1. Own the Outer Loop — substack.com

  • Why read: Engineers must shift from writing code to managing AI agents and taking responsibility for the final product.
  • Summary: AI agents are taking over the "inner loop" of writing and testing code. Engineers must now own the "outer loop" by verifying the work. We can't trust autonomous agents on their own; we need humans to judge the output, set up strict checks, and make the final call on what goes to production. Companies won't use AI until humans are clearly accountable for what changes and why it's safe.
  • Read more

2. The architect's guide to harness engineering: How to choose or build the one for you. — X (formerly Twitter)

  • Why read: A guide on when to buy or build an AI harness to get the most out of LLMs for the least cost.
  • Summary: The next big AI opportunity is optimizing models with the right "harness"—the surrounding code, tools, and constraints. Non-engineers should buy existing AI apps and focus on providing good context. Technical builders, however, should build specialized harnesses that use smaller or mixed models. Encoding domain-specific structures and removing unused tools cuts token costs and improves reliability. These custom setups will eventually evolve through reinforcement learning and fine-tuning.
  • Read more

3. Building against the big labs that are trying to eat you — X (formerly Twitter)

  • Why read: How AI startups can beat foundation model giants by optimizing domain-specific harnesses.
  • Summary: Startups can beat general-purpose tools like Claude Code by refining their agent harnesses for specific use cases. Large labs ship bloated tools for millions of users, inflating context size and token costs. Specialized agents cut this fat and focus on a few key features, running much faster and cheaper. Independent builders also stay model-agnostic, routing tasks to the most efficient model available. This lightweight approach helps startups win on accuracy and unit economics.
  • Read more

4. Making Fable Cheaper Than Opus — X (formerly Twitter)

  • Why read: Using a more expensive model to manage agents can actually reduce total workflow costs.
  • Summary: Frontier models like Fable cost more per token than older ones like Opus, but they can be cheaper to run when paired with a sidekick model. Capable models manage context and delegate well, taking fewer turns and writing fewer tokens. In coding evaluations, a Fable-led agent delegated early instead of reading or editing files directly, slashing its token usage. Weaker models micromanage and drag heavy context through unnecessary turns, which drives up costs. Agent workflow costs depend on turn count and delegation, not just the lead model's price per token.
  • Read more

5. We Cant Afford Your Inference — X (formerly Twitter)

  • Why read: The hidden costs of running multi-agent workflows and reasoning models for daily development.
  • Summary: Relying on frontier reasoning models and autonomous subagents every day can push a developer's API bill into the thousands per week. Features that spawn subagents or run long autonomous loops consume massive amounts of tokens. The models work well, but paying for every internal reasoning step or failed attempt breaks the economics. Power users are turning to cheaper, faster alternatives like Grok or smaller models. Software teams need to weigh the marginal gains of full autonomy against the steep inference costs.
  • Read more

6. The case against training your own models (for now) — X (formerly Twitter)

  • Why read: Why you should exhaust prompt engineering and orchestration before training custom models.
  • Summary: The fear of data leaking to model providers is a procurement problem. Fix it with zero-retention agreements, not by training custom models. For most teams, the best return on investment comes from capturing tribal knowledge, improving context, and optimizing agent orchestration. Training or fine-tuning only makes sense when you hit hard limits on capabilities, behavioral rules, or context costs at scale. Treat training as a last resort when cheaper methods fail.
  • Read more

7. I Paid for These Reasoning Tokens, So Why Can’t I See Them? — X (formerly Twitter)

  • Why read: Why you can't see the hidden "scratchpads" of reasoning models, and why the UI narrations are mostly theater.
  • Summary: Advanced reasoning models use thousands of hidden tokens to break down problems. These tokens aren't a clean transcript; they are messy, fragmented, and tied to the model's internal activations. Because the raw scratchpad is unreadable, products generate fake "reasoning narrations" to show progress. This narration is a post-hoc explanation and may not reflect how the model actually made its decision. You are paying for the improved final output, not for an audit trail of the model's thoughts.
  • Read more

8. GPT-5.6, Fable 5, and the AI Coding Quota War — X (formerly Twitter)

  • Why read: How the recent launches of GPT-5.6 and Fable 5 exposed capacity limits at OpenAI and Anthropic.
  • Summary: The concurrent launches of GPT-5.6 and Claude Fable 5 exposed capacity issues at both labs. OpenAI struggled with quota drains from high context limits and multi-agent loops, forcing them to reset usage and roll back features. Anthropic faced export controls and tightly rationed Fable 5 access, adjusting it weekly to keep users from defecting to OpenAI. The AI race is now as much about managing infrastructure and pricing as it is about raw capability. Developers must navigate shifting rate limits and unpredictable costs.
  • Read more

