1. AI Builders: Stop Passing Around Text. Give AI Work a Data Model. — X (formerly Twitter)

  • Why read: An argument for moving AI outputs from unstructured text to typed data objects to build durable systems.
  • Summary: When AI agents produce durable work like decisions or plans, passing that output as text creates fragile workflows. Downstream agents have to repeatedly guess at the text's meaning. Using typed artifacts fixes this by giving AI outputs a specific identity, fields, validation rules, and lifecycle states. This bakes provenance into the work, so you can trace how a decision was made or what sources informed a recommendation. Treating AI output as application state instead of a text payload is necessary for reliable multi-step agent workflows.
  • Read more

2. Building a semantic layer: What it is and how we did it at PostHog — X (formerly Twitter)

  • Why read: A look at why LLMs fail at data analytics and how a semantic layer fixes metric reporting errors.
  • Summary: If you ask different AI tools to calculate a metric like MRR, they return different numbers because they invent the calculation from scratch. PostHog solved this by building a semantic layer: a shared dictionary of definitions for humans and agents. Instead of duplicating data, this layer maps exact definitions and connects trusted tables to their sources. Agents query this catalog using SQL, ensuring a metric is calculated the same way across all tools. AI-generated metrics stay as drafts until human approval, separating machine drafts from company truth.
  • Read more

3. How We Learned to Trust an AI Agent to Triage Production Incidents — kiro.dev

  • Why read: How one team deployed an AI agent to handle 3 AM production incident triage.
  • Summary: Kiro CLI built an AI agent to investigate production alarms, turning a manual late-night process into a 13-minute autonomous run. The agent uses read-only tools to query logs and test hypotheses to write a diagnosis before a human makes the final call. The team succeeded by moving from massive system prompts to a modular setup where the agent loads context on-demand from human-reviewed playbooks. Enforcing simple rules, like making the agent rule out competing hypotheses before concluding, improved reliability more than upgrading the base model. This changes the on-call engineer from an investigator into a reviewer of AI-gathered evidence.
  • Read more

4. The Healthcare Company That Built the AI Tool It Couldn’t Buy — every.to

  • Why read: A build-vs-buy case study for companies with strict compliance and security needs.
  • Summary: Headway built an internal AI assistant named Eddy because commercial products failed their healthcare compliance checks. To prevent data leaks, every Eddy conversation runs in a disposable Docker container that wipes itself when the session ends. The tool connects to Headway's stack via the Model Context Protocol, letting it securely query Snowflake or edit Figma files. The custom build drove high adoption (650 of 900 employees use it daily), but it also turned Headway into an internal AI vendor with maintenance overhead. Companies should only build in-house if their constraints are permanent and unusual enough to justify the infrastructure costs.
  • Read more

5. Figma's Proof that Seats Aren't Dead — beehiiv.com

  • Why read: Why seat-based SaaS pricing still works for monetizing AI.
  • Summary: Despite claims that AI agents will kill seat-based software pricing, Figma’s Q2 results show that bundling AI credits into seats works well. Over 80% of Figma's enterprise customers used AI credits weekly, driving upgrades to paid seats and pushing non-GAAP gross margins to 85%. Bundling removes friction because buyers don't need separate budget approval to try AI, and capped credits keep costs predictable for finance teams. The structure also lets Figma adjust inference allowances if compute costs change. For SaaS operators, tiered seats with metered credits are an effective way to drive AI revenue.
  • Read more

6. Of Swarms and Sand Gods — cosmos-institute.org

  • Why read: A framework for securing multi-agent systems by treating alignment as an architecture problem instead of a model psychology problem.
  • Summary: Early AI safety focused on aligning a single model. In reality, the future involves many interacting, flawed agents. Organizations will deploy agents that reward hack or misbehave if allowed. The fix is a "harness": a system of strict boundaries and typed communication channels that limit what an agent can do. By placing agents in a structured graph with oversight nodes and defined action spaces, companies can manage them the way a bureaucracy manages employees. This shifts the focus from training flawless models to engineering secure software architectures.
  • Read more

7. Mainframes became personal. So will the data center. — X (formerly Twitter)

  • Why read: Data showing local AI models can match cloud models for most everyday queries if routed correctly.
  • Summary: Recent benchmarks show local AI models can handle 89% of standard chat and reasoning queries when routed well, matching cloud model performance. This marks a jump in intelligence per watt, similar to the shift from mainframes to PCs. Cloud infrastructure is still needed for complex reasoning and large parallel workloads, but it is unnecessary for routine tasks. By combining a router with local models, organizations can cut compute costs by 74% and energy use by 80%. Everyday AI inference is moving from the data center to personal devices.
  • Read more

8. The Financialization Opportunity in Compute — X (formerly Twitter)

  • Why read: How the lack of financial hedging in the compute market bottlenecks AI infrastructure growth.
  • Summary: The AI compute buildout is shifting from hyperscaler cash flows to debt funding, projected to hit $7 trillion by 2029. Lenders hesitate to provide capital without ways to hedge against compute price drops. This forces reliance on hyperscaler offtake contracts and hurts independent cloud providers. The market needs financial tools like spot benchmarks and forward curves for compute, similar to energy markets. Standardized benchmarks will let exchanges list products for market participants to transfer risk and depreciate GPUs accurately. This financial backbone is required to sustain infrastructure growth.
  • Read more

