1. ‼️突发：DeepSeek @deepseek_ai 创始人梁文锋在一场据称长达四小时的投资者会议上的发言内容曝光。 — X (formerly Twitter)

  • Why read: A leaked look at DeepSeek's roadmap, highlighting their tight focus on AGI and how cost efficiency acts as a structural advantage.
  • Summary: DeepSeek founder Liang Wenfeng says they are ignoring distractions like video generation and consumer apps to focus entirely on AGI. Their technical roadmap moves from chain-of-thought to agents, then continuous learning and self-iteration. He views continuous learning as the next major hurdle. DeepSeek relies on cheap models to train larger ones under strict compute limits. They treat open-source as a strategic wedge to prevent monopolies and build momentum. Above all, the company prioritizes team stability over maximizing revenue, betting that cost efficiency decides the winner.
  • Read more

2. Inside the Model Factory — Eiso Kant, Poolside AI — Latent.Space

  • Why read: How a small startup competes with major labs by building a highly automated pipeline for training LLMs.
  • Summary: Poolside AI's co-founder explains their "Model Factory," which runs up to 20,000 experiments a month with fewer than 70 researchers. They treat model training as a systems engineering problem, streaming data directly into training and iterating from pre-training to release in eight weeks. The team bets that agent persistence, verification, and backtracking will soon matter more than raw zero-shot capability. This setup helped their smaller Laguna S model beat much larger open-weight models, showing that tight data pipelines can beat raw scale. The takeaway is that 95% of model building boils down to data and compute efficiency.
  • Read more

3. Sahil's tweet about doing 100% of Gumroad support while having... — X (formerly Twitter)

  • Why read: A practical example of using AI agents to handle customer support and write production bug fixes.
  • Summary: Gumroad's support queue is run entirely by an AI agent that queries production databases, checks Stripe, and reads the codebase to answer tickets. When the agent finds a bug, it writes the fix, opens a pull request, waits for CI approval, and emails the customer when deployed. Escalation rules send high-risk issues to humans. This turns support from a cost center into a product engineering function. It shows what happens when you give agents read-only database access and deployment tools instead of a canned LLM script.
  • Read more

4. An opinionated guide to which AI to use to do stuff — oneusefulthing.org

  • Why read: Ethan Mollick's guide to moving from chatbots to agents that execute real workflows.
  • Summary: AI usage is shifting from chat interfaces to agents that control virtual environments. Mollick argues free models work fine for basic tasks, but serious work requires paying for top-tier models (like Claude's Fable/Opus or ChatGPT's GPT-5.6 Sol) on high thinking settings. Users can grant these agents access to local machines or virtual sandboxes to run hours of autonomous work. To get value, treat the AI like an employee instead of an encyclopedia. Give it concrete objectives across multiple files and applications.
  • Read more

5. The Bottlenecks for AI-Driven System Design: why Principal Engineers are not (yet) obsolete. — mahesh’s blog

  • Why read: A breakdown of why agents write good code but fail at designing distributed systems.
  • Summary: Agents can write software and optimize performance in simulations, but they struggle with system migration. The hardest part of software engineering is safely moving a live system from one state to another without breaking things or causing downtime. Agents working in messy environments often create technical debt because of their narrow context. Human Principal Engineers are still needed because they understand the operational realities of live systems that can't just be restarted. For now, AI is better suited for bounded coding tasks than architectural overhauls.
  • Read more

6. The Pulse: New trend - concern about massive increase in code review load — Substack

  • Why read: AI coding tools are generating so much output that human code review has become a major bottleneck.
  • Summary: Models like Opus 4.5 and GPT 5.4 have shifted the primary engineering bottleneck from writing code to reviewing it. Developers are burning out from the volume of AI-generated pull requests, leading them to rubber-stamp approvals. Companies are trying to manage this with dedicated AI code reviewers and new workflows. Engineering teams have to move away from manual line-by-line reviews and rely more on automated testing and formal verification. Sticking to traditional code review in an AI-heavy environment guarantees a drop in code quality.
  • Read more

7. Part 1: AI, Rockets, and the Return of Hard Contracts — noumena.com

  • Why read: An argument that AI makes writing code cheap, but verifying it remains expensive.
  • Summary: AI doesn't make software cheap; it only commoditizes the act of typing it out. Determining if the code is secure and fits the architecture still requires expensive human effort. Since agents write code faster than humans can review it, treating every output as a finished draft requiring manual approval doesn't scale. The author argues we need aggressive, automated testing environments to filter out plausible garbage. Engineering teams need to invest in hard contracts and test infrastructure to safely handle the volume of code agents produce.
  • Read more

8. SaaSletter - Brute-Force AI + Gross Margins — Substack

  • Why read: How soaring inference costs are wrecking traditional software gross margins.
  • Summary: As the industry focuses on making AI economically viable, inference costs are spiking. Traditional SaaS companies rely on high fixed costs and low variable costs to maintain high margins. AI products face the opposite: volatile, highly variable inference expenses that look more like energy markets than cloud computing. A company's edge now relies on its deployment economics and infrastructure efficiency as much as its model capabilities. Builders have to optimize workload routing and secure compute capacity or watch their margins collapse.
  • Read more

