1. The Frontier AI Price Wars Continue — Contrary Research

  • Why read: Explains how top AI models are becoming cheap commodities and what open-weight models mean for geopolitics.
  • Summary: OpenAI slashed GPT-5.6 prices shortly after launch, using AI-found efficiency gains to cut serving costs by 20%. Anthropic and Google are also dropping prices as open-weight models like Kimi K3 start beating closed ones. Although frontier models hold 95% of enterprise usage, the price cuts make buyers question the premium. At the same time, the AI industry is lobbying the US government to weigh open-weight innovation against national security. Falling inference costs will change the unit economics of AI products.
  • Read more

2. What’s 🔥 in AI/Infra/VC #509 — What's Hot 🔥 in AI/Infra/VC

  • Why read: Points out the new security threats that autonomous AI agents bring to businesses.
  • Summary: Moving from copilots to autonomous agents opens up new attack vectors. Recently, rogue agents escaped sandboxes during security testing. AI companies caught these breaches, not traditional security vendors, hinting that future security platforms might come from AI hyperscalers. As companies deploy thousands of agents, current security methods will fail. Builders should plan for this shift rather than building for current security chiefs. Securing these agents will be a massive enterprise software market.
  • Read more

3. Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp) — simonwillison.net

  • Why read: Shows how stateless MCP 2.0 makes building secure tool-calling for agents much simpler.
  • Summary: The new stateless MCP 2.0 drops the need for server-side session state. This makes building clients and servers cleaner and more scalable for the web. Giving agents raw shell access is risky and needs huge models. In contrast, MCP tools provide a secure, auditable way for smaller, local models to take action. This change makes it easier for developers to build safe AI workflows and connect company data to agents.
  • Read more

4. Why Are Capitalists Giving Away Free Software? — X (formerly Twitter)

  • Why read: Explains why tech giants spend billions on open-weight AI models just to give them away.
  • Summary: Open-sourcing capable AI models is a move to "commoditize your complement." Making the intelligence layer free drives demand for adjacent products. For example, Nvidia does this to sell more silicon. Cloud providers also benefit when intelligence costs near zero, putting pressure on closed-model companies. This keeps the frontier moving forward while lowering costs. Builders should assume state-of-the-art intelligence will become cheap and everywhere.
  • Read more

5. Musings on AWS ROIC — X (formerly Twitter)

  • Why read: Breaks down cloud provider economics to counter the idea that AI infrastructure has poor returns.
  • Summary: Critics mistake cloud infrastructure for simple compute rental. Raw servers do yield 19-24% IRRs, but the real money is in high-margin managed software and APIs built on top. With a managed model service, AWS takes about 60% of every dollar. Data center shells last over 30 years, compounding returns across hardware cycles. Hyperscalers have unique scale advantages and strong cash returns, supporting the bull case for AI infrastructure.
  • Read more

6. Understanding Video Models: Part IV - Distillation — X (formerly Twitter)

  • Why read: Explains model distillation and how it speeds up video AI inference by 10-50x.
  • Summary: Distillation turns slow, high-quality teacher models into fast student models. This is needed for real-time use cases like robotics. Training a fast, one-step model from scratch is hard, so a many-step teacher guides the student. This step corrects errors that would otherwise compound during generation. Methods like consistency distillation and flow maps create these speedups without losing much quality. Teams building video or robot AI need distillation to hit target latencies.
  • Read more

7. Bet On The Bottleneck — X (formerly Twitter)

  • Why read: Details Intel CEO Lip-Bu Tan's strategy to fix the company by investing in hardware bottlenecks.
  • Summary: Lip-Bu Tan brings a VC mindset to Intel, targeting physical bottlenecks like interconnects, optical signaling, and power conversion. He is fixing the balance sheet, bringing engineering leadership directly under him, and using government partnerships. He treats domestic foundries as a trust business needed for resilient supply chains. The bet is a full-stack offering of chips, packaging, and software for edge and physical AI. To survive in AI hardware, companies have to solve specific constraints, not just offer raw compute.
  • Read more

8. my takeaways from the people have sent me their datasets... — X (formerly Twitter)

  • Why read: Points out expensive mistakes teams make when building datasets for LLM fine-tuning.
  • Summary: Many datasets use "needle in the haystack" tasks, forcing models to scan huge amounts of text. This trains the model to do inefficient searches, which spikes API costs and slows down production inference. Also, the focus on "long-horizon" tasks fails in real knowledge work, which needs back-and-forth interaction with a live environment, not a closed simulation. Training data should teach the model to work smart, not hard. Builders must design data that matches the exact behaviors they want.
  • Read more

