1. 看了张小珺对罗福莉的长访谈，信息密度极高，聊了将近两万字，我尽量把核心观点压缩一下。 — loonggg

  • Why read: A rare, deep technical interview with Xiaomi's MiMo lead and former DeepSeek core author on the fundamental shift to an agent-first AI paradigm.
  • Summary: AI has transitioned from a pre-training dominant Chat era to a post-training dominant Agent era, highlighted by the OpenClaw framework and Claude Opus 4.6. A well-orchestrated agent framework can significantly compensate for a model's shortcomings, enabling even small 3B edge models to perform complex tasks. The structural gap in domestic pre-training is closing, making agent-based reinforcement learning the new competitive frontier. Compute allocation has drastically shifted from a 3:5:1 ratio to 1:1 between pre-training and post-training. Ultimately, the rapid capability gains suggest AGI could arrive within two years, radically transforming work modes first.
  • Read more

2. DeepSeek V4 is a Serious Threat — Matthew Berman

  • Why read: An analysis of DeepSeek V4's disruptive potential as a frontier-level, open-weight model offered at a fraction of the cost of US counterparts.
  • Summary: DeepSeek V4, including a 1.6T parameter MoE Pro and a 284B Flash version, has achieved parity with models like Opus 4.7 and GPT-5.5 on agentic coding benchmarks. By offering top-tier performance at drastically lower prices, it challenges the economic viability of closed-source US enterprise models. This introduces a significant geopolitical security risk if US enterprises build their AI strategies on Chinese infrastructure. Furthermore, DeepSeek's open-source innovations, driven by compute constraints, prove that algorithmic efficiency can overcome hardware limitations. The US must aggressively embrace open-source strategies to maintain its technological and cultural leadership in the global AI ecosystem.
  • Read more

3. What’s 🔥 in Enterprise IT/VC #495 — Ed Sim

  • Why read: A look at Cursor's reported $60B acquisition by SpaceX and what it signals about exit strategies and valuations in the AI era.
  • Summary: Cursor's staggering $60B deal highlights a new reality where explosive growth and strategic necessity outweigh traditional software metrics like negative gross margins. SpaceX's acquisition demonstrates the massive premium placed on controlling the developer ecosystem and AI coding tools. Coding remains the largest proven application market for AI, justifying unprecedented valuations for market leaders with high product velocity and usage data. For founders and investors, the key lesson is knowing exactly when to play your hand and capitalize on extreme scarcity value. Such generational outcomes redefine the boundaries of acceptable risk and reward in enterprise AI.
  • Read more

4. Who Gets an FDE — SaaStr

  • Why read: A sobering look at the widening gap between enterprises that get hands-on AI deployment and smaller companies left to self-serve.
  • Summary: As AI agent vendors grow, a two-tier system is emerging where only large enterprises receive dedicated Forward Deployed Engineers (FDEs), while smaller clients must rely on self-serve documentation. Real-world data reveals that agent success hinges not on the underlying model, but on expert human deployment that perfectly integrates the agent into existing workflows and data systems. Enterprises with FDEs are achieving 60-80% automation rates, whereas self-trained deployments languish around 20%. For example, Salesforce's Agentforce saw transformative results when properly deployed, yet broader adoption remains low due to the sheer difficulty of self-guided integration. If you want transformative AI results, you must secure the human expertise required to truly embed the technology into your operations.
  • Read more

5. Personal Agent — Trash Panda 🦝

  • Why read: An inspiring look at a custom-built, fully "vibe-coded" desktop agent harness that gives users control without big-lab lock-in.
  • Summary: PersonalAgent is a multi-threaded desktop application built on top of Pi's coding agent, completely developed without manually writing a single line of code. It features a desktop UI for managing multiple agent threads and an iOS companion app for monitoring long-running tasks remotely via Tailscale. The harness introduces advanced agent lifecycle management, allowing the AI to independently schedule recurring tasks, queue follow-ups, and trigger deferred resumes instead of relying on inefficient bash sleeps. It also includes an "Auto Mode" that intelligently prompts the agent to continue working until a task is genuinely complete or blocked. This setup illustrates the power of owning your own agent harness to maximize productivity and bypass the limitations of generic terminal UIs.
  • Read more

