1. The Pulse: Interesting AI coding stats from Cursor — The Pragmatic Engineer

  • Why read: Shows how top engineers actually use AI coding agents in production.
  • Summary: Cursor's two-year usage data shows a steep power law. The top 1% of users generate 30-40K lines of code per week, roughly equaling 45 average developers. Meanwhile, 90% of token usage comes from reading the codebase, not outputting code. This validates the traditional 10:1 read-to-write engineering ratio. Because of this read-heavy usage, the main cost of AI coding agents is context processing, which Cursor caches aggressively to keep costs down. If you are building custom agent harnesses, you need context-caching infrastructure. This data also complicates how we measure software value now that a single developer can produce so much code.
  • Read more

2. [AINews] SpaceXAI launches Grok 4.5, first Opus-class model post Cursor acquisition — Substack

  • Why read: Covers the newest frontier model designed for coding and agent workflows.
  • Summary: xAI released Grok 4.5, focusing on capability-per-dollar and inference speed over raw benchmark scores. At $2 per 1M input tokens, it undercuts GPT-5.6 and Claude Opus 4.8 while keeping a 500k context window. xAI partnered with Cursor to train the model specifically for software engineering and agentic workflows. This marks a shift for major labs toward developer utility and cost efficiency instead of general chat. Teams building agent systems can now get near-Opus performance at better unit economics.
  • Read more

3. The Future of Meta Superintelligence: A 1 Year Progress Update — SemiAnalysis

  • Why read: Explains why Meta could challenge the OpenAI and Anthropic duopoly.
  • Summary: Meta is scaling up talent, compute, and proprietary data for its upcoming Muse Spark and Llama models. Although its recent releases fell behind other open-source models, Meta is building a large reinforcement learning environment using real-world tasks and employee data to improve reasoning. The piece argues that reinforcement learning is the current primary scaling law, demanding custom environments and verifiers rather than more internet data. With its mix of massive compute and internal data generation, Meta could overtake its rivals in the next six months. Custom RL environments are becoming the actual moats for frontier labs.
  • Read more

4. Everyone's Watching the Wrong Benchmark — Gradient

  • Why read: Exposes the hidden industry of expert data generation separating open and closed models.
  • Summary: Public benchmarks make it look like open-weight models are catching up to frontier labs, but those tests are gameable and saturated. Real frontier capabilities, such as PhD-level scientific reasoning or expert financial analysis, are evaluated on private benchmarks where open models still fall short. To close this gap, frontier labs spend billions annually on proprietary expert data, paying professionals to document their reasoning processes. This builds a data moat that public internet scraping can't match. Going forward, the most valuable AI capabilities will depend on these private datasets.
  • Read more

5. Where the returns to AI go — economicforces.xyz

  • Why read: Explains why pure intelligence is becoming a commodity and where the financial value will actually accrue.
  • Summary: AI companies have massive valuations, but the price of their core product, intelligence, is dropping by an order of magnitude each year. As intelligence becomes abundant, it gets cheap at the margin, regardless of its total utility. The financial returns won't go to intelligence itself, but to the scarce resources needed to deploy it. This means proprietary data, compute infrastructure, specialized workflows, and distribution channels. Companies should focus on controlling these complementary assets rather than competing on raw intelligence.
  • Read more

6. Teaching a child in <1000 ms: the architecture behind a real-time tutor — ello.com

  • Why read: Details how to architect AI agents for sub-second latency to keep users engaged.
  • Summary: While building an AI tutor for children, the Ello team found that standard agent tool loops caused 3-4 seconds of latency, which made kids lose interest. They fixed this with a custom harness that separates model generation from execution. This allows the system to stream multiple actions in one response. An interpreter runs the early actions while the model generates the rest, getting an initial response to the child in milliseconds. This shows that user experience, not default LLM behaviors, should set engineering constraints. You can use this streaming-action approach to build much faster voice and interaction agents.
  • Read more

7. We 5x'd conversion and 2x'd activation with context engineering — X (formerly Twitter)

  • Why read: Shows that structuring context often improves AI agents more than upgrading model reasoning.
  • Summary: PostHog built an AI wizard to simplify its onboarding, cutting two hours of manual setup down to eight minutes. They found that advanced reasoning techniques offered little improvement compared to giving the agent accurate, structured context. They turned their technical writers into "context engineers," building a system that continuously feeds up-to-date documentation to the agent. Decoupling this knowledge from the agent itself means the AI always has the specific details needed for varied tasks. For product teams, the main bottleneck is usually missing context, not poor reasoning.
  • Read more

8. Jensen Huang x LangChain: Every Company Is Built on Intelligence, So Nobody Should Rent Theirs — X (formerly Twitter)

  • Why read: Covers NVIDIA's push for enterprises to own their AI via open-weight models and secure runtimes.
  • Summary: NVIDIA CEO Jensen Huang says AI only became truly useful in the last six months, sparking demand for controlled, local deployments. He argues that cheap, fast intelligence lets agents search larger problem spaces and iterate more often, producing better results than slower, expensive frontier models. LangChain and NVIDIA released NemoClaw, an open runtime combining agents with the Nemotron 3 Ultra model. It performs close to Claude Opus at a fraction of the cost. This allows companies to run specialized agents safely on-prem or in sandboxed clouds without leaking proprietary data. Enterprise AI is heading toward specialized, owned agents instead of rented general models.
  • Read more

