1. How The Fastest AI-First Companies Really Work — NFX

  • Why read: How to build AI-native teams for speed instead of headcount.
  • Summary: Successful AI-first companies are dropping traditional departments for "mission pods." These teams treat AI agents as coworkers. AI fills skill gaps, letting pods prototype and ship without waiting on other groups. Founders need low-ego hires who communicate well and can run many experiments at once. The main startup bottleneck isn't shipping the first version anymore; it's building a system that learns fast from real users.
  • Read more

2. Steps of AI Adoption — claude.ai

  • Why read: How companies move from gated AI chat to autonomous agents.
  • Summary: AI adoption happens in four phases. It starts with supervised pair programming where humans review everything. As models get better at testing their own code, engineers turn into orchestrators managing multiple agents. The bottleneck moves from writing code to steering the AI. To reach supervised autonomy, companies have to give agents wide context, automated tests, and safe sandboxes. Engineering leaders need to build infrastructure for asynchronous, agent-driven work rather than real-time human oversight.
  • Read more

3. Clouded Judgement 7.17.26 - Open Weights, Closed Prices? — Substack

  • Why read: Why open-weight models aren't automatically cheaper than closed ones anymore.
  • Summary: The 2.8-trillion parameter Kimi K3 model breaks the rule that open-weight means cheap. K3 hits near-frontier performance, but its $5.40 per million token API price is close to proprietary models like Opus 4.8. If you self-host it, the infrastructure costs wipe out the savings of a free license. Since base model access is commoditizing, value is shifting to token efficiency and application workflows. Operators need to model total cost of ownership before assuming open-weight is the budget option.
  • Read more

4. I thought data companies are going to die, I was... — X (formerly Twitter)

  • Why read: Why proprietary data is becoming the best competitive advantage in AI.
  • Summary: Compute is commoditizing and researchers bounce between labs, leaving proprietary data as the strongest moat. As models move into specialized industries, they need deep, domain-specific data. Companies combining research with vertical expertise will become key partners for model developers. This means enterprises will rely less on a few big labs and more on specialized data providers. Founders should focus on building unique data loops instead of just wrapping models.
  • Read more

5. FastAPI for Agents — ashpreetbedi.com

  • Why read: Why standard APIs break when serving AI agents and how to fix it.
  • Summary: Standard FastAPI assumes stateless, fast requests. AI agents need persistent memory and long-running processes, so treating them like simple HTTP requests causes dropped connections and lost state. The right setup makes a "run" the unit of work and a "session" the unit of state, checkpointing at every step. AgentOS provides a layer over FastAPI for this, allowing horizontal scaling and resumable streams. Teams need stateful abstractions to connect agents to tools like Slack or web UIs.
  • Read more

6. The Hassabis Manifesto — Sebastian Mallaby

  • Why read: Demis Hassabis's plan to regulate AI based on what models can do, not what they cost to train.
  • Summary: DeepMind CEO Demis Hassabis proposed a U.S.-based AI safety strategy: an industry-funded regulatory body that can ban unsafe deployments. He wants to evaluate models on actual capabilities, not training compute. This shifts focus to the risks of cheap, fast-following open-weight models. The plan favors fast U.S. action over slow international agreements, using market access to set global standards. By using industry money, the group could hire top talent and bypass Congress. Operators should expect future rules to focus on model outputs, ignoring training scale or open-source status.
  • Read more

7. We mapped the entire physical AI stack. 400+ companies - world models/VLAs, data loops, infra! — X (formerly Twitter)

  • Why read: A look at the 400+ companies building the physical AI and robotics stack.
  • Summary: Physical AI is the next platform shift. Machines using world models and vision-language-action architectures are now acting in the real world. A new map of over 400 companies shows a complete stack: intelligence models, hardware, data loops, and infrastructure. Many of these startups are quiet but ready to change industrial automation. Physical AI isn't a research project anymore; it's a live ecosystem. Operators and investors should look for the empty spaces in this stack to find early opportunities.
  • Read more

8. A Practical Guide to Outcome-Based Pricing — beehiiv.com

  • Why read: How to price AI products based on outcomes without confusing your customers.
  • Summary: Outcome-based pricing means billing for clear, repetitive actions instead of vague ideas like revenue influenced. A good billable outcome is indisputable and frequent, like a booked meeting or a generated image. Keep complex value pitches in the sales deck; the pricing metric has to be strictly verifiable to avoid disputes. This forces product teams to tie software directly to customer value. To do this, you need real-time billing infrastructure to track granular usage.
  • Read more

