1. The first experimental evidence of recursive self-improvement (RSI) — X (formerly Twitter)

  • Why read: Early proof that AI agents can iterate on their own code, suggesting capabilities will compound faster than expected.
  • Summary: Researchers built AIDE², an AI that improves itself through two loops: an inner loop for code and an outer loop for the agent's harness. Over eight days, the system found seven algorithmic upgrades, such as memory compression and defenses against reward hacking. These upgrades worked on unseen benchmarks and beat a baseline humans spent two years tuning. This is a milestone in recursive self-improvement, showing automated R&D can beat manual tuning. For developers, the job is shifting from building agent architectures to designing meta-optimizers.
  • Read more

2. The Economics of AI Reasoning — X (formerly Twitter)

  • Why read: A breakdown of when to pay for reasoning models and when they waste tokens on simple tasks.
  • Summary: Models like o1 and DeepSeek-R1 popularized "test-time compute," letting models think before answering. Uncapped reasoning improves complex coding scores by 20%, but costs up to 10x more in tokens and runs slower. Data shows reasoning adds no value and only slows down nearly half of routine tasks, like fetching files or checking email. Operators should treat reasoning as a toggle to balance cost and speed, not a default setting. Future systems will rely on routers to assign a thinking budget based on the prompt.
  • Read more

3. The Best Model Routing is Task Specific — X (formerly Twitter)

  • Why read: Why generalized model routers are losing to task-specific routing built for cost and speed.
  • Summary: Teams are adopting model routing to avoid paying frontier-model prices for every token. Generalized routers work well for open-ended queries but lack the context for business workflows. Top applications now break requests into sub-tasks and route them dynamically. Simple text extraction goes to cheap models, while complex table parsing goes to specialized vision models. Setting this up requires mapping which model passes the quality bar for each step. General intelligence is overkill once a workflow is broken down into specific tasks.
  • Read more

4. A Swarm of Agents for Multi-Angle Analysis: Building a Team of Experts from LLMs — X (formerly Twitter)

  • Why read: A technical pattern for using multiple specialized LLMs to avoid groupthink and improve decisions.
  • Summary: Single LLMs often default to safe, average advice. A swarm architecture fixes this by creating multiple agents with conflicting biases, like a ruthless investor and a paranoid security engineer. These agents analyze the same problem in parallel without seeing each other's work. A synthesizer agent then merges the arguments to flag strong agreements and deep contradictions. This forces structured disagreement, stress-testing high-stakes decisions instead of smoothing them out.
  • Read more

5. Claude Code: Good skills, bad skills — Fin /ideas

  • Why read: A guide to building effective skills for coding agents, based on lessons from an internal plugin marketplace.
  • Summary: The Fin team built a repository of over 260 Claude Code skills, allowing automated tasks to be shared across the company. The best skills do one thing well, use scripts for predictable actions, and expose info progressively to save tokens. Bad skills try to do too much, lack owners, rely on outdated knowledge, or just make dashboards instead of doing the work. Good skills also avoid assuming developers have specific proprietary tools and include guardrails against hallucinated workflows. Engineering teams should focus on small, tested skills with tight feedback loops.
  • Read more

6. Building a Self-Improving, AI-Native Company — X (formerly Twitter)

  • Why read: How to move beyond basic AI wrappers and build software that tests and patches itself.
  • Summary: Deel realized faster AI copilots didn't solve the core bottleneck of software maintenance. They built a closed loop with two agents: Bug Hunter and Deel Code. Bug Hunter tests production replicas for edge cases and UI bugs that standard tests miss. Once it finds a bug, Deel Code diagnoses it, writes a fix, and proves it works. Humans only step in at the end for code review and deployment. This shows the potential of AI-native software: systems that find and fix their own bugs.
  • Read more

7. How to build a self-improving code review agent — X (formerly Twitter)

  • Why read: A blueprint for adding an autonomous, learning code reviewer to your CI/CD pipeline.
  • Summary: A custom code review agent beats out-of-the-box platforms because it shares context with your existing tools. You start by defining a code review skill that tells the agent how to leave feedback, rate severity, and output JSON. Instead of loading the agent with repo rules upfront, use an outer-loop agent that watches the reviewer and updates its instructions based on human corrections. A full agent can check the codebase, test ideas, and verify that suggested code builds before commenting. Using Python scripts alongside the skill cuts token use and stops hallucinations.
  • Read more

8. RL Post-Training on Macs — pluralis.ai

  • Why read: How a distributed setup lowers the cost and hardware barriers for reinforcement learning post-training.
  • Summary: Researchers ran multi-turn RL post-training for an 8.3B-parameter model using 14 consumer Macs worldwide and one B200 GPU in a datacenter. Since RL spends most of its time generating rollouts, Apple Silicon's unified memory is a cheap way to hold the model and large KV caches. They bypassed latency and bad internet by building an asynchronous system where workers stream data to the central trainer. They used custom techniques to stabilize training despite hardware and precision differences. This setup doubled accuracy on held-out tasks, showing decentralized hardware can handle advanced post-training.
  • Read more

