1. How I use LLMs as a staff engineer in 2026 — seangoedecke.com

  • Why read: How staff engineers actually work using agentic AI.
  • Summary: Engineers have moved past smart autocomplete to handing off entire PRs and bug hunts. With cross-repo access, agents diagnose 80% of issues independently. The staff engineer's job is now skimming, rejecting bad agent attempts, and deeply reviewing the viable ones. Human time goes into gathering context, building mental models, and setting up reproductions while the agent searches in parallel.
  • Read more

2. The GenAI economy has generated $110 billion in sales over the past 12 months — Azeem Azhar

  • Why read: A bottom-up look at AI spending across consumer and enterprise markets.
  • Summary: The GenAI economy hit a $175 billion annualized run rate, growing 3x faster than past IT cycles. But with $2 trillion in CapEx committed through 2026, current revenue only covers equipment depreciation. Monthly token volumes exceed 30 trillion (up 14x year-over-year), yet the industry still hasn't figured out how to price this intelligence. Value is moving up the stack from infrastructure to models and apps.
  • Read more

3. We Added Too Many Guardrails and Broke Our Own Agent — SaaStr

  • Why read: What happens when you over-constrain production AI agents.
  • Summary: Teams deploying AI agents often over-correct with strict guardrails to stop hallucinations and bad interactions. But too many constraints break the agent's ability to act or handle edge cases. A team running 21+ agents found that excessive safety measures dragged down revenue and operations. Stripping away unneeded restrictions and fixing configuration errors led to a massive revenue recovery. Agents need autonomy to work.
  • Read more

4. OpenAI Codex lead on the new shape of product work — Lenny's Newsletter

  • Why read: How OpenAI builds products and why product taste beats technical execution.
  • Summary: AI has dropped the cost of building software to near zero. At OpenAI, almost every employee uses Codex to build tools, collapsing traditional job titles. When anyone can build anything, teams need to run a "zone defense" instead of sticking to strict roles. Knowing what to build—having refined product taste—matters more than the mechanics of building it.
  • Read more

5. TIME TO RETIRE THE CLASSIC PRD — Gokul Rajaram

  • Why read: Why traditional Product Requirements Documents fail in the agent era and what to use instead.
  • Summary: The classic 10-page PRD was built to align a room full of humans. Today, the bottleneck is giving an AI agent or engineer a spec tight enough to ship without endless follow-ups. A modern product spec should be one page: a falsifiable bet, a clear problem statement, concrete acceptance criteria, and a measurement plan. This format lets teams and agents ship faster with clear grading rubrics.
  • Read more

6. Training Physical AI on Video Game Clips — Chamath Palihapitiya

  • Why read: Training robotic foundation models on video game footage.
  • Summary: Standard video data teaches AI what the world looks like. Robots need to learn cause and effect: how their actions change the environment. General Intuition raised $320M to train action models using gameplay clips, pairing visual data with exact button presses and immediate feedback. A model trained on 100 hours of Fortnite gameplay successfully controlled a physical robot navigating an office.
  • Read more

7. Intelligence wants to be free — Christian Catalini

  • Why read: The fight between centralized lab monopolies and distributed AI networks.
  • Summary: The AI industry is split between government-approved lab monopolies and market-driven, distributed intelligence. Relying on a few large labs to gatekeep frontier models breaks free-market principles. Decentralized networks won't win by copying incumbents—they can't hoard context or build massive surveillance systems. Open networks will win by using the unmeasured, unique experiences of human participants and distributed contributions.
  • Read more

8. What is talent and talent engineering — Rich

  • Why read: Defining two new roles emerging at frontier tech labs and top startups.
  • Summary: Traditional pipeline recruiting is dead for top companies. Today, one exceptional hire sets the bar for the next ten. The Talent role now focuses on relationship-building and attracting the top 1% of operators rather than processing applicants. Talent Engineering is a rare new profile: a top-tier product engineer who prefers building recruiting tools over core products. Hiring these engineers provides a clear advantage in building talent-dense teams.
  • Read more

