New research shows AI doesn't reduce work—it makes you want to do more of it
by Katie Parrott It was lunchtime on a Friday, and I was teaching my new OpenClaw AI assistant, Margot, to manage my to-do list. I’d carved out the afternoon to get her configured. A few hours, tops, I told myself. Then I’d do something else with my evening. Twelve hours later, Margot and I had written and rewritten two essays, rebuilt my personal website, and added features to another app I’d been tinkering with. I finally tried to go to sleep around 1 a.m., but the next thing I knew, I was launching myself out of bed, rushing to my desk, and typing in all caps: “OH MY GOD, MARGOT.” You might know the feeling. Maybe you’ve stayed up too late pursuing a project that started as a quick experiment, or caught yourself prompting during lunch or in the last few minutes before you told yourself you’d be done for the day. ”One more prompt” turns into 50. “I’ll just fix this one bug” turns into a vibe coding marathon. Time flies when you’re having fun with AI. It also flies when you think everyone else is getting ahead without you, because every hour you’re not learning feels like a week you’ve fallen behind. AI is changing work—and I don’t mean how we’re working, although it’s changing that, too. I mean how the work feels in your body at 1 a.m. when you can’t stop, or at 9 a.m. when you’re afraid you haven’t done enough. Technology has been blurring the line between work and life for decades, but the old tools pulled us back through obligation—you checked that email at 10 p.m. because it felt like you had to, not because you wanted to.
Enterprise app generation that works with your stack
“Hi, Tomasz or Tomasz’s agent.” I’ve started receiving emails that begin this way. A byproduct, I suppose, of having written so much about AI. People now assume my inbox is monitored by robots. Which raises an odd question : what does it mean to write to someone when you expect a machine to answer? Gmail suggests my reply before I’ve thought it. “Sounds good!” “Thanks for sending!” “Let’s circle back next week.” The machine knows what I’d say. Sometimes I click it. Sometimes I wonder if the person on the other end can tell. Every customer support call is now with an AI agent. The voice sounds real. They are infinitely knowledgeable. The responses are fast. Does it matter that it’s not a person? A friend sends voice memos instead of texts now. “So you know it’s actually me,” he said. But how do you know? ElevenLabs can clone a voice from thirty seconds of audio. The ums, the pauses, the little laugh—all reproducible. Does it matter? But maybe the people writing “Hi, Tomasz or Tomasz’s agent” have it right. They’re not being rude. They’re being realistic. They’ve adapted to a world where the answer might come from either side of the curtain, & they’ve decided not to care which. The polite thing now is to assume the robot. The intimate thing is to be surprised when it’s not.
We’re tracking company launches as they happen and surfacing the most interesting new founders and startups (yes, all outside of this stealth activity too). If you want to stay ahead of the curve, this is where you’ll find them first.
Right now we’ve got 5 subscribers: my wife, my mom, and that high school buddy who won’t stop pitching me his app. Be smart like them and get in early 👇
We run a live feed inside Gravity that tracks founders entering stealth and companies quietly exiting it.
What you’re reading here is about 1% of the stealth activity we pick up. The full tracker updates in real time as things change and new activity emerges.
In this issue of the Stealth Startup Spy, here is what we will uncover:
Former Global Head of Aerial Imagery & 3D Mapping at Google is building a real-estate intelligence platform aggregating neighborhood-level data to help buyers make smarter property decisions
Ex-Meta Superintelligence Labs scientist launches a stealth startup
Bardeen co-founder is building AI agents that observe how teams operate across tools, reconstructing workflows to identify automation opportunities
Serial founder and former Palantir consultant who later led product at Axon is launching another company in stealth
After selling Athena to Grindr and later leading product there, this operator is now building an AI-native performance management platform that evaluates employees using real work signals instead of traditional reviews
And more…
Now let’s shine the spotlight… 💡💡💡
🕵️♂️Founders Coming Out of Stealth
Real-time updates from founders who debut what they’ve been working on under stealth mode
Prior Experience: Ex-Co-founder & CEO at Lingumi (acquired by Novakid), ex-Chief Product Officer / Product Director at Novakid, Entrepreneur First (EF5), University of Oxford
Rig builds an AI-native data automation platform that learns a company’s data warehouse and enables teams to run complex operational workflows directly on top of internal data.
HQ: United Kingdom
Industry: Data Infrastructure and Analytics | Team Size: 3
FounderDNA: Serial Founder, Technical Founder, Doctorate Degree, Former FAANG, Top 10 University
Prior Experience: Co-founder at Bardeen, ex-Senior Director of Product & Engineering at Mesosphere, ex-Engineering Director (Distributed Computing) at Qualys, ex-Software Developer & Senior Fellow (Software Group) at CERN
WIQ builds AI agents that observe how work actually happens across tools and teams, reconstructing execution flows to identify automation opportunities and measure the impact of deployed AI.
FounderDNA: Serial Founder, Masters Degree, Top 10 University, Prior Exit
Prior Experience: Ex-Founder & CEO at Athena (acquired by Grindr), ex-Senior Product Manager (New Verticals) at Grindr, ex-BCG, ex-Goldman Sachs, Duke University
Brava builds an AI-native performance management platform that captures real examples of employee work to enable continuous feedback, fairer evaluations, and faster talent development.
