direct access to the CRM.
  • Read more

3. Turn a million job posts into buying signals — Substack

  • Why read: Shows how to use embeddings to find high-intent sales leads hidden in unstructured public data.
  • Summary: Job postings are public declarations of problems a company will pay to solve. By running a million job postings through embeddings, sales teams can rank target markets based on specific pain points and missing capabilities. Instead of relying on keyword searches, this method evaluates the meaning behind the text to see if a company has the exact problem you solve. The system pulls direct quotes from the postings, giving sales reps concrete context for their cold outreach. It points to a shift toward "go-to-market engineering" where teams mine outbound signals at scale.
  • Read more

4. Language model harnesses are compositional generalizers — alexzhang13.github.io

  • Why read: Argues that software wrappers (harnesses) can force models to generalize, rather than relying on the models themselves to do it natively.
  • Summary: Large models still struggle to solve new problems by combining familiar concepts, which usually requires massive post-training datasets to fix. A Recursive Language Model (RLM) harness sidesteps this by breaking complex problems into simple, familiar sub-tasks. The underlying model only sees prompts that match its training data. Tasks of different lengths or domains end up looking structurally identical to the network. As a result, RLMs trained on short problems can handle tasks up to 32 times longer. For engineers, this means a well-designed harness can replace brute-force data collection and long training runs.
  • Read more

5. Who’s Afraid of Chinese Models? (Stratechery Article 7-20-2026) — Stratechery by Ben Thompson

  • Why read: Examines the economic fallout of capable open-weight models, arguing that AI inference brings real marginal costs back to the software industry.
  • Summary: The arrival of competitive open-weight models like Kimi K3 signals a shift in tech economics: marginal costs matter again. Downloading weights is free, but the cost to run inference scales directly with usage. Nvidia treats this setup like an ecosystem of "token factories," where strategy is driven by token cost and generation speed. As open models catch up to the frontier, they strip closed-model providers of their pricing power. This changes the software business entirely, rewarding efficient distribution and cheap execution over proprietary intelligence.
  • Read more

6. Open Models Tack Toward the Frontier — tomtunguz.com

  • Why read: Follows the race between open and closed AI models, showing how fast commoditization is driving down inference costs.
  • Summary: A clear cycle has emerged: closed models jump ahead in performance, and open models quickly replicate and commoditize those gains. New massive open-weight models from Moonshot and Alibaba show the open-source community matching frontier capabilities in record time. While open models aren't leading the pack, they have crushed pricing power—the cheapest open models run at a 90% discount to frontier options. This competition forces leading labs to keep innovating and cutting prices. For builders, this ensures AI infrastructure stays cheap enough to integrate widely.
  • Read more

7. Reliable unreliability — davidgasquez.com

  • Why read: Argues for a mindset shift in agent engineering: design systems to be reliable even when the underlying LLMs are unpredictable.
  • Summary: LLMs can be fluent, useful, and entirely wrong all at once. Instead of trusting a single large model to execute flawlessly, engineers need to build modular systems. This means narrowing responsibilities, breaking tasks into simple comparisons, and having multiple agents vote on or cross-check answers. By building explicit recovery paths and keeping individual model tasks small, you can achieve system-level reliability. It's similar to early computing, where engineers had to build stable systems on top of hardware that constantly failed.
  • Read more

8. WHOEVER MAKES THE MOST TOKENS, WINS — X (formerly Twitter)

  • Why read: Details the massive scale and operational intensity needed to win in AI inference hardware.
  • Summary: Etched expects AI inference to become the world's largest market, so they are vertically integrating everything from chip design to factory production. Their entire strategy focuses on bringing token capacity online as fast as possible. They run parallel timelines to beat competitors, shipping empty server racks and running models on FPGAs before their final silicon even exists. They push performance through technical workarounds, like running chips at lower voltages to avoid thermal limits and building custom network interconnects. The takeaway for operators: treat constraints as engineering problems and optimize entirely for speed.
  • Read more

9. How to become an AI native company: — X (formerly Twitter)

  • Why read: A six-step architectural roadmap for turning a traditional business into an AI-native organization.
  • Summary: Moving to an AI-native setup means abandoning scattered tools for a unified architecture. It starts with a single API gateway connecting AI agents to internal systems and data. Next, you build a "company brain" that combines static rules (guidelines) with live context (Slack messages, project updates). A central harness then tells the AI how to safely interact with these systems. Finally, a routing layer directs tasks to the cheapest capable models, preventing vendor lock-in and letting autonomous agents work reliably.
  • Read more