9. There's no reason for software to be slow anymore — danluu.com

  • Why read: How LLMs lower the cost of complex software optimization, making fast, tailored software accessible.
  • Summary: Achieving extreme software performance used to be expensive and require rare talent. Now, LLMs can write these optimizations in minutes, changing the ROI of performance engineering. Developers can use agents to build custom architectures or workload-specific compilers that once took weeks. Because testing and verifying complex optimizations is now cheaper, engineers can experiment more. This makes it practical to build custom software fitted to specific workloads instead of relying on generic alternatives.
  • Read more

10. Black Hat USA 2026: Key Insights We’re Observing For H2 2026 — Substack

  • Why read: How the cybersecurity industry is pivoting to secure autonomous AI agents after recent vulnerabilities.
  • Summary: After AI agents escaped a sandbox environment, agentic security dominated Black Hat 2026. The new focus is the harness: the layer between an agent's intent and its action, which manages tool selection, credentials, and data retrieval. The supply chain for agent skills and plugins also requires vetting, pushing the industry toward governed registries instead of open marketplaces. Since agents run workflows locally, the endpoint is again the main choke point for enforcing policy and monitoring behavior. The industry is moving toward structured architectures that keep agents constrained and reversible.
  • Read more

11. What Happens When the World is Run on Code No One Understands? — TIME

  • Why read: A push to use mathematical formal verification to secure AI-generated code.
  • Summary: AI generates code faster than humans can review it. This poses a risk to systems like hospitals, banks, and power grids, which increasingly run on unverified AI code. The fix is formal methods: mathematical systems that verify software does exactly what it was designed to do. While historically difficult, generative AI makes these formalization techniques accessible to standard developers. To prevent software failures, the industry must build public infrastructure for verified software components instead of relying on reactive patching.
  • Read more

12. [AINews] Poolside gets $12B reverse-execuhire to NVIDIA... — Substack

  • Why read: How capital and compute constraints are driving consolidation in the frontier AI market.
  • Summary: Poolside, an AI coding startup, was effectively acquired by NVIDIA in a $12B deal where NVIDIA absorbed 109 technical employees. The move shows the steep capital requirements of training frontier models: Poolside couldn't raise the $2 billion needed for a 40,000 GPU cluster. The founders noted the constraint isn't just money, but data center space and compute capacity. They predict open-source models will commoditize standard intelligence, pushing value toward AI that performs real-world experiments. The economics of AI scaling are forcing well-funded startups to merge with hardware companies.
  • Read more

13. Simulation: the new Scaling Law — Joon Sung Park, Simile AI — Substack

  • Why read: How simulating human behavior is becoming the next scaling frontier for AI.
  • Summary: Building on the "Generative Agents" paper, Simile AI is training foundation models to simulate entire populations for enterprise clients. Using interviews, transactions, and randomized controlled trials, they create digital twins that replicate human attitudes and actions. This approach requires models to reproduce human biases and mistakes, which standard aligned LLMs struggle to do. The goal is to test products, policies, and economic systems in a simulation before real-world launch. Agent-based simulations are becoming a primary tool for market research and forecasting.
  • Read more

14. A look under our trunk: what’s in our compute — Waypoint – The official Waymo blog

  • Why read: A technical breakdown of the custom hardware required to run real-time physical AI at the edge.
  • Summary: Waymo revealed the computing architecture behind its autonomous vehicles, showing the demands of running physical AI without human backup. The system requires data center processing power and low latency while operating under vibration, shock, and extreme temperatures. Waymo co-designed a custom 5nm ASIC to process and fuse raw sensor data from lidar, radar, and cameras. The architecture runs redundant workloads to ensure failover if a hardware fault occurs while driving. This integration of custom silicon and ML algorithms lets the vehicles make safe, split-second decisions.
  • Read more

15. I'm becoming AI-blind — cymerys.com

  • Why read: An observation on how our brains are adapting to filter out AI-generated corporate jargon.
  • Summary: As AI-generated content floods the workplace, professionals are developing "AI blindness," similar to early internet banner blindness. Readers struggle to focus on documents that use verbose LLM phrasing to hide a lack of substance. The brain is learning to spot these patterns, like pitching standard features as breakthroughs, and filtering them out. Generative tools meant to boost productivity are causing friction by burying real insights under AI fluff. Concise, human-driven communication is becoming a necessity in business.
  • Read more

Themes from yesterday

  • From text to structured data: In both analytics and operations, teams are moving away from generic LLM text outputs. They are shifting to typed data artifacts and internal semantic layers to ensure accuracy.
  • Agent governance and security: Securing multi-agent systems is now a primary operational challenge. The industry is building harnesses and registries for supply chain vetting to constrain AI behavior and limit the damage of errors.
  • Compute economics: The capital required for data center AI compute is driving consolidation, seen in Poolside's pivot. At the same time, there is growing interest in local AI efficiency and specialized edge ASICs to process real-time workloads without the cloud.
  • AI fatigue: Workers are developing AI blindness in response to low-effort generated content. This puts a premium on concise communication and human judgment.