6. CPUs: A Better Story Than Photonics? — Jason's Chips

  • Why read: A compelling thesis on why the shift towards agentic AI will drive explosive, underappreciated demand for CPUs.
  • Summary: The market currently views CPU demand as a simple linear increase relative to GPU usage, but the reality of agentic workflows suggests a far more dramatic structural shift. Using a "fingers-to-brains" analogy, as AI models (brains) become exponentially smarter and more efficient, the need for CPUs (fingers) to execute complex, multi-step actions will skyrocket. While training heavily favors GPUs, inference for complex agent loops shifts the workload mix significantly toward CPUs to manage application logic, data retrieval, and system interactions. This means CPU demand will decouple from simple feed-forward inference and scale independently alongside the rising autonomy of AI agents. Investors and operators should recognize that CPUs may become one of the most critical and supply-constrained bottlenecks in the next phase of AI infrastructure.
  • Read more

7. DESIGN: THE FIRST AI CASUALTY — Gokul Rajaram

  • Why read: A provocative prediction that traditional product design roles will be entirely absorbed by AI and cross-functional builders by 2026.
  • Summary: The standalone product design function is rapidly becoming obsolete as startups leverage AI to handle UI/UX execution. Companies are increasingly relying on temporary design consultants to establish foundational design systems, which product managers and engineers then feed into AI tools to generate fully functional prototypes. As larger enterprises follow suit, design departments are projected to shrink dramatically, retaining only a fraction of their current workforce. Designers must adapt by either becoming entrepreneurial consultants who sell design systems or expanding their skill sets to include product management and engineering. To survive this transition, design professionals need to pivot from specialized execution to holistic product building.
  • Read more

8. Deep Moats and Platform Shifts in Computing - Part 1 — Pushkar Ranade

  • Why read: A historical analysis drawing parallels between the RISC vs. CISC processor wars of the 1980s and today's AI accelerator landscape.
  • Summary: The dominance of a computing platform often feels invincible until an underlying architectural shift completely redefines the market. Just as the IBM PC's adoption of Intel's x86 architecture cemented a decades-long monopoly despite technically superior alternatives, today's AI ecosystem is heavily entrenched in incumbent paradigms. The Wintel flywheel demonstrated how software compatibility and distribution scale can overcome inelegant technical designs. However, history shows that even the deepest moats eventually succumb to new computing epochs, such as the shift from desktop to mobile. Understanding these cyclical platform transitions is crucial for anticipating when and how the current GPU and CUDA monopolies might eventually be unseated.
  • Read more

9. John Ternus, Apple's incoming CEO, on the Steve Jobs story... — Big Brain Business

  • Why read: Insight into the leadership philosophy of Apple's incoming CEO and the foundational principle that continues to drive their product design.
  • Summary: Incoming Apple CEO John Ternus emphasizes a core Steve Jobs philosophy: finishing the back of the cabinet just as beautifully as the front, even if no one will ever see it. This uncompromising standard of excellence is applied uniformly across Apple's entire lineup, from the premium iPhone Pro Max to the entry-level MacBook Neo. Ternus views this dedication to internal and external beauty as the ultimate indicator of Apple's future direction under his tenure. It reinforces the idea that true quality is defined by what a company refuses to compromise on, rather than just what is visible to the consumer. This cultural bedrock ensures that Apple's hardware continues to command premium brand loyalty through sheer craftsmanship.
  • Read more

10. The world’s most complex machine — Neil Hacker

  • Why read: A deep dive into how ASML monopolized the semiconductor industry by mastering extreme ultraviolet (EUV) lithography.
  • Summary: ASML transformed from an industry laggard into the sole supplier of the world's most complex machines by taking a massive gamble on unproven EUV technology. These room-sized machines, which require three cargo planes to ship, are the critical chokepoint for manufacturing all cutting-edge microchips globally. By perfectly orchestrating over a hundred thousand components, ASML's photolithography allows billions of transistors to be stenciled onto a single silicon wafer. This technological monopoly not only drives the relentless shrinking of computing power but has also positioned ASML at the epicenter of modern geopolitical conflict between the US and China. Their success underscores the immense strategic value of dominating a fundamental hardware capability through unparalleled precision engineering.
  • Read more

