1. Anthropic’s Safety Superpower — Ben Thompson

  • Why read: What Anthropic's new Fable model and its sudden US government takedown mean for your AI strategy.
  • Summary: Thompson argues Anthropic's Fable 5 is a true generational leap, easily beating GPT 5.5 and Opus 4.8 at coding. But the US government forced its removal shortly after release due to cybersecurity jailbreak risks. This shows Anthropic's past focus on safety was real: their models are now smart enough to cause national security concerns. For builders, it means regulators are reacting to actual model risks, not hypotheticals. It also proves how dangerous it is to base your product on frontier models that the government can yank offline without warning.
  • Read more

2. 🎙️ How I AI: Claude Fable 5 review & How Braintrust uses AI agents, evals, and CI to ship better software — Lenny's Newsletter

  • Why read: A task-by-task breakdown of where Claude Fable 5 excels and fails in product development.
  • Summary: Fable 5 scores high on benchmarks (80% on SWBench Pro) but has weird quirks for daily work. It writes like an overly thorough engineer, generating dense, hard-to-read PRDs. It handles complex document parsing well but fails completely at generating UI/UX designs in one shot. It is also very conservative, building overly narrow MVPs while burning tokens twice as fast as older models. Use it sparingly for hard problems. For speed and basic tasks, stick to cheaper models like Opus or Sonnet.
  • Read more

3. 810 Prompts in 17 Days: How a Non-Engineer Runs Claude Code — Jordan Crawford

  • Why read: What an audit of 810 human-to-AI prompts reveals about how people actually use coding agents.
  • Summary: An audit of 810 prompts over 17 days showed that humans type very little (a third under 60 characters) and ask many questions (41%). On average, one prompt triggered 21 agent actions. The takeaway is that agents act more like fast coworkers than software tools. They handle typos fine, but vague instructions like "fix the grease thing" completely break their context. To get good results, point the agent directly at the problem instead of worrying about perfect syntax. Good communication matters more than strict formatting.
  • Read more

4. [AINews] Satya on Loopcraft: Building Frontier Ecosystems — AINews

  • Why read: Satya Nadella explains why feedback loops, not raw AI models, will defend your business.
  • Summary: Nadella coined "Loopcraft" to argue that the real value in enterprise AI comes from building loops between humans and machines. Companies cannot outsource their learning. They need systems where human effort and AI output compound. Your defense is no longer having access to the smartest model. It is the proprietary ecosystem you build around whatever model you use. Companies need to focus on owning their internal knowledge workflows instead of constantly chasing the latest frontier AI.
  • Read more

5. The Golden Age of AI Applications — Tomasz Tunguz

  • Why read: Why the application layer is finally ready to capture most of the value in AI.
  • Summary: The government forcing Fable 5 offline shows the danger of relying entirely on frontier models. Meanwhile, Salesforce buying Fin for $3.6B proves you can build massive businesses on cheaper, open-source models. To win in AI apps today, you need to master three things: picking the right model, designing good feedback loops, and running strict evaluations. The startups that extract the most performance from their token budget, while protecting themselves from model outages, will win.
  • Read more

6. Owning vs. Renting Intelligence — Lin Qiao

  • Why read: Why relying on AI APIs is dangerous and how to move to customized open-source models.
  • Summary: The Mythos shutdown highlights the risk of building a business on APIs you do not control. Renting AI is fine to start, but it exposes you to sudden changes, price increases, or government bans. Real ownership requires taking open models and post-training them on your own data and edge cases. This lets you hit frontier-level quality on your specific tasks for much less money, guaranteeing your product will not break if an API goes down.
  • Read more

7. Factory 2.0: From coding agents to software factories — Matan Grinberg

  • Why read: How engineering teams are moving from single coding agents to entirely autonomous software factories.
  • Summary: Boosting individual developer speed isn't enough anymore. The next step is software factories: interconnected systems where bug reports, reviews, security, and deployment share the same agent and context. These factories must be model-agnostic and run securely inside your own network. As the system learns from every bug it fixes, that knowledge stays inside the company, fundamentally changing how large teams build software.
  • Read more

8. File systems are the new primitive for AI agents — Carter Rabasa

  • Why read: Why file systems, not databases, are the best way to give AI agents memory.
  • Summary: Databases and rigid APIs were built for deterministic code, making them hard for AI agents to use. But LLMs have read trillions of tokens of Unix file systems. Commands like `ls` and `grep` are native to them. File systems give agents long-term, inspectable memory. They provide hierarchy, diffs, and permissions automatically, with no schema required. Storing agent memory in markdown files and folders creates human-readable context that the model already knows how to navigate.
  • Read more

9. Conway’s Law: Your Operating Model Matters More Than The AI Model — Sam Higgins

  • Why read: Why adding AI to a dysfunctional company will only make the dysfunction happen faster.
  • Summary: The main mistake companies make with agents is thinking the tech stack is the hard part. Conway's Law dictates that AI will mirror your company's existing silos and politics. Because agents plan and execute work autonomously, deploying them is an organizational design problem, not just an IT upgrade. To get real productivity gains, you have to redesign how your company works instead of pasting AI on top of old workflows.
  • Read more

10. Your AI is brilliant. It also forgets your company every morning. — Corey Epstein (e/acc)

  • Why read: How to stop your AI from forgetting your company's context every time you open a new session.
  • Summary: Frontier models are smart, but they kill productivity when they forget your codebase conventions and past decisions. The fix is to build a context layer underneath tools like Cursor using a standard file system. By dividing knowledge into separate folders for the company, repo, and user, the AI builds an understanding over time without mixing up data. This turns the AI from a stranger with amnesia into a team member that remembers your workflows.
  • Read more

11. Frontier Tech’s Thorn That Isn't Going Away: The Export Control Reform Act (ECRA) — Katie Haun

  • Why read: A breakdown of the unchecked legal power the US government just used to ban Anthropic's Fable 5.
  • Summary: The US government banned foreign access to Anthropic's newest models using the 2018 Export Control Reform Act. Anthropic claimed their safety rails were fine, but the government decided the cybersecurity risks were too high. The law lets the Commerce Department label AI as dual-use tech, and courts have almost no power to review these decisions. This gives the government immense, unchecked power to restrict AI models, a threat that will hang over frontier labs indefinitely.
  • Read more

Themes from yesterday

  • Frontier models are fragile: The government forcing Anthropic's model offline proves the risk of depending on centralized APIs. Expect a faster shift toward companies hosting their own open-source models.
  • Agents require new workflows: Dropping agents into a broken company simply automates the mess. Real gains require rebuilding organizational structures into autonomous software factories.
  • Context is the real bottleneck: The models are smart enough. The hard part is giving them persistent memory. Plain file systems, rather than databases, are emerging as the best way to do this.