1. Sweat The Tokens — Gokul Rajaram

  • Why read: How Anthropic shifted from benchmarking models to shipping consumer products, and what that means for AI product management.
  • Summary: AI product management demands treating token-level evaluations like pixel-level UI design. Because model capabilities jump unpredictably, the main job is measurement: turning failed user transcripts into research evals. Evaluations are replacing Product Requirements Documents. A model without an interface is a benchmark; an interface without a model is a demo. Small, opinionated teams are best equipped to build these products.
  • Read more

2. We're not using enough inference — Rhys

  • Why read: A look at how near-zero inference costs will change software engineering.
  • Summary: As inference gets cheaper and agents get better, the engineering bottleneck shifts from human attention to compute allocation. AI agents will constantly monitor systems, instantly handling tasks like security patching and debugging that currently require human response. Defense will require automated fixes to counter automated attacks. Autonomous deployment has risks, but relying on human speed will be riskier. Existing tools need to adapt to a world of unlimited, cheap inference.
  • Read more

3. Path to continual learning: Kimi K3's KDA as Fast Programmable Weights & NoROPE for infinite context — GDP

  • Why read: How Moonshot's Kimi K3 architecture enables continual learning by updating world models instead of freezing them at a training cutoff.
  • Summary: Continual learning requires models to update their worldviews through interaction, rather than freezing after training. Kimi K3 hits high performance using a fixed-size, compressed memory for 75% of its attention layers, rather than caching exact tokens. This approach mimics human memory, storing a compressed world model instead of a verbatim archive. By dropping ROPE to get infinite context and using linear transformers as fast weight programmers, K3 shows exact token caching isn't required for high performance. This could change how models learn after deployment.
  • Read more

4. every company needs a cassandra — Sunil Pai

  • Why read: The case for deploying AI agents as designated skeptics to counter corporate groupthink.
  • Summary: Companies struggle with self-criticism because employees who disagree lose social capital. An AI agent called "Cassandra" can act as a skeptic in Slack without worrying about performance reviews or ego. The agent observes and speaks up only when it disagrees with the consensus, like noting a competitor's overlooked feature. It builds trust by sometimes confirming it can't find a counterargument. This uses AI to institutionalize dissent and improve decisions.
  • Read more

5. The playbook for building high talent density teams | Adam Ward, Head of Talent at Cursor — Lenny's Newsletter

  • Why read: Tactical advice from Cursor's talent lead on hiring top engineers in a competitive market.
  • Summary: Standard recruiting funnels produce average hires through passive filtering. To build dense talent, treat every role like an executive search: scope, map, and pursue aggressively. The market is split, and founders have to work hard to win top candidates. The "forward deployed engineer" role is changing how technical staff interact with the business. Founders should avoid mis-hiring their first recruiter, which can ruin company culture.
  • Read more

6. there are some really interesting rumors going around related to... — Jack Morris

  • Why read: Rumors on how Chinese AI labs might be extracting hidden reasoning traces to improve open-weight models.
  • Summary: Rumors suggest Chinese AI labs reverse-engineered reasoning traces from models like Claude Code and Codex. Distillation relies on these hidden reasoning chains. By collecting long-horizon data with these traces, open-source models have improved rapidly. Research shows that training a reasoning inverter can reconstruct traces from final outputs alone. If accurate, extracting proprietary reasoning could threaten closed-source models.
  • Read more

7. vibecoding gpu kernels — maharshi

  • Why read: How to use LLMs to automate writing GPU kernels.
  • Summary: Writing GPU kernels is highly verifiable, making it ideal for autonomous LLM loops. The workflow checks for compilation errors, verifies correctness against a slow reference, and optimizes for performance. Giving the agent access to Domain-Specific Language directories (like NVIDIA cutlass) helps it write better first-pass code. Since correctness is mathematically verifiable, the agent can optimize performance without supervision. Verifiability is the key requirement for deploying AI coding agents.
  • Read more

8. A Bitcoin Miner Switches to AI for a $4.7 Billion Lease — Chamath Palihapitiya

  • Why read: How the AI compute boom is changing data centers and energy use.
  • Summary: Bitdeer leased a Norwegian Bitcoin mining facility for $4.7 billion to convert it into an AI data center. Because securing grid connections takes years, pre-permitted land with power is highly valuable. This site runs on hydropower and has strong cooling, which AI workloads need. "Land Power Shell" (LPS) is currently the fastest way to generate returns in AI. As inference costs approach zero, physical infrastructure and software harnesses will be the main competitive moats.
  • Read more

