1. Kimi K3 Stuns the Tech World — Contrary Research

  • Why read: Chinese open-weight models are unexpectedly competitive. This changes the balance of power in global AI.
  • Summary: Chinese startup Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that reportedly beats Claude Opus 4.8 and GPT-5.5. It offers top-tier reasoning and coding at a fraction of closed-API costs. This suggests US chip export controls aren't stopping Chinese AI progress, as architectural tweaks like MoE routing make up for compute limits. For enterprises, a cheap, frontier open-weight model completely changes the math of relying only on closed labs. If verified, this forces a rethink of global AI supply chains.
  • Read more

2. [AINews] not much happened today — AINews

  • Why read: The technical AI community's reaction to Chinese open-weight models and the shift in focus from raw compute to data.
  • Summary: Builders are debating if Kimi K3's success comes from better data and post-training rather than raw FLOPs. Benchmarks place K3 in the frontier tier, especially in coding. This challenges the idea that frontier capabilities require massive compute. Smaller, highly optimized stacks can close the gap with larger Western models. The lesson for builders is to focus on data and reinforcement learning loops over raw scale.
  • Read more

3. What’s 🔥 in AI/Infra/VC #507 — Ed Sim from What's Hot 🔥 in AI/Infra/VC

  • Why read: The bottleneck in AI infrastructure has moved from compute to human-verified reinforcement data.
  • Summary: Founders are confusing early investor interest with product readiness, trapping themselves in fundraising cycles. In AI infrastructure, the main bottleneck is no longer compute, but high-quality, human-verified reinforcement data. Startups that train directly on user workflows will build defensible products. Meanwhile, open-source AI is gaining political support against broad national security restrictions. Operators need to secure expert data traces and reinforcement learning pipelines to stand out.
  • Read more

4. Claude Became Our AI VP of Product — SaaStr

  • Why read: A case study of replacing expensive legacy software with a cheap AI agent in production.
  • Summary: The divide in B2B is between those talking about AI and those running it in production. One operator replaced a $10,000 legacy app in an hour with Claude, migrating ten years of Marketo workflows for $14. AI agents are already handling real go-to-market tasks, changing the math on software budgets and headcount. Separately, over half of all AI and B2B funding is now concentrated in the Bay Area. Startups need to stop pitching AI in slide decks and start shipping agents that cut costs.
  • Read more

5. Knowledge Bases — drew.tech

  • Why read: The engineering reality of building accurate AI knowledge bases for enterprise teams.
  • Summary: Unifying Slack, email, CRM, and call data into an accurate knowledge base is a hard engineering problem that basic prompting can't fix. Vercel solved this by indexing customer-specific data with vector embeddings and full-text search, using their Workflow SDK for background tasks. They used Turbopuffer to manage unstructured data cheaply. This system lets AI apps, like sales coaches, pull accurate context for deal reviews. Builders have to invest in data ingestion before trying to deploy context-aware agents.
  • Read more

6. Capital Allocation Is Dead — Kyle Harrison from Investing 101

  • Why read: Why standard spreadsheet models for venture capital fail in an AI-driven market.
  • Summary: Standard venture investing relies on linear growth and margin improvements, but AI breaks these models. Capital allocation is disconnecting from historical metrics. Investors sticking to old spreadsheets will misread a market that ignores five-year plans. Successful investing now means understanding how new tech rewrites entire categories. Founders looking for funding need to pitch massive shifts, not just incremental SaaS growth.
  • Read more

7. A British coder built the world's first network of fully... — Tymofiy Mylovanov

  • Why read: A solo developer built an autonomous drone network, showing how AI democratizes asymmetric warfare.
  • Summary: A British coder built a network of fully autonomous drones that operate without human input or radio signals. Because the AI makes decisions onboard, operators can deploy the swarm and leave. The drones create localized kill boxes, acting like intelligent minefields. By removing the need for remote pilots, the system neutralizes enemy jamming and counters numerical superiority. This shows how quickly solo developers with AI tools can upend military tactics.
  • Read more

8. 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗪𝗼𝗿𝗸 𝗧𝗶𝗽 #𝟭: Don’t start by trying to one-shot... — Kyle Kober

  • Why read: How to turn your existing work into reliable AI templates instead of writing new prompts from scratch.
  • Summary: Don't try to generate new materials with single, massive prompts. Start with a document, slide, or spreadsheet you already know works well. Break down your formatting rules and judgment calls, then build those specific constraints into an AI prompt. This turns vague AI features into tools that meet your exact standards. Codifying what already works scales your productivity better than starting from scratch every time.
  • Read more