Herd builds an AI-powered collaborative travel planning app that helps groups coordinate trips, vote on plans, manage bookings, and split shared expenses in one place.
iHuus builds a real estate intelligence platform that aggregates neighborhood-level data to help buyers make more informed home purchasing decisions.
HQ: Switzerland
Industry: Technology, Information and Internet | Team Size: 3
Time Spent in Stealth Mode: 2 Months
🕵️♂️Key Talent Going Under Stealth
Illuminating clues left behind by world class talent and influential innovators who just went into stealth mode
Jason Wu - Co-Founder & CTO at Stealth AI Startup
FounderDNA: Technical Founder, Masters Degree, Former FAANG
Prior Experience: Ex-Staff Applied Research Scientist & Tech Lead Manager at Meta, ex-Software Engineer at Google, ex-Applied Research Scientist at Sanas, Cornell alum
Prior Experience: Ex-Staff Software Engineer at Ramp, ex-Senior Software Engineer at Ramp, ex-Infrastructure Engineering Intern at Facebook, ex-Software Engineering Intern at Apple
Prior Experience: SOF veteran, Ex-Founder & CEO at Native, ex-Founder & CEO at Viceroy, ex-Founder & CEO at Guildsmith, ex-Principal Product Manager at Axon, ex-Consultant at Palantir Technologies
🚨Here’s the deal 🚨This email has gotten too big. Exciting, but with more people following it, the edge diminishes. I’ve thought long and hard about what to do to preserve the value in the signals. I’m not sure about the final direction yet, but in the meantime I’ve been sending an email 48 hours earlier to a select group of paid subscribers. The feedback has been pretty positive so I’m going to open up the list for another 100 spots. To get signals early, Apply here!
Stay Stealthy,
Drake
Thank you for reading. If you liked it, share it with your friends, colleagues and everyone interested in staying ahead of the hidden developments in tech. Subscribe below and follow us onX / Twitter to never miss a company operating under stealth again.
Stealth Startup Spy is a data-driven newsletter for investors, journalists and tech enthusiasts interested in uncovering the next big move for key talent, real-time stealth company launches and technology advancements not in plain sight. We leverage the technology built at Gravity to shine a light on the hidden world of stealth startups.
ben's bites · Tuesday, March 10 2026 · 8 min read · ↑ top
workshop recording inside
Hey folks,
The ‘become a builder’ workshop last week went well-ish 😊 (Codex crapped out on us). The recording is available, but I’m working on a thorough guide to cover everything properly (plus the bits we didn’t get to cover). I’m ~50% through it so hope to have it out this week.
Also, Factory is hosting a hackathon this thursday, everyone gets 200M tokens, and a mac mini is on the line.
OpenAI released GPT 5.4 in “thinking” and “pro” variants. It brings the coding power of GPT-5.3-Codex to the main model series, with better vision, tool use efficiency and a context window of 1M tokens. It’s now much better at computer use (see demo) and financial tasks. It’s also a bit more expensive vs GPT-5.2 ($1.75/$14 → $2.5/$15 per million input/output tokens). OpenAI expects to keep this naming and capacity difference between instant models (GPT-5.3 Instant) and reasoning models moving forward.
More from OpenAI:
ChatGPT for Excel - An extension to use ChatGPT in a sidebar right from your workbooks.
Codex Security, an AI app security agent evolved from Project Aardvark—free for a month to Enterprise customers.
Codex for Open Source - A program for open-source maintainers, giving them 6 months of ChatGPT Pro with Codex, conditional access to Codex Security and API credits.
It’s also acquiring Promptfoo, an open-source AI security testing tool (popular among Fortune 500, stays OSS).
New built-in skill in Claude Code -/loop lets you schedule recurring tasks in a single session, for up to 3 days at a time. Plus, you can now schedule tasks using Claude Code Desktop - these tasks run regularly as long as your computer is awake. They also launched a community ambassadors program for Claude.
For enterprises, Anthropic released Code Review by Claude and Claude Marketplace. The review tool uses a team of agents to review every PR and, on average, costs $15-25 per review. The marketplace lets enterprises consolidate their AI spending by using their Anthropic commitments to pay for other AI apps like GitLab, Harvey, Replit, etc.
Karpathy releasedautoresearch — agents autonomously iterate on LLM training code. Ran 2 days on 8xH100, found 20 real improvements with an 11% speedup. 630 lines, single-GPU, open source. I assume this approach of agents coming up with ideas and implementing them will see much more activity this year.
Yann LeCun, Meta’s ex-Chief AI Scientist, along with other researchers, has raised over $1B at a $3.5B valuation for their new startup, Advanced Machine Intelligence (AMI Labs). They are already operating from Paris, New York, Montreal and Singapore with a strong focus on world models and research that goes beyond LLMs.
Go stackless and get back to selling. Remember when selling meant talking to people? Before the tab-switching and endless sync errors. Reevo brings it all back to one platform. Prospecting, calls, pipeline, and reporting all in a single tab. From prospect to close. Go Stackless. reevo.ai*
🌐What I’m consuming
Cursor’s third era - Cloud agents have overtaken tab autocomplete in the IDE.
Building for trillions of agents - They will need their own infra, access to files, identities, while maintaining security, compliance, and governance.
Anthropic is suing the DoD to block its supply chain risk designation, calling it unlawful. Meanwhile, the White House is preparing an executive order to formally ban federal agencies from using Anthropic’s tools.