9. Lots of buzz recently on compute capital markets — Jay Yu

  • Why read: How the financial markets for AI compute will actually work.
  • Summary: Compute acts more like electricity than fungible commodities like gold. There is no standard pricing for GPU access. Since reservations are tied to specific SKUs, times, and models, capacity forwards will trade as bilateral over-the-counter (OTC) agreements. Providers (neoclouds) and consumers (inference platforms, agentic apps) will transact through OTC dealers. Dealers will hedge their exposure on generalized basket exchanges, forming the backbone of the new compute capital markets.
  • Read more

10. As engineering, product, design, DS, etc — Boris Cherny

  • Why read: How traditional job titles are dissolving into five product lifecycle archetypes.
  • Summary: In advanced AI teams, domain-specific roles are fading into five cross-functional archetypes: Prototyper, Builder, Sweeper, Grower, and Maintainer. Team makeup shifts based on product-market fit. Pre-PMF products need Prototypers and Builders; mature products need Growers and Maintainers. Individuals often cover two or three archetypes regardless of their official title. This model is a better way to structure modern product teams than legacy silos.
  • Read more

11. Using Local Coding Agents — Sebastian Raschka, PhD

  • Why read: Why and how to move from proprietary cloud APIs to local, open-weight coding agents.
  • Summary: Proprietary services like Claude Code dominate, but local coding harnesses offer a transparent and controllable alternative. Local stacks protect against API price hikes, keep data private, and ensure reproducibility when proprietary models change. Open-weight models are now capable enough to drive local development. Hooking these models into local harnesses that read files, edit code, and run commands creates a strong, customizable daily driver.
  • Read more

12. TPU Deep Dive — henryhmko.github.io

  • Why read: How Google's TPUs scale and out-throughput NVIDIA GPUs.
  • Summary: GPUs use tens of thousands of small cores and large caches. Google's TPUs optimize for matrix multiplication throughput using systolic arrays and larger on-chip memory. TPUs hit 42.5 ExaFLOPS per pod by co-designing hardware and the XLA compiler. The architecture relies on pipelining and Ahead-of-Time compilation, betting that deep learning math maps cleanly to systolic arrays. This explains Google's infrastructure advantage in AI training and inference.
  • Read more

13. 10 lessons from working at startups — nic

  • Why read: Practical advice for engineers and operators at startups.
  • Summary: Startup time is scarce, making prioritization essential. Trivial features like dark mode add friction with no upside. Product decisions should follow data. If users aren't complaining about a UX flow, don't rewrite it for engineering perfectionism. Treat early customers as partners, as they hold most of your credibility and revenue. Finally, treat marketing and distribution as engineering problems; they compound over time and often decide a product's success.
  • Read more

14. I Built a Cheap Channel Finder for Any Market — Jordan Crawford

  • Why read: Using AI to find cheap, high-converting marketing channels based on buyer behavior.
  • Summary: Standard marketing playbooks push the same few channels like SEO and LinkedIn. Real go-to-market success requires finding uncrowded places your buyers already frequent. AI can synthesize market case studies and sales call transcripts to spot association networks and local reputation loops that competitors miss. Starting with a buyer's required routines uncovers specific, cheap places to run tests. This approach provides a ranked roadmap instead of guesswork.
  • Read more

15. Your Job as CEO is to Become a Coach — Regina Gerbeaux

  • Why read: Why founders must shift from player-managers to executive coaches to scale their teams.
  • Summary: CEOs often confuse being "hands-off" with empowering their direct reports. To the team, it just looks like absence. Real empowerment means explicitly communicating trust while coaching executives through hard problems, like firing underperformers. Without active coaching, leaders burn out and direct reports struggle with their own management blindspots. Becoming an executive coach is a necessary evolution of the CEO role.
  • Read more

Themes from yesterday

  • The collapse of traditional roles: AI lowers the cost of software creation, blurring rigid titles like PM, Engineer, and Designer into hybrid builder-operator archetypes.
  • Hardware and compute markets: The industry is shifting toward specialized inference chips like TPUs and forming over-the-counter capital markets to trade compute.
  • New GTM and product playbooks: The agent era replaces long PRDs with one-page specs. Growth relies on AI-driven channel discovery rather than standard marketing channels.
  • The push for local and open AI: Resistance to centralized API monopolies is driving developers toward local coding harnesses and decentralized intelligence networks.