9. It’s important to understand that Harvey and Anthropic have completely... — Gordon Cassie

  • Why read: Why standard SaaS subscriptions don't work for AI applications that pay per token.
  • Summary: AI model providers and AI application layers have conflicting economics. Providers like Anthropic have fixed server costs and use rate limits, which fits subscription pricing. Applications like Harvey pay per token, so their margins shrink as usage grows. This forces low-usage customers to subsidize power users, which fails when a product becomes heavily used. AI applications will need to adopt usage-based billing or build their own infrastructure.
  • Read more

10. Owner of Everything, Owner of Nothing — Regina Gerbeaux

  • Why read: Why top operators in scaling startups often end up with no clear domain ownership.
  • Summary: In fast-growing companies, strong operators become general problem solvers with no defined ownership. They are expected to keep everything running but lack the authority or focus for deep strategic work. By constantly putting out fires, they hide process and organizational flaws. This causes burnout and stops the company from building scalable systems. The fix is to draw clear boundaries and build processes that don't depend on a single generalist.
  • Read more

11. Stop Guessing Job Titles, Find Every Job Title and Filter Down — Jordan Crawford

  • Why read: A B2B sales strategy that relies on scraping employee data instead of guessing job titles.
  • Summary: Standard sales prospecting searches for assumed job titles, which limits results. If you reverse the process—pulling every employee at a target account and filtering down—you find thousands of missed prospects. In one case, searching by title found 60 leads; pulling all employees found 79,000 people with 23,000 distinct titles. The seller's vocabulary shouldn't limit the search. Looking at the raw data reveals the actual organization and improves reach.
  • Read more

12. There Are Three Types of AI Users — David Brooks

  • Why read: How different knowledge workers adapt to AI tools.
  • Summary: Early data shows AI doesn't create leisure time; it makes work more intense. Users spend twice as much time on messaging and software. Instead of relaxing, they take on tasks they used to outsource and work longer hours. The main differentiator won't be intelligence, but a willingness to do mental work. People who use AI to avoid thinking get short-term gains but stagnate. Those who push back and wrestle with the tools will develop new capabilities.
  • Read more

13. The Only Good Ending: Why Human Flourishing Has a Single Economic Architecture — Carlos E. Perez

  • Why read: Why zero-marginal-cost AI forces a redesign of how purchasing power works.
  • Summary: The economy relies on human labor earning wages to fund consumption. AI breaks this loop by pushing the cost of cognitive work toward zero. The economic gains go to infrastructure owners, leading to high production but a shrinking consumer base. Because debts are fixed but incomes are falling, this risks debt deflation. Society will need a new economic structure, as the wage-based model is breaking.
  • Read more

14. The Victorverse — David Hoang

  • Why read: How Bret Victor’s interface theories apply to LLMs.
  • Summary: Bret Victor predicted computing should be a responsive medium, not a set of manual controls. He argued software should infer context and present tailored information instead of relying on rigid interfaces. With LLMs, static UIs are shifting to context-aware environments. The future of tooling is visual, direct-manipulation systems rather than faster syntax editors. Good software will act as an anticipatory environment, not a mechanical tool.
  • Read more

15. Kaushik Subramanian: How to Find Promising Fintech Startup Ideas — eqtgroup.com

  • Why read: How to find B2B software ideas by looking for process friction instead of picking a sector.
  • Summary: Founders fail when they pick a sector and try to invent a product. Instead, follow the money through a business and look for where it slows down due to manual entry or disconnected systems. A good idea passes two tests: the workflow can be automated with existing data, and the solution improves the customer's P&L. If a company spends heavily on a bottleneck, they will buy an automated solution. The best opportunities sit at the intersection of technical feasibility and existing corporate spending.
  • Read more

Themes from yesterday

  • AI forces organizational change: Companies are treating AI evaluations like product design and using AI agents to break group consensus.
  • The AI SaaS pricing model is breaking: Startups paying per token lose margin as usage scales. They will need to shift to usage-based billing or build their own infrastructure.
  • Macroeconomic shifts: Zero-marginal-cost AI threatens the wage-consumption loop by decoupling human labor from productivity, raising the risk of debt deflation.
  • Data scraping over assumption: Operators are using raw, bottom-up data—like full org charts or money flows—instead of relying on top-down assumptions for sales and product development.