We run a funeral home and a medspa platform—not exactly Y Combinator darlings. Here's what we've learned about implementing AI in the real world.
by Sam Gerstenzang Sam Gerstenzang. Every illustration. The AI hype cycle has mostly rewarded software companies, butSam Gerstenzangis betting on the opposite: operationally complex, real-world service businesses—funeral homes, a medical spa platform. Fresh off an appearance on our podcastAI& I, the Boulton and Watt partner and former Stripe product leader shares four hard-won lessons for injecting AI into Main Street businesses, including why humans remain stubbornly hard to replace. If you enjoy the piece, watch his episode on X or YouTube , or listen on Spotify or Apple Podcasts.— Kate Lee__ I imagine it wasn’t easy being Costco’s management in the late 1990s. When dot-com darlings like Webvan, Amazon, Pets.com, Kazoo, and eBay—some still with us, others not—emerged, many doubted the warehouse retailer would survive the digital age. But Costco carried on. I started an incubator called Boulton and Watt, where we start a new software-enabled business every year or two that looks nothing like a Y Combinator darling and a bit more like Costco. Our portfolio includes Meadow, a contemporary funeral home with no physical locations (we use wedding venues during the day), and Moxie, a platform for nurses starting aesthetic medicine clinics (which just raised a $25 million Series C funding round). These are high-margin businesses in regulated industries—operationally complex, outside the San Francisco zeitgeist, and typically more likely to attract private equity than venture capital, which favors asset-light, high-growth tech businesses. For a while, the AI hype cycle made us look even more counter-trend. But now, as software gets increasingly easy to build and public SaaS companies’ share prices are down more than 70 percent, entrepreneurs and investors are looking for companies that will continue to be resilient when anyone can build software. Our thesis is that real-world service-based businesses are going to continue to flourish, and being on the cutting edge of AI will matter as much for the next Costco as it does for the next Lovable. We’ve had real successes, long journeys, and total flops as we bring AI to Main Street. Here are four lessons from bringing AI to a funeral home, a medical spa platform, and an incubator.
A single database for your entire AI stack
1-From three months to three weeks: AI accelerates research
We incubate new businesses with the same steps each time: Pick a market, interview people in that market, and test a solution. Each one of these steps is now AI-assisted, shaving days or weeks off of how we would have run this process previously. In our most recent search for our third company, we used:
ChatGPT to pull stats like how fragmented and large a new market was, pull data from public companies to understand margin profiles, and then suggest similar markets for the things we were looking for
Data-enrichment platform Clay to find relevant customers and kick off a series of targeted emails, which pointed to a landing page that Claude Code had spun up in an hour
Granola to record all early lead calls, and Claude to extract ideas, sentiment, and open questions from the transcripts
When we had interest from potential customers, we’d screen-record the customers’ existing workflows, and use Claude to analyze how the time was spent and opportunities to build new products, and have those products ready to test the next day. Each step could have taken weeks before. Using AI not only allowed us to move faster, it also reduced the cost of each tiny pivot required to find product-market fit.
2-Reward outcomes, not AI usage
We tried to implement “AI initiatives” in the companies we founded, but the first attempt often fell flat and became a “check the box” activity. We struggled with two challenges. First, AI became an excuse for team members to outsource their judgment to the hallucinations of a madman. If you don’t hold people accountable for the result—not just the method—the critical thinking you hired them for goes out the window. Your team needs to understand that using AI to generate a bunch of bad copy isn’t a good thing. The use of AI itself is not the goal—using AI only matters if it can improve the quality and the speed of the work. Hallucinations are your problem, not the LLM’s. And second, AI often requires re-thinking what the work can be, and real examples help show what is possible. Without practical examples of how AI tools could be implemented across the business, the use of AI was incremental rather than transformational. AI usage was limited to typing into ChatGPT and pasting the result into a spreadsheet. Here’s one example: It was natural for analysts to use AI to clean up spreadsheets or write formulas. But the breakthrough came when a team member used Claude Code to create an entirely new, interactive map-based approach for looking at customer density data. Prominently highlighting this example allowed everyone else on the team to better understand AI’s possibilities and how it could help them solve their own problems. Similarly, for our engineering team, the first incremental value of AI coding was simple: autocompleting code. But to drive change, we had to teach our engineers how to think differently. It used to be that a senior engineer would scope and break down a complex project into small chunks that a more junior engineer would take on. Now the junior engineers need to be trained to think like senior engineers to build out a more detailed plan that AI agents build.
3- Customer acquisition: The same story, but different this time
But so far, it looks a lot like Google search did: people come with high-intent queries, and businesses play the same cat-and-mouse game trying to show up in results. . It’s not a new paradigm—it’s another channel. For nearly two decades, this was the story of SEO: gaming Google’s algorithms to drive more traffic instead of paying for ads. Similarly, with “Generative Engine Optimization, ”companies are paying to fake interest on Reddit that will filter into ChatGPT. But this opportunity will come to an end much faster: OpenAI has started selling ads on ChatGPT , and showing up in results will quickly become pay-to-play. Meanwhile, Google’s search revenue is up 17 percent over last year. So much for the end of search! AI-powered lead generation tools like Clay, using signal data from sources such as LinkedIn, are flooding inboxes—you’ve probably noticed a rise in email spam over the last year. We think this will be short-lived: Either you will get so many emails you’ll learn to ignore them, or email providers will get smarter and hide them for you. The lesson for businesses like ours is to treat every new AI channel the way you’d treat any new channel, period. Test it, measure it, and don’t bet the farm on it just because it has “AI” in the name.
