1. DIscussing tokenomics with Gemini — X (formerly Twitter)

  • Why read: How Chinese models cut agent workflow costs through architecture changes instead of just distillation.
  • Summary: Multi-agent frameworks and coding tools are replacing standard chat and will drive enterprise AI billing. Western models use GQA, requiring large uncompressed KV cache storage that makes long context windows expensive. Chinese models use Multi-head Latent Attention (MLA) to compress the cache into a dense vector. This turns cache hits into cheap distributed storage reads, cutting context costs by over 99%. Factor these design differences into your infrastructure and API pricing plans.
  • Read more

2. I dont know how to explain this to yall without... — X (formerly Twitter)

  • Why read: The case for building REPL-first applications instead of relying on the Model Context Protocol (MCP).
  • Summary: For complex agent applications, design a spec-driven, REPL-first architecture over MCP. MCP works for exposing limited tools in existing apps, but a REPL approach lets agents write and run code to interact directly with underlying data. This lets agents build their own logic, loops, and functions instead of relying on rigid APIs. This flexibility requires strict security measures and automatic feedback loops to catch bad code execution.
  • Read more

3. AI agents are everywhere at @Uber — X (formerly Twitter)

  • Why read: How Uber scales and secures 50,000 daily agent sessions.
  • Summary: Traditional Endpoint Detection and Response (EDR) fails for agents because it misses the prompt and reasoning behind network calls. Uber built Agentic Detection and Response (ADR) to capture the full chain: prompt, reasoning, tool call, and outcome. It triages safe sessions cheaply and reserves expensive LLM analysis for suspicious events, hitting zero false positives on enterprise benchmarks. Security teams need to stop monitoring isolated tool calls and start observing the full workflow, as attacks can hide in steps that look safe on their own.
  • Read more

4. Self-Improving Agents Are Event-Sourced — Lobu

  • Why read: How to build reliable, queryable memory for agents.
  • Summary: Most agents use filesystem memory, which lacks schemas and struggles with conflicting facts or tracking data origins. Agents need append-only memory that logs every action and correction, similar to Git history. In this event-sourced model, you correct a fact by writing a new one that replaces the old. This keeps the history walkable and treats human review like a merge request. This database-backed approach provides the transparency and state tracking needed for enterprise deployment.
  • Read more

5. Gavin Baker: "if you're a foundation model company and you... — X (formerly Twitter)

  • Why read: The physical limits on AI infrastructure and the short shelf life of foundation models.
  • Summary: Foundation models without unique data or internet-scale distribution depreciate quickly as open weights catch up. The main bottleneck for AI progress is no longer GPUs, but the power required to run them. Future mega-data centers will need to sit near unregulated nuclear reactors, severely limiting US locations. Additionally, upcoming memory upgrades in mobile devices could spark a massive new memory cycle, shifting hardware investment.
  • Read more

6. 🏰 The Only Moat That Survives AI — Superintelligence.

  • Why read: Why human verification is the last defensible advantage in AI.
  • Summary: Open-weight models lag the frontier by months, erasing traditional moats like execution and routing. If a task can be measured, it will be automated. Value is shifting to the last mile: human verification of machine output. Owning this feedback loop and feeding it back into training builds a network effect that strengthens with each model update. Stop trying to own the best model and start building a proprietary measurement and validation layer.
  • Read more

7. when you experience a language model doing a task that... — X (formerly Twitter)

  • Why read: How to handle the identity crisis that follows AI automating your work.
  • Summary: When AI automates a task central to your identity, it triggers a loss of agency. The reaction follows a pattern: first seeing the AI do the work, then realizing the human setup required to guide it, and finally rebuilding a sense of agency. In the end, the AI becomes background infrastructure, and the human shifts to high-level direction. Recognizing this cycle helps teams get through the friction of AI adoption.
  • Read more

8. Devtools must be open source — exe.dev

  • Why read: How agents make custom software environments easy to maintain.
  • Summary: Customizing software is getting easier because agents can download, modify, and recompile source code. Agents run nightly cron jobs to fetch upstream changes, rebase local edits, check functionality, and deploy the update on their own. Developers can reshape their tools with prompts instead of wrestling with plugins or config files. This personalization only works if the underlying devtools are open source, giving agents access to the code.
  • Read more