9. 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗪𝗼𝗿𝗸 𝗧𝗶𝗽 #𝟮: Skills get better with every... — Kyle Kober

  • Why read: Why AI workflows require constant feedback to reach automation.
  • Summary: New AI workflows usually require heavy human oversight, starting around a 70/30 split of AI to human effort. When the AI fails, don't scrap the prompt; use the mistake to tighten the instructions. Each cycle of review improves accuracy and shifts more work to the model. A tuned process can eventually hit 95% automation. Building useful AI tools requires ongoing refinement, not instant magic.
  • Read more

10. Longreads + Open Thread — Byrne @ The Diff

  • Why read: How automation changes the value of human connection and coordination in business.
  • Summary: Coordinating complex systems is hard. Air travel in developing markets shows industries only mature when everyone prioritizes reliability over personal convenience. In AI, models can copy outputs but can't form human relationships. As AI handles the busywork, tasks where consumers value human connection will become rarer and more expensive. Also, industries mature by adapting expertise from nearby peers, not just importing it from pioneers. Businesses need to figure out what they sell: raw output or human connection.
  • Read more

11. A tumor the size of 2 fists was growing inside my chest. I had no idea. — Conno Christou

  • Why read: How a founder managed the psychology and uncertainty of a surprise cancer diagnosis.
  • Summary: A founder who followed strict health protocols discovered a massive tumor in his chest. Facing conflicting advice from top doctors on treatments and survival rates, he had to make a fast, life-altering decision. He survived the treatment by treating it like a project with a deadline and refusing self-pity. He found that sleep, protein, and mindset were the only variables he could actually control to get through the day. You can't prevent black swan events, but you can control how you execute your response.
  • Read more

12. Fujifilm X100 vs Leica Q vs Ricoh GR — Simon Sarris

  • Why read: How hardware limits remove friction and increase creative output.
  • Summary: Fixed-lens cameras like the Fujifilm X100, Leica Q, and Ricoh GR stop users from stressing over gear. By removing the option to zoom or change lenses, they force constraints that make people shoot more and focus on composition. Good default processing also skips the hassle of RAW editing. This shows that limiting choices and nailing the default experience drives higher use and better results. Sometimes, removing options is better than offering endless flexibility.
  • Read more

13. "A New “Movement Class” Is Emerging" and 4 more — Substack

  • Why read: Why users fail to get the most out of advanced AI models like Claude.
  • Summary: Power users are realizing they aren't ambitious enough with models like Claude. Many use AI just to clear trivial tasks for a quick dopamine hit. The real value comes from handing the model complex, multi-step problems, not basic chores. Operators should re-engineer their workflows to offload entire processes, not just busywork. Stop using frontier models for a dopamine backlog.
  • Read more

14. “Could I really claim these hallmarks of adulthood if the only thing making them possible is my father’s generosit… — The Substack Post

  • Why read: How hidden financial safety nets change our definitions of adulthood and risk.
  • Summary: An essayist realizes her main markers of adulthood—buying a home and raising a kid in an expensive city—are subsidized by her parents. This financial safety net changes her career choices, stress, and risk tolerance compared to peers without it. It points to a tension where the milestones of adulthood are separating from actual financial independence. It's a reminder to audit the hidden safety nets that allow for entrepreneurial risks. Knowing your dependencies helps you assess your real resilience.
  • Read more

15. taiwan travelogue — Ava from bookbear express

  • Why read: Managing the psychological toll of long-term creative work and the benefits of changing locations.
  • Summary: A writer details the anxiety of a months-long novel revision with slow feedback loops. Hiring an editor kept her grounded, showing that solitary work needs outside structure to prevent burnout. A trip to Taiwan served as a circuit breaker, proving the value of moving physically to separate work from life. The contrast between isolated tech work and a foreign environment helps manage fatigue. Creators need to force geographic and social changes to keep their stamina up on long projects.
  • Read more

Themes from yesterday

  • Chinese open-weight models are closing the gap with US frontier models, changing the math on enterprise AI costs.
  • Iterative AI workflows with continuous feedback are replacing zero-shot prompting as the standard for automation.
  • Constraints—like fixed-lens cameras, strict prompt formats, or rigid project scopes—force focus and improve output.