1. The Future Worth Building Is Human — thinkingmachines.ai

  • Why read: Mira Murati's new AI lab is betting on customizable, specialized models over centralized AGI.
  • Summary: Thinking Machines argues centralized AI falls short for real-world tasks because valuable knowledge is local and distributed. Instead of renting generic AI, they want users to train and customize model weights for specific needs. The goal is building AI that extends human judgment rather than replacing it. It is a move away from "one model rules all" toward specialized, owner-operated systems. Moving forward, competitive advantage will mean owning custom AI.
  • Read more

2. Meta's New Model & The AI Pricing Debate — Contrary Research

  • Why read: Meta enters the frontier model race with Muse Spark 1.1 and severely undercuts the competition on price.
  • Summary: Meta's Superintelligence Labs launched Muse Spark 1.1 to compete with GPT-5.5 and Opus 4.8. At $1.25 per million input tokens, it costs a fraction of Anthropic's Fable 5. Mark Zuckerberg called this a direct strike at competitor margins. As enterprises question the ROI of expensive frontier models, Meta's aggressive pricing is poised to trigger a race to the bottom for AI inference costs, altering the unit economics of AI-native apps.
  • Read more

3. 30x Fewer Agents, 87% Agreement — Substack

  • Why read: A case study showing how separating data collection from AI judgment improves agent architecture.
  • Summary: An author tried scoring 1,145 companies by deploying 1,145 AI researchers with large context windows. It was slow and error-prone. They fixed it by using standard code to gather public data, then passing it to a small crew of AI graders. The new setup matched the original approach 87% of the time while saving compute and time. It turns out massive context windows distract models and degrade reasoning. For better scaling, keep data gathering separate from AI evaluation.
  • Read more

4. The Bitter Lesson and Hayek's Revenge — Soren Larson

  • Why read: Applies Hayek's economic theories of distributed knowledge to current AI compute constraints.
  • Summary: Hayek argued central planning fails because knowledge is dispersed, requiring local decision-making. Massive LLMs attempt to aggregate all context centrally, but energy and latency constraints make this inefficient for localized tasks. Moving data to centralized compute is simply too slow and expensive. Instead, AI is shifting toward "compute on the spot"—specialized, context-aware instances running at the edge. Future infrastructure will favor distributed, specialized models over single, centralized systems.
  • Read more

5. "Token usage" is the least thing that you should worry about with AI — Substack

  • Why read: The real cost of AI isn't token usage; it's the false confidence generated by authoritative but flawed outputs.
  • Summary: Teams worry about API costs while ignoring the damage of AI-driven false confidence. Executives and PMs are building roadmaps on shallow, unverified AI research. Because LLMs sound authoritative, people trust them with work they would double-check if a junior employee submitted it. This blind trust leads to bad strategy and wasted engineering time. Teams need strict verification rules before using AI research to allocate resources.
  • Read more

6. What’s 🔥 in Enterprise IT/VC #506 — Substack

  • Why read: A VC take on why enterprises are moving from renting generic AI to owning specialized models.
  • Summary: As new models from Meta, X, and OpenAI drive down prices, a new enterprise thesis is forming: specialized intelligence. Companies are building custom AI using their own data. This lets them use their specific advantages instead of relying on rented models. Startups are applying this playbook to robotics, biology, and automation. Going forward, a founder's best moat will be deep domain expertise and proprietary reinforcement learning environments.
  • Read more

7. AI 2040 and the Cult of Intelligence — the singularity is nearer

  • Why read: An engineer argues against the "hard takeoff" AGI narrative by pointing out physical constraints.
  • Summary: The idea that superintelligent AI will rapidly take over the world ignores the reality of hardware and supply chains. Tokens cannot bypass the months it takes to fabricate a chip or manipulate physical matter. The author argues that AGI regulation mostly serves to create bureaucracies that restrict compute access. They suggest building hyper-local, uncensored personal AI that serves individuals directly. Ignore sci-fi doomerism and focus on local utility.
  • Read more

8. The King Bleeds, Long Live The King — substack.com

  • Why read: The strategic realities of the semiconductor industry and the real story behind the Groq and Cerebras exits.
  • Summary: Nvidia maintains its dominance in the semiconductor market despite competition from alternative architectures. Groq's $20B sale to Nvidia appears to be a defensive move reacting to Cerebras partnering with OpenAI. These hardware startups survived by managing supply chains and personnel, not just engineering. While startups take on massive risk, the moves of giants like Nvidia and OpenAI dictate the market. In hardware, business strategy routinely beats technical superiority.
  • Read more