4-Humans are harder to replace than you think
Thirty years ago, I believed the internet and computers would eliminate manual paperwork. But it hasn’t happened yet, and the reasons why tell us a lot about AI adoption. At Meadow, our software emails dozens of insurance claim PDF files a day to insurance companies. The data from these PDFs are then manually entered into their systems. It was an inefficient process even before LLMs arrived, but it’s the only way we can work with these insurance companies. The insurance partners aren’t dumb—the investment of creating, implementing, and moving over customers to an electronic system simply doesn’t outweigh the benefit for them, even though it makes my software engineering brain go crazy. We’ve also worked with many small businesses that avoid automation for a different reason: They enjoy some of the repetitive work. And even if they saved time, it’s not obvious where they’d invest it. Then there’s a third barrier: Humans are easier to forgive. Moxie operates a receptionist service for hundreds of aesthetic medicine clinics, powered by dozens of humans picking up the phone to book appointments. We thought it would be a perfect use case for AI—replace this service with voice and chat, and reduce the cost. But when we shared the AI-based service, we saw greater churn, even when the human offering was far more expensive. Both made the same number of errors, but customers were much more forgiving of the humans. The pilot taught us a broader lesson: Getting AI to work in a test is easy. Getting it to work reliably inside your business is a different problem. Even though most of our code is now AI-generated and human-reviewed, the surrounding work—discovering edge cases, rolling out releases, and maintenance—still requires engineering time. And despite the headlines, good engineers are still expensive. So many tasks that seem like obvious AI candidates—like a landing page—only get done if the person who needs them can finish the whole thing without engineering help. The moment it requires integration into our real systems, it goes to the bottom of the queue.
The bridge between AI and the real world
Costco survived the dot-com boom by being relentlessly good at what it already did and what its customers needed. It adopted technology when it served its customers better, not because everyone else was doing it too. That’s our stance with AI. The goal was never to use it—it was always to use the best tools to run great businesses that serve their customers.
Lean more about how Gerstenzang builds businesses at Boulton and Watt on theAI& Ipodcast. Watch on X or YouTube , or listen on Spotify or Apple Podcasts_. _
Stop asking “What’s your biggest pain point?” in customer discovery
First Round Review · Tuesday, March 10 2026 · 2 min read · ↑ top
This week, we’re back with another installment of our Paths to Product-Market Fit series with Serval, an AI startup taking a swing at a hundred-billion dollar ITSM incumbent.
Serval's Path to Product-Market Fit — Win Enterprise Buyers by Treating Them Like Consumers
When hunting for startup ideas, Jake Stauch opened with the textbook discovery question: “What’s your biggest pain point?” It got him nowhere.He’d had dozens of conversations with IT buyers, a persona he’d spent a lot of time with as a product leader at security platform Verkada and wanted to build for at his new startup.“Nowadays people have mostly solved the problems they're aware of. They've already got some tool in place,” says Stauch. “I did a lot of interviews where I’d ask, ‘What keeps you up at night?’ And I just didn’t hear anything very interesting.”So he swapped in a new question: “If you could hire somebody today to sit next to you and do your work for you, what would you have them do?”“When you frame the question as, ‘Hey, if you had somebody else here to help you, what’s the work that you'd give them?’ That's a nonjudgmental way of asking for pain points because you're saying, ‘What would you push over to this new person?’” says Stauch. “That way, they can be much more free to say, ‘I don't like to do these things or I am doing a lot of this and I think somebody else could do it for me instead.’”The answers to that question sparked the idea for Serval , an AI platform that automates help desk requests and other IT workflows.Stauch had these discovery conversations in April 2024, while still at his day job at Verkada. Serval’s now a billion-dollar startup that nabbed a $75M Series B just one month after announcing its Series A, with customers like Notion, Clay and Vercel.On The Review, Stauch shares his biggest PMF lessons two years into building, before the early decisions blur into a glossy timeline.| | Continue reading on The Review
One of my favorite days of the year is the NYC Computer Science Opportunity Fair. The idea behind the CS Fair is pretty simple: Put students together with companies, universities, and organizations that care about technology and opportunity and let the conversations happen. The energy in the room is always incredible. The CS Fair has been going on for thirteen years now, and as a result, over 25,000 NYC public school students have been exposed to the opportunities that a career in tech in NYC offers.
This year's Fair will take place at the 168th Street Armory in Washington Heights on April 21st. We are expecting 2,400 high school students who will walk the track meeting engineers, founders, college representatives, and nonprofit leaders and learn about career paths and continued learning opportunities. More importantly, they will see how the coding and AI skills they are learning in their classrooms today can turn into real opportunities tomorrow.
That kind of exposure matters a lot. When students can see themselves in the field, it changes what they believe is possible.
New York City has made a big commitment to computer science and AI education in public schools over the past decade. The CS Fair is one of the places where you can really see the results of that effort.
Your team will leave inspired.
I always do.
Support AVCI am a VCShow you appreciate this writer, help support their work, and share in their growth over time by buying their writer coin.Support
“Since last November, 100% of my code has been written by Claude Code. I have not manually edited a single line, shipping 10 to 30 PRs per day.”