9. engineer away the slop — Geoffrey Huntley

  • Why read: An argument that formal verification and deterministic testing are the only ways to handle low-quality AI code.
  • Summary: Because anyone can generate code now, we are facing a flood of defects and unreviewable output. To counter this AI "slop," the industry needs formal verification and deterministic testing so we can rely on automation instead of manual review. Tools that emulate deterministic computers will be necessary to catch deep system errors that standard tests miss. Writing code is basically free now; the new engineering bottleneck is automated verification.
  • Read more

10. __________________ __________________ Frontier — Google Docs

  • Why read: How the shift from "foundation" to "frontier" models reveals the precarious economics of AI labs.
  • Summary: The term "foundation" suggests a durable asset that generates steady returns. "Frontier" describes an expensive, temporary lead that has to be constantly bought back. This language shift shows that labs know the base layer will eventually be commoditized by open weights; profits will move to whatever remains scarce. Closed-source labs justify massive capital burn by treating the race itself as a moat, hoping for a breakthrough in recursive self-improvement. But if capabilities plateau, "good-enough" open models will eat their core business. The current AI business model is a high-stakes treadmill, not a stable utility.
  • Read more

11. this system will change your life... — X (formerly Twitter)

  • Why read: A guide to building a self-hosted knowledge base to give your local agents durable memory.
  • Summary: Most people suffer from "scattered context," repeatedly explaining their projects and preferences to AI models. The fix is having an agent maintain a centralized markdown vault as a single source of truth. Instead of relying on a model's built-in memory, you tell the agent to use its memory as an index pointing to this local folder. By structuring files methodically and requiring the agent to read them before answering, you give the AI real context. This setup cuts friction and makes local agents far more useful.
  • Read more

12. Series A is Discontinuous — X (formerly Twitter)

  • Why read: A warning that Series A funding has changed from a predictable milestone into a binary bet on outliers.
  • Summary: Hitting $1M to $5M in ARR used to guarantee a Series A. That path is dead. Today, even companies with $10M ARR and high retention fail to raise if they don't fit favored categories or seem sufficiently AI-native. Investors aren't looking for solid companies at reasonable prices; they only care if a startup can become a massive outlier. Lowering your valuation doesn't help because capital is strictly concentrated on a few perceived winners. Founders have to build massive narrative momentum while running the business as if the next round will never happen.
  • Read more

13. How to Win the Largest Market in AI — X (formerly Twitter)

  • Why read: Why the explosion in inference demand will push the industry from general GPUs to specialized chips.
  • Summary: Model training is a fixed capital expense, but inference is an ongoing operational cost that scales with usage. Thanks to coding agents and consumer apps, inference is becoming the largest workload in computing history. GPUs dominate now because their flexibility was needed during early architectural discovery. But generating tokens is memory-bound, not compute-bound, making GPUs inefficient for the job. As inference workloads stabilize, economics will force a shift toward custom silicon designed specifically for massive memory streaming.
  • Read more

14. A NEW NARRATIVE FRAMEWORK FOR DATA CENTERS — substack.com

  • Why read: The growing political clash between tech companies building massive data centers and the rural communities hosting them.
  • Summary: There is a disconnect between the tech industry pitching data centers as vital infrastructure and rural Americans seeing them as extractive. As AI demands massive build-outs, these facilities are landing in rural areas, moving the digital economy's footprint away from cities. But locals are fighting back over water usage, rising energy bills, and a lack of good jobs. Drawing on Rem Koolhaas’s theories, the author argues the countryside is the new frontline of technological change. To avoid moratoriums, tech companies have to figure out how to align infrastructure growth with tangible local benefits.
  • Read more

15. How to deploy a Cerebras-style Knowledge Base this week — X (formerly Twitter)

  • Why read: Why simple enterprise search beats complex AI knowledge graphs for internal company tools.
  • Summary: Cerebras built an internal knowledge base answering 15,000 queries a day by ignoring the trend of complex context graphs. Instead of freezing knowledge into a graph that immediately goes out of date, they use intelligent retrieval to search Slack, GitHub, and Jira at runtime. Knowledge stays in the systems where it was created. Building the infrastructure for syncing, ranking, and access control is hard and should usually be bought, not built. Connecting AI agents to a solid enterprise search layer is the best way to deploy a company brain today.
  • Read more

Themes from yesterday

  • The Shift from Creation to Verification: Writing software isn't the bottleneck anymore; reviewing it is. The industry is moving toward formal verification, deterministic testing, and strict contracts to handle the volume of AI-generated code.
  • The AI Infrastructure Squeeze: Heavy inference workloads and volatile token costs are wrecking traditional software margins. This is forcing a shift toward specialized inference chips and causing political friction over data center expansion in rural areas.
  • Agents in Production: Agents have moved from chatbots to actual workers. They are running full customer support loops (including writing and deploying bug fixes) and managing personal knowledge bases using runtime retrieval instead of static graphs.
  • The End of Middle-Class Tech Economics: From the language shift of "foundation" to "frontier" models to the harsh new reality of Series A funding, the market is punishing "good enough." Capital is concentrating entirely on extreme outliers that can survive the constant race for capability scaling.