9. Rat's Nest Problems — Substack

  • Why read: A framework for understanding problems that seem unsolvable due to complex dependencies.
  • Summary: A "Rat's Nest Problem" is a web of dependencies where pulling one thread tightens the knot elsewhere. Solutions that start with "if everyone would just..." always fail because coordinating humans at scale is hard. Product builders need to recognize these problems to avoid wasting time on naive fixes. You cannot cut the knot. You have to map the dependencies and accept that solving one issue creates another. Build systems to manage trade-offs instead of looking for silver bullets.
  • Read more

10. AI Can't Recreate Thrust (But It Can Help You Understand It) — jamesdrandall.com

  • Why read: LLMs fail at generating game physics but excel at decoding legacy assembly code.
  • Summary: The author asked Claude to recreate the 1986 game Thrust, but it failed to match the original physics and momentum. However, when given the disassembled 6502 assembly code, Claude easily explained the legacy systems and level data. The AI wasn't good at writing the game from scratch, but it was excellent at analyzing low-level code to guide a human developer. Right now, AI is better as an interactive decompiler than an autonomous software architect.
  • Read more

11. The Builder's Economy — SaaStr

  • Why read: Metrics and case studies showing how AI and embedded fintech drive modern B2B SaaS growth.
  • Summary: The latest ICONIQ report shows AI moving from an experiment to a revenue driver, nearing 50% of revenue for surveyed companies. Meanwhile, ServiceTitan hit $1B+ ARR, with their embedded fintech arm growing faster than core software subscriptions. Presentation app Gamma reached $100M ARR with 50 employees and zero marketing spend. B2B SaaS growth is shifting away from traditional seat-based models toward embedded financial services and lean, AI-driven teams.
  • Read more

12. TIL: Bob Proved It Was Impossible. Two Months Later He Shipped It. — Semi Doped

  • Why read: A historical lesson on ignoring conventional wisdom to build breakthrough hardware.
  • Summary: In the 1960s, engineer Bob Widlar changed integrated circuits by refusing to design them like discrete circuits. While experts tried to fit large resistors onto silicon, Widlar used cheap transistors instead. He invented the Widlar current source and the first op-amp on a chip. He succeeded by designing natively for a new medium instead of porting over old constraints. For AI builders, the lesson is to design for the realities of the new technology rather than forcing old paradigms onto it.
  • Read more

13. David Shelley: How to Get Your Book Published — Substack

  • Why read: The CEO of a Big Five publisher explains the mechanics of traditional publishing in the self-publishing era.
  • Summary: David Shelley, CEO of Hachette Book Group, details what publishers actually do today. They manage the logistics of shipping 150 million physical books a year, oversee manufacturing, and drive demand. Their real value is getting bookstores to stock a title and convincing consumers to buy it. It highlights why physical distribution, quality control, and curated marketing still matter.
  • Read more

14. Longreads + Open Thread — Ghost

  • Why read: How algorithmic feeds damage reading stamina and change how we process information.
  • Summary: Algorithmic feeds condition people to expect immediate comprehension, eroding their ability to read challenging texts. Physical books survive as high-density information stores that demand sustained attention. The author also predicts that as AI agents learn to anticipate our needs, our interactions with them will drop pleasantries in favor of dense information. Our media diet shapes our cognitive habits, and long-form reading remains necessary for deep understanding.
  • Read more

15. The Proportional Web — owickstrom.github.io

  • Why read: How to apply classic print typography principles to modern web design.
  • Summary: Using Robert Bringhurst’s The Elements of Typographic Style as a guide, the author built a CSS framework for print-quality, justified text on the web. By using modern CSS features like `word-break` and `hyphens`, they achieved text alignment long considered impossible or taboo for websites. The system scales across devices using relative measurements and a single font family. It is a practical guide for developers wanting highly readable long-form content.
  • Read more

Themes from yesterday

  • Local and Specialized AI: Companies are moving away from renting centralized, generic models. They prefer owning custom, local intelligence trained on their own data.
  • AI Pricing and ROI: Pushback on AI costs is causing aggressive price cuts, like Meta's Muse Spark 1.1. At the same time, teams are discovering the hidden costs of AI-driven false confidence.
  • Architecting for Reality: Builders are finding that LLMs are better at targeted tasks, like code analysis, than autonomous generation. They are adjusting architectures to separate standard data gathering from AI evaluation.