9. Snacks are cheaper and tastier than ever — Justin Duke

  • Why read: Why you shouldn't use LLMs to procrastinate on hard problems.
  • Summary: LLMs make it easy to dodge difficult, high-impact problems by "snacking" on trivial tasks. Prototyping, writing boilerplate, and generating charts are now frictionless, creating a false sense of accomplishment. Teams risk optimizing output without actually moving the needle for the business. Real progress means redefining "done" from shipping code to delivering customer value.
  • Read more

10. Agent Experience — proofofconcept.pub

  • Why read: Agent Experience (AX) is joining UX and DX as a core design discipline.
  • Summary: Agents are becoming primary users of software systems, requiring designers to consider Agent Experience (AX). Unlike humans, agents don't complain about friction; they silently retry, brute-force, or fail. Good AX means building parallel interfaces and logging so humans can see agent confusion and dead ends. Product teams need to stop cramming humans and agents into chat interfaces and build decoupled apps where the agent is the controller.
  • Read more

11. few people have had more foresight than wei dai: — X (formerly Twitter)

  • Why read: The AI industry's lack of strategic foresight on scaling and reinforcement learning.
  • Summary: The AI industry struggles with long-term strategic planning. Improving an AI's strategic capability through long-horizon reinforcement learning also increases misalignment risk. The community has trouble building projects that extract strategic advice from AI because progress is hard to measure. Scaling compute without fixing these evaluation bottlenecks increases risk.
  • Read more

12. Few thoughts on how Astra results relate to algorithmic progress... — X (formerly Twitter)

  • Why read: What AI solving new math theorems means for AI research and deep learning theory.
  • Summary: Frontier models solving old math conjectures points to automated AI research. Since AI R&D is verifiable, these math skills will likely improve deep learning theory, architectures, and optimizers. If investing in "theory compute" drastically improves training, it could create a feedback loop that speeds up timelines. Prepare for algorithmic progress driven by automated R&D loops, not just human researchers.
  • Read more

13. How to Measure Intelligence Beyond Human Scale? — lesswrong.com

  • Why read: How to evaluate AI once it beats human benchmarks.
  • Summary: Human-written benchmarks fail when AI gets smarter than humans. Researchers suggest "adversarial psychometrics": agents generate questions to test each other without a human judge. Agents get rewarded for writing valid problems that differentiate peer capabilities. This Elo-style system scales as models improve and provides a way to track progress past human-level intelligence.
  • Read more

14. Some initial thoughts, and a complicated mix of feelings — X (formerly Twitter)

  • Why read: A mathematician's reaction to AI solving decades-old theorems.
  • Summary: An internal OpenAI model resolved major math conjectures, prompting existential questions from experts. The formal verification is impressive, but the AI's documentation hides the key insights under boilerplate text. This makes it hard for researchers to follow the AI's intuition. As AI solves verifiable problems, experts have to accept answers that lack human clarity.
  • Read more

15. We are not cleared for takeoff. Yet. — X (formerly Twitter)

  • Why read: Why recent math breakthroughs don't mean AGI is here.
  • Summary: Despite huge wins in verifiable fields like math and code, the conditions for an AGI takeoff aren't met. Models still have uneven capabilities, struggle in open-ended environments, and fail at long software projects. Labs haven't cracked the credit assignment and long-term reinforcement learning required to generalize outside narrow domains. The tech handles short tasks well, but broad autonomy is still far off.
  • Read more

Themes from yesterday

  • Agent Infrastructure Costs: MLA architectures are cutting context caching costs, but power limits will constrain data center expansion.
  • App Architecture: Developers need to build REPL-first systems where agents act as controllers, treating Agent Experience (AX) as a core design discipline.
  • Automating R&D: AI is hitting superhuman levels in math and verifiable research, leading to faster progress but harder-to-read proofs.
  • Verification and Security Moats: With model capabilities commoditizing, the remaining advantages are human verification loops and full-workflow security observability.
  • AI Psychology: Automation triggers identity crises, tempting workers to use LLMs for easy tasks and avoid hard strategic work.