1. A global workspace in language models — anthropic.com

  • Why read: Anthropic found Claude developed a silent "J-space" for internal reasoning, changing how we understand model cognition.
  • Summary: Anthropic researchers found Claude naturally developed an internal "global workspace" (or "J-space") during training. This lets it reason silently without outputting text. This internal scratchpad directly affects the model's performance on multi-step tasks. If you suppress the J-space, the model loses its high-level reasoning but keeps its basic conversational fluency. For developers, this means a model's limits depend on invisible internal states. It opens the door to new prompting and steering methods that target these hidden reasoning processes directly.
  • Read more

2. The Stable Era: The Modularisation Of Post-Training — Akash Bajwa

  • Why read: A framework explaining the shift in the AI industry from vertical integration to modular, specialized parts.
  • Summary: The AI industry is entering a "stable era" defined by standard interfaces like the transformer architecture, OpenAI-compatible APIs, and agent frameworks. This stability allows different layers of the stack to innovate independently, much like the PC revolution. Open-weight models are accelerating this by making intelligence cheaper and reducing the advantage of closed models. Developers need to realize the cost of shaping intelligence is dropping fast. Long-term moats will move from the models to the application layer and the agent workflows built on top of them.
  • Read more

3. Continual Learning for Agents — Michele Catasta

  • Why read: Replit's Head of AI explains how to build agents that improve continuously, even if you can't edit the underlying model weights.
  • Summary: Continual learning isn't just about updating model weights. When using closed models, builders need to use harness-level and context-level learning to improve their agents. Replit evaluates its agents using offline benchmarks for regressions, A/B tests for production impact, and trace clustering to find failures. This makes evaluation a continuous feedback loop rather than a launch checklist. Teams need to stop looking at single-score metrics and start mining production traces to update their agent's tools, code, and prompts.
  • Read more

4. Everyone is wrong about open source AI in the enterprise — Jesse Zhang

  • Why read: Why total enterprise spending on open-source AI is falling, even as advanced startups move their production entirely to open weights.
  • Summary: Despite the hype, open source only accounts for 11% of enterprise LLM spending. This is because most enterprise AI use cases are still experimental, so companies use massive, general-purpose frontier models to explore. But for mature, high-stakes deployments like Decagon's customer service agents, small, heavily fine-tuned open-source models are the only way to hit speed and quality targets. As more enterprise AI projects mature, companies will shift from renting frontier models to running their own specialized open-source models in production.
  • Read more

5. Open source will win AI, and it won't even be close. — Xiaoyin Qu

  • Why read: A historical argument that price collapse, not ethics, will drive open-source AI to market dominance.
  • Summary: Historically, major technological shifts like electricity and computation were triggered by a massive drop in price, not the initial invention. Right now, AI is a metered luxury controlled by a few labs. Open source is the only force capable of commoditizing intelligence into cheap infrastructure. As open source drives the cost of AI to zero, we will see new industries and jobs emerge that don't make economic sense under today's pay-per-token pricing.
  • Read more

6. Introducing Crawford — A Claude Code AI Agent for Cheap Right Answers — Jordan Crawford

  • Why read: How to optimize AI research agents for accurate answers per dollar instead of raw model speed.
  • Summary: Most AI research tools compete on speed and accuracy but hide the cost. This leads to expensive mistakes, like failing to notice a domain is parked. Crawford was built to optimize for correct answers per dollar, achieving 34 correct out of 39 rows for a total of 77 cents. This treats AI research like chip design: controlling exactly when money is spent, what runs first, and how data is logged. Developers building agent workflows need this mindset. Run cheap validation steps, like checking a webpage's actual content, before using expensive frontier model tokens.
  • Read more

7. A Stargate for Data — will depue

  • Why read: AI scaling is moving from compute limits to data limits, which will require massive investments in new data collection.
  • Summary: Deep learning grew on the back of the open internet, but we are running out of high-quality public text. As AI gets closer to automating knowledge work, progress will hit a wall without new data in specialized, undocumented areas. Capturing this "dark matter" of tacit knowledge and internal company data will take a massive collection effort, costing over $100 billion a year by 2030. Companies with offline, proprietary, or hard-to-digitize data hold highly valuable assets. This data will fuel the next wave of pretraining and reinforcement learning.
  • Read more

8. 10 Lessons for Agentic Coding — Drew Breunig

  • Why read: Practical rules for engineers working in an environment where code generation is practically free.
  • Summary: With AI reducing the cost of writing code, developers need to focus on architecture and intent rather than typing syntax. Teams should use spec-driven development: keep markdown plans updated to guide agents, and treat the actual code as disposable. End-to-end behavioral tests are now reliable contracts that let you safely delete and rewrite features using AI. With boilerplate automated, engineers can spend their time on hard, high-value problems like system design, performance, and user experience, where human judgment is still required.
  • Read more

