Search the archive

TRY MERCURY COMMAND →

_*Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A., Members FDIC._

SO YOU WANT TO SELL INFERENCE

AI companies selling inference face a zero-margin, cost-plus pricing dilemma. They can achieve better margins through value-based pricing, charging based on outcomes rather than raw inference costs. Optimizing inference costs using model routing, caching, and distillation can also enhance margins and offer a competitive edge.

COMPANIES THAT BUILD COMPANIES

Polsia, a startup with AI agents instead of employees, claims $10M in annual revenue, illustrating the rise of companies building companies using AI. Both Polsia and YC-backed Thomas focus on AI-driven business models promising to innovate startup creation. While most ventures may fail, the few successful ones could prove lucrative, resembling Shopify's model where only a small percentage of ventures succeed significantly.

HOW TO STEAL A CUSTOMER BASE

The fastest way to take a competitor's customers is to leave them nothing to walk away from. Cursor didn't ask developers to switch off VS Code, it forked VS Code, so every extension, shortcut, and bit of muscle memory came along for free, then rebuilt the engine around AI underneath. Microsoft can't simply copy that back, because rebuilding VS Code that deeply would put its whole extension ecosystem at risk, so it won't, and that refusal is the counter-positioning that holds the door open.

7 MISTAKES FOUNDERS MAKE WHEN CHASING THEIR FIRST DEALS

Founders often delay selling for product perfection, underprice their offerings, and rely too much on a single customer. They should identify decision-makers, listen more than they speak, and adapt their pitch continuously. Early-stage technical founders must personally handle sales to fine-tune their approach before hiring dedicated salespeople.

INTRODUCING CLAUDE TAG

Slack users can now tag Claude in chats on selected channels. The bot can connect to tools and data to complete requested tasks. It builds context by remembering relevant information from the channels it's in, and it can plan out tasks to complete in the future. Claude Tag is now available in beta for Claude Enterprise and Team customers.

COLLABORATIVE POSTS ARE COMING TO LINKEDIN — WHAT THEY ARE AND HOW TO MAKE THEM WORK

LinkedIn will soon implement a feature that allows two or more people to share one post together, with every collaborator listed at the top. The feature is currently in beta with creators and brands. It will roll out over the coming months.

WHAT YOU NEED TO KNOW ABOUT LAMBDA MICROVMS

Lambda MicroVMs provide VM-level isolation with fast startup and stateful sessions. They are useful for coding assistants that need to run AI-generated code safely, interactive AI notebooks, agent sandboxes, and vulnerability scanners. The VM automatically suspends when idle and resumes when traffic arrives. Vertical scaling within a MicroVM is automated.

AI AGENT HYPE MEETS REALITY: THE PRODUCT DOESN'T WORK (AND CHURN IS COMING)

Instead of building a superagent that does everything, make the focus as narrow as possible. Make sure your product has been extensively tested and that it works in real-world conditions. Companies that promise everything but produce AI slop will lose customers quickly. This churn makes growth impossible to maintain.

EVERY MOAT BECOMES MOOT

A good reframe of what a moat actually is. It's not a static wall like a patent or a network effect, it's an engine that only keeps working if you stay willing to do the uncomfortable, repetitive thing rivals won't. Amazon is the example, the thin margins that look like weakness are the wall, because fat margins are really just a fund for your competitors' R&D. The moat that lasts isn't any single advantage, it's being able to build the next one faster than the current one erodes.

DEAD PLAYERS

Aiming for the stars now feels more impersonal and lower-risk than ever.

SIX MONTHS INTO THE “HERO'S JOURNEY”

The transition from VC to founder offers a new sense of autonomy.

SLOP AS A SERVICE

The "AI-generated" label doesn't make a bad video good, doesn't stop the feed from filling with junk, and doesn't tell you whether something is accurate, ethical, or just stolen and spun up for ad money.

Start for free →

WORKFLOWS ARE KING

The primary moat in SaaS was never data storage itself, but the workflows built around systems of record, which made switching costs prohibitively high. In an agentic world, the moat shifts from data ownership to workflow orchestration, where platforms coordinate, route, and govern agents across business processes.

EVALS: THE STRATEGIC IP THAT WILL DEFINE THE NEXT ERA OF AI

Evals are comprehensive frameworks that systematically measure and improve AI systems. Making production-quality agents that can actually do work starts with evals. A strong evaluation suite captures the nuances of judgment, tone, and taste, assesses agent use of tools, and breaks tasks down into specific, scorable dimensions. The best companies treat agentic evals as a core quality, reliability, and governance layer.