Boris Cherny, creator of Claude Code, ships 20-30 pull requests per day. Major code changes, not typo fixes. He runs five parallel AI instances, each on a separate branch.1 Compare that to a traditional engineer : 3 PRs per week.2 Cherny isn’t 10% more productive. He’s 30x more productive. That productivity gap compounds at the company level. Anthropic generates ~$5 million per employee.3 Cursor, $3.3 million. Midjourney, $2 million.4 Traditional SaaS considers $200-300k strong. A 10-20x difference. One explanation : communication overhead. The math follows Metcalfe’s Law.5 Each new team member adds n-1 new connections. Coordination drag doesn’t grow linearly. It explodes. Now consider what AI does to this equation. A traditional 150-person organization runs four layers deep. The org chart creates 11,175 potential communication channels. Meetings multiply. Alignment decays. An AI-enabled team producing equivalent output might need 30 people. Communication channels drop to 435. A 96% reduction. This is one reason AI-native startups are pulling ahead, and why building AI companies feels fun. The advantage comes from organizational structure. Fewer humans, fewer channels, faster iteration, compounding speed.6 R&D adopts this fastest. AI writes the code. Human communication becomes the bottleneck. The span of control debate shifts from “how many people can one manager oversee?” to “how many AI agents can one human orchestrate?” Small teams have always paid less coordination tax. AI cuts it further.
1. Cherny, Boris. Claude Code creator landed 259 PRs in 30 days, Hacker News, 2025. ↩︎
2. Seporaitis, Julius. What Can 75,000 Pull Requests Tell?, 2021. Median developer opens 3 PRs per week; consistent with Google’s internal data. ↩︎
3. Estimated from Anthropic’s ~$20B revenue run rate (Bloomberg, March 2026) divided by ~4,300 employees (LinkedIn). ↩︎
4. Dealroom estimates. AI startups revenue per employee : Cursor $3.3M, Midjourney $2M, OpenAI $1.5M per employee. ↩︎
5. Metcalfe’s Law, Wikipedia. ↩︎
6. How to start a Lean, AI-Native Startup in 2025, Henry the 9th, 2025. ↩︎
by Dan Shipper If you use AI seriously to think or work, a surprising share of your documents are probably already written by agents. Most of us at Every are using Codex to generate plan documents, Claude Cowork to write research reports, and OpenClaws to create strategy memos. But the current process for collaborating and iterating on agent-generated writing is weirdly primitive. It mostly takes place in Markdown files on your laptop, which makes it reminiscent of document editing in 1999. Because these documents are stuck on your laptop, it’s hard for other agents to help iterate on them—and they’re hard to show to your team, too. That’s why today we’re officially launching Proof, an online document editor built for agents and humans to collaborate. Fast, free, and no login required. (And it’s open source!) Try Proof
What is Proof?
Most word processors still assume a human is doing the writing and AI is helping at the margins for brainstorming, making rewrite suggestions, or producing a first draft. Proof flips that around. It’s a document editor built for the kinds of documents agents are increasingly writing: bug reports, product requirement documents, implementation plans, research briefs, copy audits, strategy documents, memos, and proposals. It supports live edits with multiple collaborators, allows you to see and leave comments, and lets you track changes. But it also has what you would expect of a document editor built for agents. Proof:
Is agent-native: Anything you can do in Proof, your agent can do just as easily.
Tracks provenance: A colored rail on the left side of every document tracks who wrote what. Green means human, purple means AI.
Is login-free and open-source: We want Proof to be your agent’s favorite document editor.
You can try it yourself right now. Give the prompt below to your agent of choice and have it generate a Proof document for you.
Try Proof with your agent
Send your agent this link: Try Proof with your agent It will ask them to write a Proof document describing what they’ve learned about how to work with you best. You’ll get a Proof link with their insider’s perspective on your work style.
How we use Proof at Every
We’ve been using Proof internally for over a month and, a couple of weeks ago, opened it up to Every subscribers as an experiment. Internally, we’ve found that plans, like the ones generated when Claude Code or Codex maps out a new feature for a product, are the most common type of Proof document, along with strategy documents. Austin Tedesco , Every’s head of growth, whose job involves writing a lot of campaign strategy documents, might go back and forth on the drafts with one or more agents. When the document feels ready, he’ll send it to the rest of the team on Slack for feedback. Austin also uses Proof for his own writing. He sends notes for his food newsletter throughout the week to his Claw, which organizes them in a Proof document and generates an outline, so that when he sits down to write, everything is already in front of him. I use Proof documents as my daily to-do lists. I pin the page for that day in my Slack account, and anytime a new task pops up for me to do, I’ll tell my Claw, R2-C2, to add it to the document in the appropriate section. At the end of the day, I can see which tasks I’ve completed and which I haven’t, and I can tell my agent to either push my incomplete tasks to the next day, help figure out how to complete them, or just go ahead and do them for me. Our editorial team is also exploring how to integrate Proof into their AI workflow. The initial outline and draft of this post started in Proof.
What’s available at launch
Today, Proof is available to everyone, for free, even if you don’t have an Every account. Those who do sign in with an Every account have access to a continually updating library with all of the Proof documents they’ve created. And it’s the first one of our products to be open-source, so that anyone can inspect it, build on it, and help shape where it goes next.