9. The Treadmill: Why Frontier AI is Starting to Rhyme with Semiconductors — X (formerly Twitter)

  • Why read: Why the capital-intensive economics of frontier AI labs mirror the semiconductor industry.
  • Summary: Frontier AI development requires massive capital expenditures that depreciate instantly when a competitor ships a better product. A $15 billion chip fab amortizes over years, but a $500 million training run loses its premium value the moment a better model drops. Training costs are growing 2.4x per year, pushing future runs into the billions. Labs must raise massive amounts of capital just to stay relevant, with no guarantee of a moat. If the semiconductor industry is any guide, this treadmill will force brutal consolidation.
  • Read more

10. Kevin Deierling (Nvidia, SVP Networking): The Data Center Is Now a Factory That Prints Tokens — X (formerly Twitter)

  • Why read: Data centers are shifting from IT cost centers to AI factories built solely to produce tokens.
  • Summary: The traditional data center is dead. The new model is the "AI factory," a facility built to mint tokens as efficiently as possible. Because Moore's Law has stalled on power and cost, the entire stack—from power plants to networks—must be co-designed. Any watt or dollar that doesn't produce a token is wasted. This shift means AI infrastructure needs a complete rebuild, not just incremental upgrades.
  • Read more

11. My biggest takeaways from @noamseg, diving into the results of... — X (formerly Twitter)

  • Why read: Survey data shows AI is causing an epidemic of burnout among tech workers despite massive productivity gains.
  • Summary: A survey of tech workers shows their biggest fear isn't losing their jobs; it's being squeezed to produce more for the same pay. AI has made 97% of respondents faster, but management instantly absorbed these gains into higher expectations, driving a 10-point surge in burnout. The data also points to "cognitive rot": workers blindly accepting AI outputs and letting their critical thinking atrophy. The workforce is splitting between those amplified by AI and those destabilized by it, with designers and researchers hit hardest. Good management is the only protection against this productivity squeeze.
  • Read more

12. ARMCF: Introducing A Practitioner Control Framework for AI and Agentic Risk Management — Substack

  • Why read: A new governance framework tailored for the specific risks of autonomous AI agents.
  • Summary: Companies are adopting AI faster than they can secure it, relying on vague principles instead of technical enforcement. The ARMCF framework maps the standard NIST cybersecurity lifecycle to AI-specific risks. It focuses on material threats: what data the AI can access, what tools it can use, and what actions it can take on its own. Security teams can use this to assess blast radiuses, track shadow AI, and recover from incidents. It helps leaders balance rapid adoption with hard security controls.
  • Read more

13. The Bad Apple Problem — Ghost - The Professional Publishing Platform

  • Why read: The legal and strategic fallout of IP theft and employee poaching in AI hardware.
  • Summary: Apple is suing former employees for allegedly stealing trade secrets for OpenAI. The legal filings describe ex-employees using Apple devices to exploit network bugs and coach others on leaking data. This poaching strategy highlights the intense pressure AI companies face to build hardware quickly. But hiring people with a history of corporate espionage introduces massive legal and cultural risks. It forces companies to decide how much bad behavior they will tolerate in the race for AI dominance.
  • Read more

14. Latent Space as a New Medium — KK

  • Why read: Why we should view Large Language Models as highly compressed, explorable maps of human knowledge, rather than just chatbots.
  • Summary: Large Language Models compress human knowledge into a small "latent space." They don't store copies of texts or images; they store mathematical relationships between concepts. This dense space is a new medium for human creativity. Artists and scientists can navigate it to find novel combinations and generate new works. Viewing latent space as an explorable medium, rather than a smart answering machine, reveals the real potential of AI.
  • Read more

15. How to never get writer's block with ChatGPT Voice and Codex — X (formerly Twitter)

  • Why read: A workflow for using voice AI and custom agents to accelerate writing.
  • Summary: Typing and editing often ruin creative flow. You can beat writer's block by speaking your ideas to ChatGPT's voice mode, acting as a silent ghostwriter to produce a transcript. Feed this transcript to Codex with strict instructions to draft a piece using only your spoken words. This preserves your actual voice and avoids generic AI writing. You can then use other agents to format, edit, or generate video timestamps. This setup turns raw thoughts into finished content without the mechanical drag of writing.
  • Read more

Themes from yesterday

  • Agent architecture: Builders are shifting from chasing the largest base models to optimizing domain-specific harnesses that handle context, delegation, and costs.
  • AI economics: High inference costs and steep capital requirements mirror the semiconductor industry. Operators are mixing models and trimming context to cut their token bills.
  • Burnout and cognitive rot: AI makes workers faster, but management absorbs the gains into higher expectations. This causes burnout and lets workers' critical thinking atrophy as they blindly accept AI outputs.
  • Governance and accountability: As agents act autonomously, security teams are pushing for harder frameworks. Engineers must now verify outputs and take responsibility for what ships, rather than trusting AI blindly.