THE MINIMUM VIABLE UNIT OF SALEABLE SOFTWARE

There is a zone of viability in which, when priced within reason, it makes sense to buy software over building it, even with AI. Software in this zone is sufficiently novel to make a rebuild with AI non-trivial, with some ongoing maintenance burden, and priced appropriately to discourage rebuild-by-LLM. Somewhere along the zone of viability is the minimum viable unit of saleable software. Below this level, a rebuild is the same or less effort compared to going through the purchasing process for a third party and not cost-effective over the long run.

THE FAKE TRACTION PROBLEM

One of the most dangerous moments in fundraising is when a founder starts to believe their own traction before the market has fully validated it. Founders see early interest as proof their companies are working, while investors are more interested in whether that early interest reduces risk. The difference matters as fundraising is not won by proving that people are curious, but by proving that the company is becoming more investible. Fake traction might get founders into investor conversations, but it often fails during diligence.

WORK WHIPLASH

"Work whiplash" occurs when constant changes in priorities, poor communication, and unclear roles lead to employee burnout and wasted time. Effective communication and clear delegation can prevent this by ensuring everyone knows decisions, goals, and their responsibilities. Leaders must proactively close the information gap, making sure changes are communicated promptly to prevent employees from feeling undervalued.

LIGHTHOUSE AGENTIC BROWSING SCORING

The Lighthouse Agenting Browsing category evaluates how well sites are constructed for machine interaction. The standards for the agentic web are still emerging, so the project is currently focused on gathering data and providing actionable signals rather than a definitive ranking. Lighthouse uses a set of deterministic signals to evaluate pages to ensure audits are reproducible and suitable for integration into CI/CD pipelines. Websites can improve their agentic readiness by adopting WebMCP, ensuring a sound a11y tree, and optimizing for stability.

RECALL

Recall is a tool that keeps a local log of Claude Code sessions and condenses it into a resume-ready summary, entirely locally. The summary is generated using a local algorithm, not an LLM call, so it doesn't cost anything to run. Resuming from a compact memory instead of re-explaining each project from scratch can save a lot of tokens. Recall keeps everything on-machine, providing privacy, and it works offline.

RISK IS SEVERELY MISUNDERSTOOD

Top candidates overestimate the risk of joining elite early-stage startups by conflating uncertainty with actual downside. For startups with exceptional teams, strong investors, and important markets, the financial upside can be massive while the downside remains limited because employees can typically return to established companies if the startup fails.

ON THE PRECIPICE

AI is weakening the three forces that historically protected cybersecurity incumbents: switching costs, proprietary data advantages, and distribution. AI agents can automate migrations, reduce the value of exclusive telemetry by improving reasoning over available data, and make product evaluation fast enough for buyers to test and replace vendors more frequently.

BEFORE IT WAS OBVIOUS

The most successful founders stayed in the game long enough for their insights to compound.

LOST CONFIDENCE

Stop pretending you can quantify confidence and instead use techniques that work even when the future is unpredictable.

THE NEW RULES OF MEDIA WITH MARC ANDREESSEN & BEN HOROWITZ

AI has collapsed content production costs, shifting media power from institutional publishers toward individual creators and domain experts with direct audience relationships.

THE MOM-AND-POP SAAS ERA HAS ARRIVED

Creating software used to require access to a small, highly specialized group of people concentrated in a handful of tech hubs. This meant that only ideas large enough to justify the costs ever made it into the world. That complexity and cost is collapsing. Now, domain experts of every kind can start solving problems they understand deeply. There will likely be an explosion of products from unexpected places built by local experts with experience.

AI CODING TOOLS ARE QUIETLY RE-PRICING THEMSELVES

THE 37SIGNALS GUIDE TO MAKING DECISIONS

Companies are essentially made up of a group of people and a collection of decisions. How these people make these decisions is the art of running a business. This post contains a collection of general principles on how to make decisions. They serve as frames, considerations, and shared practices to draw upon when making a decision.

HOW STRIPE, GOOGLE, CANVA, CLOUDFLARE, AND HIGGSFIELD ARE ACTUALLY SELLING IN 2026

Top-performing tech companies increasingly use founder-led content, customer education, community building, and product visibility to drive demand before traditional sales engagement. The playbook emphasizes creating trust and mindshare at scale through content, events, and social channels rather than relying solely on outbound sales motions.