Try Proof
Most of the writing that will happen in the next decade will be generated by AI agents. Now they have a document editor built for them. Watch me talk about Proof with Every’s Brandon Gell , Kieran Klaassen , and Austin Tedesco on X or YouTube , or listen on Spotify or Apple Podcasts on the latest episode of AI& I. You can also read the full transcript. Try Proof
AI eliminates the marginal hire. Tech job openings are down 45% from the 2022 peak, but up 16% since the start of 2026 - from 227k to 264k. Why the narrative violation? Companies are hiring again, just fewer people than before. A reset to a lower baseline. A team that would have added two engineers to hit next year’s roadmap now ships with the headcount they have. Cursor, Claude Code, Copilot close the gap. The job postings never go live. The offers never extend. Inside most organizations, headcount stays flat. No layoffs. No restructuring announcements. Just fewer new hires than planned. Block slashing 40% of its workforce showed what happens when a company acts on this logic all at once. Jack Dorsey explained : “Intelligence tools we’re creating & using, paired with smaller & flatter teams, are enabling a new way of working which fundamentally changes what it means to build & run a company.” Most companies won’t restructure so dramatically. Until an economic shock, a missed quarter, or pressure from the board forces the question. What AI made possible, AI makes necessary. The restructuring that might have happened gradually over five years happens in one quarter. The seismic shock isn’t coming out of nowhere. It’s building invisibly, one unposted job at a time.
ben's bites · Thursday, March 12 2026 · 6 min read · ↑ top
web access CLIs, sandboxes and another openclaw clone
Hey folks,
Google released Gemini Embedding 2 , and it is multimodal, so you can embed text, audio, images, video and PDF documents using the same model. It’s a little expensive compared to other options in text, but videos at low fps and audio are really cheap with the unmatched feature of embedding them all at the same time. This should open a lot of startup ideas that are basically “search over a large amount of non-textual data.”
Replit released its Agent 4 with multiple parallel agents, live collaboration with teammates, and an interactive design canvas that both you and the agent can edit on. Agent 4 can make more than just web apps; it can create animations, slides, mobile apps, data visualisations, and more. All of it is possible in a single project. Plus, Replit raised $400M and is now valued at $9B.
Perplexity teased Personal Computer - They say it’s always on version of Perplexity Computer with access to your files, apps, and sessions through a continuously running Mac mini. That sounds kinda like openclaw, doesn’t it?
Async Voice API is a human-like, low-latency text-to-speech API for real-time apps and agents. 15 languages, streaming-ready, integrations with n8n, LiveKit, Twilio, and more. Top-ranked on the Hugging Face TTS Arena. From just $0.50/hour with a 24/7 SLA. Try it now.*
NVIDIA plans to spend $26B over the next 5 years to build the world’s best open-source models. They just released Nemotron 3 Super - 120B params (12B active) model with similar performance to GPT-oss 120B and Qwen 3.5 122B
Two new interesting benchmarks:
PostTrainBench - Measuring how well AI agents can post-train language models
RuneBench - Long-horizon goal optimisation across 14 AI coding models inside Runescape.
The Anthropic Institute - New team from Anthropic with a focus on communicating the impact of AI to the world.
Runway Labs - A generative AI incubator to explore use cases of AI video and general world models.
Every · Thursday, March 12 2026 · 7 min read · ↑ top
New research on writing style reveals that the most distinctive parts of your prose are the ones you don't even think about
by Marcus Moretti TL;DR:Why does AI writing still sound like AI writing, even as the models get smarter? In his first piece since joining Every asSpiral’s general manager,Marcus Morettiexplains why the answer is more complicated than you’d think. The most reliable fingerprints of your personal style come from the words you write subconsciously: articles, pronouns, and function words that emerge in a distinctive pattern as you focus on the meaning of a sentence. His piece explores what new research in machine learning and stylometry—the study of style—means for the future of writing tools like Spiral. If you want to go deeper, Spiral has several updates, including creating a writing style from your website or X account (even taking post engagement into account) and a cleaner, faster editor. — Kate Lee OpenAI models demonstrate Ph.D.-level knowledge across physics, biology, and chemistry. Anthropic staff have claimed its Opus 4.5 model “largely solved coding.” Yet AI writing remains stubbornly detectable: “It’s not an idea. It’s a breakthrough.” “Delve.” Lists of threes with no “and.” If you’re a regular Every reader, you may already know why this is. LLMs are trained on an unfathomable amount of words and learn generally how to speak. Post-training, which refines a model after initial training on large datasets, makes the models friendlier and safer, so they end up speaking in a kind of generic politeness. Ted Chiang ’s description from a few years ago remains apt: “ChatGPT is a blurry JPEG of the web”—a tool that approximates human insight without ever landing on the mark. I’m interested in the relationship between LLMs and writing style because I’m the general manager of Spiral , Every’s AI co-writer. Writing sessions in Spiral begin as a chat: You describe what you intend to write, and Spiral helps you hone your message and gather relevant research. Then it produces one or more drafts, offering several approaches for your piece. Our aim is for Spiral’s written output to reflect your personal writing style, not the generic politeness of the foundational model. To this end, I’ve been reading papers on natural language processing, linguistic forensics, and stylometry—the study of writing styles. It wasn’t until I started working on Spiral that I became aware of the century-plus history of stylometry, or of the fastidiousness with which researchers have catalogued the elements of style. In recent years, researchers in these fields have flocked to LLMs, finding new ways to expand our understanding of human writing. Here are some findings that I found interesting and even counterintuitive, and that provide a hint as to where AI writing might be headed.
Looking for an AI notetaker for your meetings?