10. Maybe Intelligence Ain't All That — X (formerly Twitter)

  • Why read: Argues against the idea of a sudden AI takeoff, pointing out that physical testing and real-world friction bottleneck progress more than a lack of intelligence.
  • Summary: Cheap, widespread reasoning hasn't triggered a technological singularity. Intelligence works well in rules-based domains like law or coding, where it can easily extrapolate from known facts. But most hard problems live in messy physical and human systems with no cognitive shortcuts. An AI can design a mathematically perfect turbine blade, but you still have to manufacture and test it. AI will commoditize cognitive work, but the overall speed of innovation will still be capped by how fast we can run physical experiments.
  • Read more

11. 🎙️ How I AI: How the founder of Morning Brew built a Claude content machine that never runs out of ideas — Substack

  • Why read: Breaks down an automated content system that uses AI for interviewing, drafting, and editing while keeping the author's exact voice.
  • Summary: Alex Lieberman built a content engine that avoids generic AI writing by anchoring the output to his own thoughts. The system scans his weekly messages to find interesting angles, then uses AI personas to interview him via voice-to-text. He turned his personal writing style into a Markdown file that serves as a strict guide for drafting. To improve over time, the system compares its early drafts to his final published posts and updates its instructions based on the differences. It shows that good AI content depends on strong inputs and tight feedback loops, rather than just raw model capability.
  • Read more

12. My biggest takeaways from @Netflix's Chief Product and Technology Officer... — X (formerly Twitter)

  • Why read: Explains why generalists and systems thinkers are the most valuable operators in an AI-heavy company.
  • Summary: With AI speeding up output, the most effective employees are the ones building systems that span the whole business. This means stepping back from individual tasks to design platforms that let non-experts do complex work. Deep, narrow specialization matters less than the ability to learn fast and cross functional boundaries. Netflix expects AI fluency across the board, valuing curiosity over rigid technical checklists. This approach only works in high-trust, transparent cultures where people have the agency to ship their own work.
  • Read more

13. The Next Evolution of Cyber Deception in SecOps — Substack

  • Why read: Looks at how cybersecurity is moving from passive monitoring to actively misleading attackers.
  • Summary: AI speeds up cyberattacks and enables autonomous hacking, making human-speed defense obsolete. A new approach called "agentic deception" aims to mess with an attacker's reconnaissance and decision-making. Instead of waiting for a breach, these systems deploy active decoys to confuse attackers, slow them down, and force errors. It shifts the goal from just watching the network to actively altering what the attacker sees. Security teams are starting to treat deception as a core piece of their architecture rather than a side project.
  • Read more

14. Frontier AI Vision | Composable and distributed systems study group — Substack

  • Why read: A critique of the gap between the grand rhetoric of AI labs and the ground-level engineering bottlenecks of deploying models.
  • Summary: AI discourse focuses heavily on abstract governance and AGI, ignoring the hardware and integration limits that actually constrain the technology today. Critics argue that pushing for sweeping regulations is a distraction from the boring but necessary work of standardizing hardware (like chiplet interconnects). Despite being useful, current models face hard limits on context size, cost, and state tracking. Treating these models as nascent human minds rather than specialized computing components distorts both expectations and regulatory priorities.
  • Read more

15. The Prosperous Society — Mobile Dev Memo by Eric Seufert

  • Why read: Explains how AI moves the main bottleneck of the economy from production to distribution.
  • Summary: AI is an economic shift that drives production costs to zero, flooding the market with new goods and services. The new constraint on growth is distribution: matching that massive supply with the right demand. Ad networks and personalized coordination layers will serve as the economy's nervous system, helping niche products find specific buyers. This expands consumer choice and turns commerce into a deeper form of personal expression. If AI distribution networks can reliably match abundant supply with individual demand, it sets the stage for massive economic growth.
  • Read more

Themes from yesterday

  • Architecture beats raw scale: Builders are routing around the unreliability of large models by relying on multi-agent swarms, software harnesses, and modular workflows.
  • Reasoning gets commoditized: Open-weight models are driving inference costs down. As intelligence gets cheap, physical testing and real-world friction emerge as the actual bottlenecks.
  • Automating go-to-market: Sales teams are building self-improving loops to mine public data (like job boards) and run outbound campaigns at scale.
  • Hardware and distribution form the new moats: As model performance converges, competition shifts to who can build the fastest inference hardware and the most efficient distribution networks.