WHY AI FOR GTM HASN'T DELIVERED (AND HOW TO FIX IT)

AI promised to supercharge sales, and to be fair, it delivered the easy stuff. We can write great cold emails now, build account lists in seconds, and summarize every call. However, none of that actually wins deals. What wins is still the boring upstream call of deciding which accounts to chase and why now. Most teams are quietly letting an AI vendor make that call for them, so they end up with the same targeting as every competitor on the same tool.

pay attention. Learn more about sponsorship opportunities.

FRAMER 3.0

Design, build, and manage websites with AI-assisted content creation, collaborative branching workflows, and a community platform for sharing and distributing site resources.

SHIP VIBE-CODED APPS TO PRODUCTION

Retool is a platform that allows developers to build software on whatever AI coding platform they want and ship it through a single, governed runtime that enforces authentication, permissions, and data governance automatically. Developers can deploy any React app on Retool, and it will map the app's data connections to resources the team has already configured and secured. Apps can be deployed on Retool Cloud, in a VPC, or fully on-prem. Retool handles versioning, release management, environment promotion, access changes, and monitoring. Free hosting, plus bonus AI credits on every paid plan, is available until July 1.

COPILOT COWORK IS NOW GENERALLY AVAILABLE

Copilot Cowork executes complex, long-running multi-tool tasks. Users define the work, and Cowork runs it end-to-end and returns a completed result. The tool is designed to be accurate, secure, and low-cost. The feature is now generally available for users on the Microsoft 365 Copilot User Subscription License.

AI WORK NEEDS A PULL REQUEST

Human workflows usually have checkpoints, like a review, that make coordination possible. Agents need the same kind of containment. Agents can do much more than just write code. A pull request would give agents a familiar framework to work with humans. It would allow people to review changes, accept useful work, reject weak work, and leave a record for the future.

DECISIONS AND DOLLARS

The smarter the models get, the less your software is worth alone, so every app company now has to become a data company and a fintech company. The data half is judgment, the corrections your users make that no benchmark captures, which is why xAI took a $60B option on Cursor mostly to get inside the token flow, not for the editor. The survival rule for the "what if Anthropic builds this" question is to let agents read but guard the writes, because the decision being made right now is the only thing a rival can't scrape.

BUILDING AN AI-NATIVE STARTUP

This playbook remaps the four core stages of the startup journey (Idea, MVP, Launch, and Scale) according to the AI era.

ARE YOU BETTER AT RUNNING THE COMPANY OR RUNNING YOUR BOARD MEETING?

Founders often excel at running their companies but struggle during board meetings due to different skill requirements.

I SOLD MY AI STARTUP BEFORE REVENUE: HERE'S WHAT INVESTORS MISSED — AND FOUNDERS SHOULDN'T

AI startups should focus on resolving deep scientific challenges, as significant advancements will drive future technological frameworks.

CHAT-BASED ONBOARDING

Opal has introduced a chat-based onboarding UX that encourages users to interact with the product more.

Register for the free webinar →

HOW TO COMPETE IN "CROWDED" AI MARKETS

Decagon CEO Jesse Zhang argues every AI market feels crowded because only a few players can win, and walks through the brutal customer-service AI sales cycle where Decagon gets compared to Salesforce and Google adding AI, pre-AI Intercom, AI-natives like Sierra, and the customer's own MVP. Speed of shipping is an advantage but not a durable one, because AI tools have compressed build-and-copy timelines so much that a feature lost in a deal goes straight onto every competitor's roadmap.

AI CAN'T SURVIVE YOUR C-SUITE'S MAGIC THINKING

Corporate culture and agenda setting have failed engineering organizations. Data wrangling and cleansing have become more about making the data align with the business rather than making data more valuable to improve decision-making. Executives are relying too much on magic thinking. Shareholders and the board are now holding CEOs accountable for choosing evidence over comfort to right the ship.

HOW TO PIVOT WELL

The iconic stories of the Web 2.0 era give founders the hope that unimaginable riches are just an iteration away. AI makes faster pivots more possible. Knowing when to pivot, how fast, and toward what has become a core skill for founders in the AI era. This post looks at what founders need to think about when considering a pivot.

A TECHNICAL GUIDE TO BUILDING YOUR OWN LEARNING LOOP

The winning move in AI now is to build a learning loop on top of the best model. This allows data and usage to compound into IP that nobody can rent back to you. This post looks at why this should be done and provides a blueprint for how to create a winning learning loop.