Subconscious decisions define writing styles
Stylometry has had a few moments of glory. In the 1800s, stylometrists gave sold-out lectures about whether William Shakespeare wrote those plays. In the 1960s, two stylometrists isolatedAlexander Hamilton ’s contributions to TheFederalist Papers based largely on the presence of the word “upon.” In the 2020s, LLMs have introduced new ways of studying style. Last year, two Cornell University researchers systematically manipulated text snippets to see how it affected LLMs’ ability to guess their authors. They removed an attribute of the text one at a time—such as proper nouns or capitalization—and measured the effect on attribution accuracy. They found that removing the more functional features of the text caused the models to misattribute authorship more often, proving that those features are most helpful for attribution. In particular, removing “stop words” made it a lot harder to guess who wrote something. In natural language processing, stop words are common, functional words like articles (“a,” “the”) or pronouns (“I,” “she”). These words are often filtered out of text analysis because they don’t convey much meaning, but it turns out that they appear in patterns that can help identify who wrote something. This is why Hamilton’s use of “upon” tipped off those researchers to his Federalist contributions. Things like stop words and word order turn out to be some of the most distinctive markers of someone’s writing style. These purely functional aspects of writing mostly reflect subconscious decisions. When we write, we focus on choosing meaningful words, and our subconscious tends to fill in the rest. But the way our subconscious contributes to our sentences is to be distinctive.
Most people are wildly inconsistent writers
“How Well Do LLMs Imitate Human Writing Style?” asked another paper co-authored last year by a researcher at Bucknell University. The authors used various methods to get an LLM to copy someone’s style, finding that just a few writing samples increased style fidelity over the base output by about 23 times. A smattering of examples went a long way. They then tested whether AI text could be identified as such even if it accurately reproduced someone’s style. They discovered it can be. Natural language processing has something called a “perplexity” score, and the greater the linguistic variance, or the diversity and unpredictability of word choice and sentence structure across a writing sample, the higher its perplexity. The researchers found that, on average, humans are twice as varied in their writing as machines.
Writing styles change dramatically over time
An LLM is just a string of numbers. When you interact with ChatGPT or Claude, you’re “talking” to a static set of digital files. LLMs are point-in-time snapshots of human language, which is why they need to search the web for information after their training cutoff date. Language itself, however, rapidly evolves. In the book Algospeak , the self-styled “etymology nerd” and linguist Adam Aleksic argues that our vocabulary is evolving faster than ever, due to social media and hyperconnectivity. This poses a problem for LLMs. What good is a model if its training run ended before we started saying “skibidi”? In October, Sushil Khairnar , a graduate student at Virginia Tech, tried to quantify models’ “temporal drift.” He found that GPT-2 and GPT-3’s manner of speaking lagged behind the general lexicon by about 15 percent a year after its release and 28 percent after two years. Ironically, LLMs themselves are altering language, including in academic research. Post-ChatGPT papers include significantly more AI-ese : words like “underscore,” “highlight,” and “showcase.” “Delve” is the biggest culprit, with usage in papers skyrocketing by more than 2000 percent between 2022 and 2024.
What’s next
As foundation models and methods of style transfer improve, computers will get better at mimicking individual writing styles. It’s an open question of how close they can get. The better we understand the science of style, the more we can bridge the gap between model output and manual output. The research guides Spiral’s roadmap. As an example, we recently updated how Spiral generates a style guide from a user’s writing samples, which then stylizes Spiral’s drafts. We previously generated hundreds of descriptions of the user’s writing, but now we focus on the key textual identifiers and quintessential phrases from the source material. And we’re building connections to writing sources—blogs, newsletters, and social media—so Spiral can keep up with how your personal style evolves over time. For any LLM-generated writing, though, there will always be some gap—after all, a person didn’t write it. That fact may be harder and harder to detect in the output, but it’s always worth considering what your reader would think upon learning that your piece was AI-assisted or -generated. This piece, for example, was written the old-fashioned way, despite the em-dashes , which Every’s style guide allows. Unless you want to analyze the stop words, you’ll have to take my word for it.
__Spiral has been busy the past few weeks. New to the app: connect your Twitter account and get an engagement-weighted style guide, allowing Spiral to draft bespoke tweets for your audience. Link your website or RSS feed to teach Spiral your style via bulk post import. Workspaces now make it easy for you to share styles across your team, so you can write in one unified voice.
Drake Dukes · Thursday, March 12 2026 · 7 min read · ↑ top
Ex-DoorDash sustainability exec launches AI platform to accelerate energy transactions, Amazon/SpaceX hardware engineer enters stealth, & Former Meta researcher builds data agent for self-evolving AI
We’re tracking company launches as they happen and surfacing the most interesting new founders and startups (yes, all outside of this stealth activity too). If you want to stay ahead of the curve, this is where you’ll find them first.
Right now we’ve got 5 subscribers: my wife, my mom, and that high school buddy who won’t stop pitching me his app. Be smart like them and get in early 👇
We run a live feed inside Gravity that tracks founders entering stealth and companies quietly exiting it.
What you’re reading here is about 1% of the stealth activity we pick up. The full tracker updates in real time as things change and new activity emerges.
Co-Founders:Guillaume Nozière (Applied AI researcher at Meta), John Bragg (exited to CoStar and previously network security engineer at Capital One)
Vor Systems is an AI-enabled transaction platform designed to surface the material nuances of complex energy deals so teams can move viable projects toward the grid faster.