11. Longreads + Open Thread — Byrne @ The Diff

  • Why read: Fascinating commentary on how AI models like Sonnet 4.7 can unerringly identify the authors of pseudonymous texts, threatening online anonymity.
  • Summary: Advanced LLMs are now capable of analyzing a piece of text and accurately identifying its author based on their writing style across various historical contexts and genres. This capability implies that true pseudonymity may soon only be available to those who have zero pre-existing digital footprint or those who heavily sanitize their writing through AI tools. As inference costs plummet, casually deanonymizing colleagues or whistleblowers could become trivial. The defense against this requires using AI to explicitly strip personal stylistic markers before publishing sensitive information. This marks a profound shift in digital privacy, where your unique voice is as identifiable as a fingerprint.
  • Read more

12. every tech executive is talking about making it so anyone... — dax

  • Why read: A reality check on the executive dream of empowering non-engineers to ship code safely.
  • Summary: Tech executives are currently obsessed with creating environments where anyone in the company, such as marketing teams, can push changes directly to production. While framed as a new AI-driven capability, this has always been the fundamental responsibility of senior engineers: building guardrails and design patterns to make less experienced team members productive without breaking the system. To actually realize this vision of democratized development, a company's core engineering team must be exceptionally skilled at system architecture and robust testing. Ultimately, AI doesn't eliminate the need for great engineering; it shifts the focus entirely toward designing foolproof infrastructure that can withstand chaotic, automated, or amateur inputs.
  • Read more

13. The best companies and teams in the world become great... — Fahd Ananta

  • Why read: A succinct framework on talent density, featuring an incredible anecdote about Elon Musk's extreme measures to hire a top engineer.
  • Summary: The foundation of any world-class organization rests on three equally difficult pillars: attracting exceptional talent, developing individuals with high ceilings, and rigorously avoiding hires who interview well but underperform. To illustrate the lengths leaders will go for the right talent, the author shares a story of a SpaceX candidate whose wife worked at Google in SF. Elon Musk personally called Larry Page to arrange her transfer to LA just to secure the hire. Building an elite team requires a relentless, almost unreasonable commitment to removing every obstacle between you and top-tier talent. This level of aura and intensity is what truly differentiates legendary companies from the rest.
  • Read more

14. An MIT professor taught the same math course for 62... — Ihtesham Ali

  • Why read: The inspiring story of how MIT's Gilbert Strang quietly revolutionized global math education by democratizing his linear algebra course.
  • Summary: For 62 years, Gilbert Strang taught MIT 18.06 Linear Algebra, ultimately becoming the default math teacher for the entire planet when he released his lectures for free online in 2002. While most professors taught abstract vector spaces first, Strang inverted the curriculum to start with concrete matrix multiplication, demanding intuitive understanding before introducing complex proofs. His profound respect for students and his insistence on clear, jargon-free explanations allowed countless self-taught programmers and engineers to actually understand the math underlying AI. By the time he retired, entire universities globally had reshaped their curricula around his YouTube playlist. His legacy is a testament to the power of open knowledge and teaching with immense empathy and clarity.
  • Read more

15. Revisiting The Outcome Distortion Complex — Kyle Harrison from Investing 101

  • Why read: A reflection on how market outcomes consistently reshape our perception of reality, strategy, and decision-making quality.
  • Summary: "The ends rewrite the means"—when outcomes are exceptionally positive, we tend to post-rationalize the chaotic or flawed processes that led us there. Conversely, during market downturns, we assume our foundational strategies were inherently broken simply because the macro environment shifted. It's crucial for operators and investors to decouple their evaluation of process from the ultimate outcome, especially in highly speculative environments. Believing that positive outcomes validate bad habits is just as dangerous as letting negative outcomes breed unnecessary self-loathing. Mastering this distortion complex is essential for maintaining strategic clarity regardless of where we are in the economic cycle.
  • Read more

Themes from yesterday

  • The Open Source AI Geopolitical Threat: The rise of DeepSeek V4 and state-level algorithmic breakthroughs are forcing a structural rethink of closed-source AI moats and US enterprise dependencies.
  • The Shift from Human Capital to Human Expertise: While AI automates execution (like product design or entry-level coding), the premium on deep human expertise—such as elite deployment engineers or 10x architectural talent—is higher than ever.
  • Agentic Architectures Reshaping Compute: The transition from Chat to Agent paradigms is massively shifting compute ratios from pre-training to post-training, and simultaneously pointing toward a looming CPU bottleneck for complex agent loops.