9. Gavin Baker x Generating Alpha: If AI Eats the World, Silicon Eats the World — X (formerly Twitter)

  • Why read: Offers an investor's case for why the AI infrastructure buildout isn't a bubble.
  • Summary: Tech investor Gavin Baker argues that current AI valuations are backed by cash flow and positive ROI, unlike the dot-com era. Despite a temporary dip in ROI from massive training spending on Blackwell chips, the performance of models like GPT-5.2 points to high future returns. He notes a global shortage of power and manufacturing capacity, highlighting TSMC as the main bottleneck that will prevent overbuilding. Also, while AI will cut software and game development costs, inference and rendering will still demand heavy local GPU usage for years. Expect long-term constraints on compute resources, not an oversupply.
  • Read more

10. I have a fun new paper today w/ @joshgans: what... — X (formerly Twitter)

  • Why read: A framework for optimizing AI models based on how humans actually interact with them.
  • Summary: Researchers Kevin Bryan and Josh Gans offer an economic model that defines AI value based on human decision-making. They show that maximizing a model's accuracy on standard benchmarks doesn't maximize its real-world utility. The best AI design depends on the user's cost to verify the output and the penalty for mistakes. Since people switch between trusting the AI and double-checking its work, the ideal training target isn't a smooth curve. Product builders should optimize their models for the end-user's specific workflow and risk tolerance, not general leaderboards.
  • Read more

11. Adam Mosseri: AI is a tailwind for authenticity — Substack

  • Why read: The Head of Instagram explains why AI-generated content pushes users toward authentic human connection.
  • Summary: Adam Mosseri argues that as synthetic media floods the internet, users will place a higher premium on verified human experiences and genuine identity. He also notes that Meta's product teams are shifting from large, specialized groups to small pods of generalists. This creates a new "product staff" role combining design, product, and data science. With AI handling detailed execution, operators who can work across disciplines and set product strategy are becoming more valuable.
  • Read more

12. Cult Holdings Co-Founder @willhmayer, who led the Rick Rubin-Polymarket ad... — X (formerly Twitter)

  • Why read: How to use psychology and branding to build a deeply loyal user base.
  • Summary: Will Mayer outlines five tenets of a "cult" brand: doctrine, ritual, symbols, a charismatic leader, and an enemy. Brands need to define their identity before shipping features. They should create daily rituals that turn user habits into part of their identity. Specialized language and symbols establish an in-group, similar to CrossFit terminology. Finally, picking an ideological enemy, like Apple fighting conformity, unifies the community against a shared threat. Founders can use these ideas to turn a utility product into a lifestyle.
  • Read more

13. The top 10 GTM mistakes founders make — Substack

  • Why read: Advice on the common go-to-market mistakes that sink early-stage startups.
  • Summary: After advising over 150 founders, Arnie Gullov-Singh lists the GTM errors that product builders often make. A major one is using an investor pitch on buyers, who only want to solve immediate, expensive problems. Another is targeting multiple customer profiles at once, which keeps messaging too broad to build a repeatable sales process. He also advises against scaling cold outbound before using warm intros to learn what buyers actually care about. Focus on a single customer profile and listen during discovery calls before trying to scale sales.
  • Read more

14. 8.9 Million AI Users — tomtunguz.com

  • Why read: Details the rapid growth of local, open-source AI in the enterprise.
  • Summary: Ollama has reached 8.9 million developers, adding nearly a million users a week by making it easy to run open-source models locally. Driven by data privacy requirements, 85% of the Fortune 500 use Ollama in secure settings. Developers can run inference locally and offload complex tasks to the cloud, making Ollama effectively the "Docker for AI." Its recent $65M Series B suggests that enterprise AI will rely heavily on locally owned, open-weight models rather than just centralized API providers.
  • Read more

15. [Job Market Alert] AI Isn't Killing Jobs. Here's What's Really Happening. — beehiiv.com

  • Why read: A look at the data on tech jobs, challenging the idea of an AI-driven employment collapse.
  • Summary: Despite headline layoffs at large tech companies, an analysis of a billion job postings shows AI is often just a cover for cutting post-2021 bloat. The market has split in two. Jobs based on easily automated tasks are shrinking. But roles where AI amplifies human judgment are seeing double the job growth and 42% faster salary growth. Companies that are hiring are growing faster than those cutting staff. Workers need to move away from routine execution and into roles that require strategic judgment.
  • Read more

Themes from yesterday

  • The true AI moat is data and context: Raw intelligence is becoming a commodity. The real advantage goes to those who own high-quality expert data (Private Benchmarks, Meta's RL) and can feed that context directly to agents (PostHog).
  • Enterprises want local and open AI: Companies are pushing to own their intelligence. They prefer running efficient open-weight models locally or in sandboxes (NVIDIA/LangChain, Ollama) over renting generalist APIs.
  • Agent UX is maturing: We are moving past basic chat interfaces and slow execution loops. Builders are designing systems for sub-second streaming actions (Ello) and tuning them for specific workflows like coding (Cursor, Grok 4.5).
  • Data corrects the macro narrative: Despite the hype, the numbers tell a straightforward story. The compute buildout is tied to real ROI and physical constraints (Gavin Baker), and the job market is reorganizing around AI-amplified roles rather than disappearing.