9. Microsoft is Finally Saying the Quiet Part About AI Adoption Out Loud — Brad Groux

  • Why read: Microsoft's $2.5 billion investment in an AI implementation team shows that out-of-the-box AI tools aren't enough to drive real enterprise productivity.
  • Summary: The idea that companies can drop a chatbot onto their documents and instantly boost productivity is dead. Microsoft's new 6,000-person "Frontier Company" is an admission that adopting enterprise AI takes serious change management. You have to understand tribal knowledge, shadow IT, and office politics. Models can summarize a documented process, but they don't know if employees actually follow it. AI vendors need to stop pushing generic tool stacks. Instead, they must map a business's actual friction points before deploying agents.
  • Read more

10. LLM Wiki's Don't Work — Jacob Posel

  • Why read: Why pointing an LLM at your company's Notion or Google Drive fails to create a useful company brain.
  • Summary: LLM wikis look great on day one, but they degrade quickly. Business context isn't like a static codebase. Real corporate knowledge—intent, rejected ideas, unwritten constraints—lives in Slack threads and customer calls, not sanitized documents. If teams have to update a separate wiki outside their normal workflow, the data gets stale, sources get lost, and people stop trusting the AI. A real company brain needs to be integrated directly into the communication channels where work actually gets done.
  • Read more

11. Moving from firms to tasks — Kinjal Shah

  • Why read: How AI agents as "programmable labor" will unbundle the traditional company into a network of task-based transactions.
  • Summary: Past technological shifts changed the boundaries of the firm by making it cheaper to coordinate labor. Now, the mix of programmable labor (AI agents) and programmable money (crypto) is pushing coordination costs to near zero. Soon, an agent will be able to hire a specialized sub-agent for a specific task, check the work, and pay for it on-chain in milliseconds. The primary unit of economic value is shifting from full-time employees to fast, automated task execution that happens outside of normal corporate structures.
  • Read more

12. Good Judgment Beats Good Prompts — Aadil Maan

  • Why read: AI agents can't fix broken organizational design. They just highlight the existing mess.
  • Summary: The author tried building an AI agent to summarize morning briefings, only to find the real problem was a bloated org chart full of unnecessary meetings and scattered data. AI is great at execution, but it doesn't know if a meeting should exist in the first place. Good judgment requires knowing the context of why and when things happen. That remains a human skill. Operators need to fix their basic workflows and clean up their data before layering AI on top of a mess.
  • Read more

13. The Token Economy - State of AI — John Furrier

  • Why read: A look at how the AI industry has moved from software demos to heavy industrial manufacturing.
  • Summary: The AI boom is no longer just about model updates. It's about pouring concrete and securing power grids. Companies are building gigawatt-scale "AI factories" next to cheap, stranded energy because moving power is harder than moving data. Intelligence is now a manufactured commodity sold in tokens, and countries view GPU access as a national security issue. The AI economy is now limited by physical things like energy, land, and steel, not just software breakthroughs.
  • Read more

14. Replit Head of AI @pirroh says we're entering the "post-prompting era" — TBPN

  • Why read: How long-horizon AI agents are moving developers away from micromanaging prompts toward setting broad verification rules.
  • Summary: Prompt engineering is fading as models get better at long-horizon tasks. Instead of writing strict instructions, developers now set a high-level goal, give the agent tools, and let it iterate. The agent framework handles verification, measuring success and adjusting until the job is done. Builders need to adapt to this "post-prompting" environment by defining clear success metrics and building closed-loop systems where agents can experiment and correct themselves.
  • Read more

15. 🎙️ How I AI: Sonnet 5 review & How to run autonomous coding agents from your phone — Lenny's Newsletter

  • Why read: How to build personal, repeatable benchmarks to test new models against your specific tasks.
  • Summary: Instead of trusting public leaderboards, build your own benchmarks based on your actual work. The author used Claude Code to build an HTML scoring tool, then ran blind tests of Sonnet 5 against other models using past chat logs. Even with good pricing, Sonnet 5 didn't always win on personal preference. Cost efficiency only matters if the model actually does your specific job well. Engineers should spend an hour building automated tests on their own data before switching to a new foundation model.
  • Read more

Themes from yesterday

  • Models to systems: Attention is shifting from foundation models to the agent frameworks, workflows, and evaluation loops built around them.
  • The cost of intelligence: There is a growing divide between massive, energy-heavy frontier model training and the push for cheap, specialized open-source inference.
  • Enterprise reality: Pointing an LLM at messy corporate data doesn't work. Real enterprise AI requires fixing broken workflows, cleaning up org structures, and capturing the unwritten context of how work gets done.
  • The agentic mindset: Software development is moving from granular coding and prompting to setting goals, building automated loops, and treating code as disposable.