9. How to Run 7-Figure ABM Campaigns on LinkedIn With a Team of One — Substack

  • Why read: How one person used AI to manage a six-figure LinkedIn ad budget.
  • Summary: A skeleton marketing team used Claude to run a $900,000 pipeline playbook. The setup uses AI to audit live ads, balance messaging, check revenue goals against budgets, and generate reports. This replaces spreadsheet work that used to take large teams hours to finish. With AI handling the manual work, operators can focus on strategy and creative direction. Growth leaders should build custom AI workflows for their own bottlenecks instead of trying to automate strategy completely.
  • Read more

10. Scaling to 1 million concurrent sandboxes in seconds — X (formerly Twitter)

  • Why read: How Modal rebuilt container orchestration to handle AI agents at massive scale.
  • Summary: AI agents and reinforcement learning require launching millions of isolated sandboxes at once. Kubernetes fails here because its central state bottlenecks under high-frequency scheduling. Modal rebuilt its platform to drop global coordination, allowing them to spin up 1 million sandboxes in under a minute. They push scheduling straight to load balancers and worker nodes. Teams building agent workflows need to plan for these infrastructure limits and look at serverless platforms built for burst workloads.
  • Read more

11. How Agents actually use your browser: A guide to CDP (Chrome DevTools Protocol) — X (formerly Twitter)

  • Why read: A technical primer on Chrome DevTools Protocol, the system AI agents use to control browsers.
  • Summary: CDP is the core of browser automation tools like Puppeteer, Playwright, and AI web agents. It uses WebSockets to give external programs access to Chromium’s rendering, network, and JavaScript runtime. Since CDP was built for debugging, not autonomous control, it’s a messy stream of asynchronous events that is hard to manage. Engineers building web agents need to understand CDP targets, domains, and event lifecycles to make their agents reliable. Mastering CDP is a major advantage for teams building web workers.
  • Read more

12. The gateway drug for AI at work — Substack

  • Why read: Why an AI Chief of Staff is the best way to get non-technical employees using agents.
  • Summary: An automated AI Chief of Staff that triages email, Slack, and meetings is a great entry point for agent workflows. It runs overnight and delivers a morning brief of urgent decisions, follow-ups, and forgotten tasks. This saves time and catches things humans miss. It also flips employees from skeptics to people looking for more ways to use AI. Setting this up requires careful prompting upfront so the agent understands the user's priorities and context.
  • Read more

13. Don’t touch the weights until necessary — X (formerly Twitter)

  • Why read: Why you should focus on agent infrastructure before trying to fine-tune models.
  • Summary: People argue over open vs. closed models, but base models are usually smart enough. The real bottleneck is the surrounding infrastructure. Teams building agents should start with top closed models to get a baseline. They should spend their time on orchestration, tool integration, and evaluations. Only after fixing routing, context, and debugging should they swap in open models to save money. Fine-tuning weights is a late-stage optimization most companies can skip. AI success comes from reliable systems and context, not just model capability.
  • Read more

14. How to properly build a credit ledger — X (formerly Twitter)

  • Why read: How to build the low-latency billing systems needed for token-based AI apps.
  • Summary: Standard asynchronous billing doesn't work for AI apps. You need real-time quota checks before tokens are spent. A good credit ledger handles business logic like promo grants and expirations while making fast access decisions. To do this, separate the main database state from an in-memory cache for fast reads. This setup tracks usage accurately and handles the high concurrency of agent interactions. Teams have to build this specialized infrastructure to monetize without slowing down the app.
  • Read more

15. Why are some AI companies hitting $10M ARR in six months? — X (formerly Twitter)

  • Why read: Why AI infrastructure startups are hitting eight-figure revenues so fast.
  • Summary: A new group of AI startups is hitting $10M ARR in months by selling infrastructure directly to top AI labs, not traditional enterprises. These labs scale so fast they hit memory, inference, and security bottlenecks before anyone else. Since they are racing for capability, they pay a premium for anything that saves time and compute. Startups solving these hyperscale problems get big contracts because their ROI is immediate and obvious. Go-to-market here means finding and fixing the sharpest technical pain points at the frontier.
  • Read more

Themes from yesterday

  • Infrastructure over models: Engineers are shifting focus from base models to agent harnesses, decentralized sandboxes, and low-latency billing.
  • Agents taking over work: Companies are moving from chat interfaces to asynchronous "mission pods" and personalized AI assistants.
  • New AI moats: As model prices converge, the best competitive advantages are specialized data loops and fixing hyperscale bottlenecks for AI labs.