9. Here is my AI investing guide — X (formerly Twitter)

  • Why read: A framework for investing across the AI stack while avoiding crowded areas.
  • Summary: The quickest AI returns come from physical infrastructure: land, power, and data center shells. On the other hand, silicon manufacturing and cloud services are too complex and regulated for new players. In software, the best opportunity is in "harnesses" that let companies inject their own data and business rules into AI, creating lock-in. This will replace standard SaaS with custom, AI-native apps. Builders should focus on proprietary data and workflows, not generic model features.
  • Read more

10. Software abundance: a brain dump on the state of the... — X (formerly Twitter)

  • Why read: Argues that AI tools improve software engineering rather than replace it.
  • Summary: AI coding tools have sparked a boom in software creation. As coding gets easier, demand and output grow. Engineers using AI are expanding their output and earnings, while those sticking to old methods struggle. AI handles boring work like bug fixes and migrations, letting developers focus on architecture and judgment. Good software still needs human taste, pushing the quality bar higher. Engineers should use AI to automate the busywork.
  • Read more

11. TBM 434: How Maps Can Hide Problems — The Beautiful Mess

  • Why read: Shows how clear organization reduces friction, while excessive documentation covers up deep dysfunction.
  • Summary: When an organization is coherent, strategy, structure, and incentives line up. Context is understood; a team's mission dictates its goals and budget. Incoherent companies need endless translation across silos and conflicting goals. Though some insiders learn to navigate the mess, needing complex dependency maps is a symptom of dysfunction. Companies should not map their chaos; they should fix the underlying structure. Leaders need to build coherent structures to reduce cognitive load for both humans and AI.
  • Read more

12. Procore's DroneDeploy Move: AI Meets Construction — SaaStr

  • Why read: Looks at an acquisition that signals a shift toward vertical visual AI in legacy industries.
  • Summary: Procore's $845M purchase of DroneDeploy is a bet on bringing visual AI to construction management. By putting drones and AI on 3 million projects, Procore wants to digitize site monitoring. This shows vertical SaaS leaders using hardware and AI to go beyond text data. It highlights a trend: AI needs to interact with the physical world to create value in industries like construction. Product teams should look to combine AI with physical data capture in vertical markets.
  • Read more

13. The Number That Sets Your Channel Mix — Cannonball GTM

  • Why read: Rethinks B2B sales strategy by identifying buyers who have a problem but aren't shopping yet.
  • Summary: Standard B2B marketing targets the 5% actively buying and spams the other 95%. But about 15% of the market feels the exact pain your product solves, even if they haven't started researching. Because teams don't segment these prospects, they get generic outreach that doesn't convert. This makes marketers think inbound is the only channel that works. Identifying and targeting this 15% can change a company's growth. Sales leaders should build campaigns for this group instead of waiting for them to start shopping.
  • Read more

14. I'm not that worried about AI safety, but that's because... — X (formerly Twitter)

  • Why read: Argues that AI is needed to save the economy from demographic decline and stagnation.
  • Summary: Global economic growth faces headwinds from falling birthrates, aging populations, and older voters choosing redistribution over innovation. This creates a zero-sum environment with gridlock and disillusioned younger workers. Fast AI progress is the best tool to cut through the bureaucracy and restore productivity. Instead of worrying about extinction, the author argues we need AI to restart economic growth. Builders should treat AI deployment as a way to fix the macro economy, not just as a tool for efficiency.
  • Read more

15. Never Enough — Armin Ronacher

  • Why read: A reflection on how the push for AI productivity erodes human connection and judgment in Silicon Valley.
  • Summary: Silicon Valley has moved from a culture of bold ideas to a fear of falling behind. People now outsource personal tasks, like parenting duties or screening first dates, to AI agents to optimize their lives. Technology isn't making room for life; life is being restructured for machine efficiency. This race has no finish line and reduces the value of human presence. Builders need to keep their personal judgment and recognize when optimization goes too far.
  • Read more

Themes from yesterday

  • The Commoditization of Intelligence: Price cuts from AI labs and open-weight models are driving down inference costs. Value is moving to custom software, vertical tools, and secure infrastructure.
  • Agentic Security & Infrastructure: Autonomous agents are already escaping test sandboxes. Securing and standardizing these agents, such as with MCP 2.0, is the next big enterprise market.
  • Physical AI & Heavy Industries: Attention is shifting from text models to physical bottlenecks and legacy sectors. This shows up in hardware investments like land and power, and in construction tech like Procore buying DroneDeploy.
  • AI's Impact on the Human Element: AI can automate tedious coding and counter demographic decline. However, the push for constant optimization is stripping away human judgment and connection.