9. Context engineering with Dex Horthy — Substack

  • Why read: Why engineers are moving from generic AI frameworks to custom context pipelines for software development.
  • Summary: Dex Horthy notes that AI engineers are dropping frameworks like LangChain for custom, deterministic pipelines. A major lesson: letting models write unreviewed code creates massive technical debt. Humans can't debug systems they don't understand. Models built for benchmarks like SWE-bench are good at fixing bugs but make bad architectural choices that degrade code over time. Good context engineering means curating exactly what the model sees, giving it precise scope without useless history. As code generation speeds up, strict tooling and human reviews are necessary to prevent software rot.
  • Read more

10. My team spends all day talking AI with enterprise execs — X (formerly Twitter)

  • Why read: A look at the tactical questions and anxieties stalling enterprise AI adoption.
  • Summary: Enterprise leaders are past asking whether to use AI. Now they are stuck on deployment and governance. Their main problems: securely exposing fragmented data, measuring the ROI of custom agents, and testing non-deterministic features. Executives want to let employees build tools while controlling token costs and protecting core systems. They also worry about vendor lock-in and how to transition non-technical staff to new workflows. Anyone selling to or leading enterprise teams needs to understand these friction points.
  • Read more

11. How to measure the impact of AI search the right way — beehiiv.com

  • Why read: Why traditional SEO traffic metrics are obsolete in the era of AI search.
  • Summary: Marketers struggle to track visibility as buyers switch to AI chatbots. Users rarely click citations, and when they do, 70% of that traffic shows up as untraceable "Direct" visits, breaking last-click attribution. Chasing raw citation counts is also a mistake, as models hallucinate or swap domains frequently. The fix is dropping traffic metrics and linking leading indicators, like where you rank in LLM answers, to business outcomes. Tools that track brand placement in AI-generated shortlists are becoming standard.
  • Read more

12. Every AI Company Is Failing at the Same Thing: Do This Instead. — beehiiv.com

  • Why read: Why early revenue for AI startups often masks a lack of real product-market fit.
  • Summary: Many AI startups think they have product-market fit because they close pilots and generate revenue. In reality, they have "problem-market fit." Buyers are desperate for AI and will meet with anyone in the space. Revenue from custom, unrepeatable deals across different customer profiles does not equal a scalable business. Real product-market fit means your product solves the problem better than alternatives and sells through a highly repeatable process. Founders should wait for a repeatable conversion engine before hiring more sales staff.
  • Read more

13. Builder-Executives Are Getting Paid Like Pro Athletes — Substack

  • Why read: Why technical product leaders who can build with modern AI tools are commanding massive salaries.
  • Summary: Pay for "builder-executives" is detaching from standard market rates. Product leaders with executive strategy skills and the technical chops to build with modern AI are seeing offers up to $10 million a year. This scarcity is dragging up compensation across the industry, pushing leaders to stay technical rather than chase immediate raises. Operators aiming for this tier should stop worrying about incremental title bumps and focus on staying hands-on.
  • Read more

14. Free GTM Data You've Never Heard Of — Substack

  • Why read: How to find high-intent B2B lead lists in public government filings.
  • Summary: Go-to-market teams spend heavily on data brokers while ignoring free government datasets. NYC housing records force LLCs to list corporate officers, giving proptech companies direct landlord names. Department of Labor Form 5500 filings force companies to list their benefits vendors, handing you a competitor's customer list. SEC Form D filings for unannounced raises include executive contact info long before the news breaks. Operators can scrape and filter these filings to build targeted outbound pipelines.
  • Read more

15. Absolutely beautiful rant about AI in Linux Kernel from Linus... — X (formerly Twitter)

  • Why read: Linus Torvalds settles the debate on using AI in open-source development.
  • Summary: Linus Torvalds stated the Linux Kernel project will use AI and will not tolerate performative anti-AI sentiment. He noted AI's utility is not up for debate, and ignoring it means refusing to adapt. He acknowledged AI generates bugs and overhead, but said the fix is better tooling, not ignoring the technology. He rejected ideological arguments against AI, stating open-source decisions run on technical merit. This signals a green light for integrating AI into legacy and foundational projects.
  • Read more

Themes from yesterday

  • Specific architectures over general models: From task-specific routing to multi-agent swarms and Claude Code skills, teams are dropping monolithic prompts in favor of modular systems.
  • Autonomous loops: Systems are starting to improve themselves without human intervention, seen in AIDE²'s R&D and Deel's bug-hunting pipelines.
  • Enterprise reality check: Adoption is stalling on governance and data fragmentation. Startups are realizing early revenue often signals buyer desperation rather than repeatable product-market fit.
  • New playbooks for careers and growth: The rise of the technical "builder-executive" and the death of traditional SEO metrics show how AI is rewriting the rules for operations and hiring.