1. Your context is the bottleneck, not your model — Substack

  • Why read: See how managing your team's shared context gets better results out of AI.
  • Summary: An AI model's output depends entirely on what you feed it. Without good context, a capable model still produces bad work. Building a "Context OS" lets teams capture undocumented knowledge and prevent information silos. You can start with standard file directories and basic rules for agents. This structure turns isolated tasks into a system that gets smarter over time, keeping operations organized as work gets done.
  • Read more

2. React for Agents: Astro Creator Brings Hooks to his Meta-Harness, Flue — Latent.Space

  • Why read: How React-style hooks are making AI agents easier to build and update on the fly.
  • Summary: Flue 2 adds "Agent Hooks," letting AI manage its own state and swap tools while running. Instead of hardcoding behavior, this modular setup lets bots adjust mid-conversation. Flue runs agents inside a "harness" that handles context, replacing opaque file-based setups with standard web development patterns. This makes it easier to build adaptable, headless bots for support and triage.
  • Read more

3. The Unfalsifiable Economy — Investing 101

  • Why read: Why measuring business success by subjective metrics damages accountability.
  • Summary: Companies used to make physical products that either worked or didn't. Now, businesses chase fuzzy metrics like engagement and brand lift. This shift has created an economy where failure is easy to hide. When outcomes are subjective, there is no penalty for being wrong, which slows actual innovation. Organizations need to return to goals that can be proven false to stay disciplined.
  • Read more

4. What’s 🔥 in AI/Infra/VC #511 — What's Hot 🔥 in AI/Infra/VC

  • Why read: A look at the barbell market for AI venture funding and how to survive the Series A crunch.
  • Summary: VCs are funding massive seed rounds and late-stage winners, leaving mid-stage AI startups struggling. Application startups need steep revenue curves, sometimes $10M+ just to hit Series B. Infrastructure startups can raise on vision but must show they can reach massive scale. Companies stuck in the middle need to focus on reaching cash flow breakeven to avoid relying on a tough funding market.
  • Read more

5. Klaviyo's CEO on Building at $1.5B With Agents — SaaStr

  • Why read: How Klaviyo is shifting from reporting dashboards to active, revenue-generating AI agents.
  • Summary: Klaviyo grew by tying email metrics directly to sales rather than opens or clicks. Now, they are taking the same practical approach to AI. The company is moving away from passive software and building autonomous agents that execute tasks. It shows how a large enterprise integrates working AI instead of shipping theoretical features. For B2B software, automated action is becoming the new baseline over simple data reporting.
  • Read more

6. Why SpaceX Acquired Cursor — Contrary Research

  • Why read: The reasoning behind SpaceX’s $60 billion purchase of Cursor.
  • Summary: SpaceX bought Cursor to pair its own compute power with the startup’s post-training methods. SpaceX had the GPUs but struggled to build top-tier models internally. Cursor knew how to train models but lacked the hardware to scale. The deal quickly led to Grok 4.6, advancing their coding capabilities. The acquisition shows that frontier AI requires both massive infrastructure and specialized talent.
  • Read more

7. The Channel Layering Playbook — Cannonball GTM

  • Why read: How to replace volume-based cold outreach with signal-based selling.
  • Summary: Blindly calling everyone on an email list burns leads and hurts your brand. Sales teams get better results by waiting for intent signals, like website visits. A quiet prospect browsing your pricing page is often a better lead than someone who politely replies to an email without checking your site. Running targeted ads alongside these behavioral signals increases meeting book rates. Focusing on buyer behavior improves conversion without exhausting your prospect list.
  • Read more

8. Tyler Cowen interviews Daron Acemoglu on liberalism and economic growth — The Diff

  • Why read: A debate on how automation affects society and why we have to manage its fallout.
  • Summary: Automation usually increases overall wealth but can wipe out specific jobs and industries. Acemoglu argues that ignoring these stranded groups leads to political instability. If displaced workers are left without a safety net, society will push back against technological progress. The interview outlines the political friction we can expect as AI deployment scales.
  • Read more

9. Training a robot to play the guqin — The Diff

  • Why read: How researchers are using robotics to record complex, undocumented human skills.
  • Summary: A team trained a robot to play the guqin, a Chinese instrument usually taught person-to-person. Standard sheet music doesn't capture its timing or nuance, making the art hard to record. By teaching a machine the physical movements, researchers found a way to digitize and preserve skills that normally rely on human memory. It points to a use for robotics outside of basic physical labor.
  • Read more

10. Abraham Thomas on being the correct amount of paranoid when using AI for financial tasks — The Diff

  • Why read: How to force AI to be accurate enough for financial analysis.
  • Summary: LLMs are good at finding patterns across data but bad at being exact. This makes them dangerous for finance, where a small mistake ruins the analysis. To get useful answers, you have to strictly format your prompts and force the model to show its work. You essentially have to bully the AI into thoroughness. Using AI in finance requires a baseline level of suspicion to catch errors.
  • Read more

11. Apple Is the King of AI and Nobody Knows It — Substack

  • Why read: An argument that Apple's on-device AI strategy beats NVIDIA's compute dominance.
  • Summary: While the market focuses on NVIDIA and large models, this piece argues Apple has the actual advantage. Apple bakes AI directly into hardware that people already use every day. By prioritizing local, practical features over massive cloud models, Apple owns the interface where the value is delivered. The author suggests this lock on the consumer ecosystem matters more in the long run than selling raw compute.
  • Read more

12. Why compute might get 10x+ more expensive in coming years — Dwarkesh Podcast

  • Why read: The math on why GPU prices will surge if AI reaches human-level coding.
  • Summary: If an H100 GPU can replace a software engineer, its market value will match the engineer's salary. At standard rates for good developers, a single H100 would be worth $250,000 a year, about 15 times what it rents for now. As models hit human parity, hardware pricing will re-anchor to the cost of the labor it replaces. Companies building on cheap compute need to plan for these prices to spike.
  • Read more

13. Economists Reconsider Industrial Policy — workfutures.io

  • Why read: Why mainstream economists are changing their minds about government intervention in markets.
  • Summary: Economists have traditionally opposed industrial policy and subsidies. But recent data shows that funding specific sectors can strengthen the economy over time. Subsidizing upstream suppliers, rather than consumer products, lowers the risk of government waste and secures supply chains. This research provides the groundwork for recent legislation aimed at domestic tech and manufacturing. Expect more state involvement in strategic industries.
  • Read more

14. A homogeneous internet for all — The Substack Post

  • Why read: How AI tools and templates are making the internet look exactly the same.
  • Summary: With AI generators and templates, anyone can produce polished graphics and video. The result is a flood of competent but identical content. Making things look professional is now the baseline, meaning polish alone no longer captures attention. To stand out, creators have to stop relying on defaults and develop a distinct style that software can't immediately copy.
  • Read more

15. Underwriting GPUs — Contrary Research

  • Why read: How Wall Street is turning AI compute into an asset class like real estate.
  • Summary: NVIDIA is working with firms like BlackRock and Goldman Sachs to raise $500 billion for AI infrastructure. The market is starting to treat GPU clusters like power plants or office buildings: hard assets that generate yield. While NVIDIA backs some of the risk, the heavy lifting is moving to institutional finance. This shift means the AI buildout will rely on debt and project finance rather than just venture capital.
  • Read more

Themes from yesterday

  • Software is shifting from static reporting tools to autonomous agents that act on context.
  • The economics of AI are changing as infrastructure becomes a formal asset class and compute costs map to human labor.
  • Companies and creators need hard, falsifiable metrics to cut through the noise of AI-generated content and vanity data.