HQ: United States
Industry: Software Development | Team Size: 3
Latest Funding: $3M Pre-Seed Round on 3/5/2026
Key Investors: Gigascale Capital, Virta Ventures, Christopher Payne, Hank Couture, Badrul Farooqi, Joe Song, Paul Grana, and Titiaan Palazzi
FounderDNA: Serial Founder, Technical Founder, Doctorate Degree, Former FAANG, Top 10 University
Prior Experience: Ex-Applied Science Manager & Senior Applied Scientist at Amazon, ex-Deep Learning Researcher in Robotics & Autonomous Driving at Berkeley DeepDrive, South Park Commons member
Prior Experience: Ex-Product Manager at Stripe, ex-Lead Engineer at Recko (acquired by Stripe), ex-Senior Software Engineer at Practo, ex-Co-founder at Furriez
Prior Experience: Ex-Product Owner at Viessmann Climate Solutions, ex-Managing Director at OneClimate, ex-Venture Architect at wattx, ex-Management Consultant at Accenture, ex-Co-Founder at Awad Getränke
Routine Labs builds automation infrastructure for regulated environments, ensuring reliability, compliance, and full auditability for critical operations.
HQ: Germany
Industry: Technology, Information and Internet | Team Size: 8
FounderDNA: Technical Founder, Doctorate Degree, Former FAANG, Top 10 University
Prior Experience: Ex-Research Scientist at Meta, ex-Staff Research Scientist at Together AI, ex-AI Research Scientist at X (the moonshot factory), ex-Research Engineer at Microsoft, ex-PhD Student at Northwestern University
Analogy AI builds a data agent that automates the creation, validation, and delivery of high-quality training datasets for self-evolving AI, enabling repeatable, auditable, and continuously improving model training.
HQ: United States
Industry: Technology, Information and Internet
Time Spent in Stealth Mode: 6 Months
🕵️♂️Key Talent Going Under Stealth
Illuminating clues left behind by world class talent and influential innovators who just went into stealth mode
Matt H. - Founder at Stealth Startup
FounderDNA: Technical Founder, Former FAANG, Top 10 University
Prior Experience: Ex-CTO & VP of Engineering at Grounded, ex-Hardware Development Engineering Manager at Amazon Leo, ex-Senior Manufacturing Engineer at Amazon Project Kuiper, ex-Lead Propulsion Manufacturing Engineer at Astra, ex-Manufacturing Engineer at SpaceX
FounderDNA: Serial Founder, Masters Degree, Top 10 University
Prior Experience: Ex-Co-Founder & CEO at Lever, ex-Product Lead at Marathon Digital Holdings, ex-Product Manager at Airwallex and mesha, ex-Research Fellow at Harvard University
Prior Experience: Ex-Senior Director of Customer Experience Engineering at Snowflake, ex-Senior Director & Director of Engineering at Cloudera, ex-Engineering Manager at Conductor
🚨Here’s the deal 🚨This email has gotten too big. Exciting, but with more people following it, the edge diminishes. I’ve thought long and hard about what to do to preserve the value in the signals. I’m not sure about the final direction yet, but in the meantime I’ve been sending an email 48 hours earlier to a select group of paid subscribers. The feedback has been pretty positive so I’m going to open up the list for another 100 spots. To get signals early, Apply here!
Stay Stealthy,
Drake
Thank you for reading. If you liked it, share it with your friends, colleagues and everyone interested in staying ahead of the hidden developments in tech. Subscribe below and follow us onX / Twitter to never miss a company operating under stealth again.
Stealth Startup Spy is a data-driven newsletter for investors, journalists and tech enthusiasts interested in uncovering the next big move for key talent, real-time stealth company launches and technology advancements not in plain sight. We leverage the technology built at Gravity to shine a light on the hidden world of stealth startups.
In September 2024, Hurricane Helene flooded Baxter International’s plant in Marion, North Carolina, which produced 60% of the nation’s IV fluids. Within a week, more than 80% of U.S. healthcare organizations reported shortages. One plant, one flood, one week. That disruption made headlines. Most don’t. Eighty-five million packages arrived damaged in the U.S. in 2024, up 30% from the prior year, costing businesses $4 billion. Sean McCarthy saw those failures accumulate during his years at Amazon Shipping, where he was one of the early hires. The investigation process never varied. Query the warehouse management system, often two decades old. Cross-reference the carrier portal. Call the driver, who doesn’t pick up. File a claim: seventeen fields. Four hours pass. Sometimes the problem gets solved. The obstacle was fragmentation. A single shipment can touch 40 to 60 processes across multiple vendors. Connecting them would mean hundreds of bespoke integrations. The project never got funded. Sean partnered with Henry Ou, who led ML teams at Apple and built ranking systems at ByteDance. Together they founded BackOps, which deploys AI agents that read emails, click through portals, call drivers, and file claims. When a customer reports a problem, BackOps traces it across every system involved, escalating to a human only when a judgment call is required. Your browser does not support the video tag. Watch the video We’re leading BackOps’s $26 million Series A. The product works in two stages. Employees record their screens while solving problems; BackOps converts those recordings into automated workflows. Then Relay, the automation engine, runs continuously: filing claims, initiating reshipments, responding to customers. Customers report 93% faster response times and 60% time savings. BackOps files 100% of eligible carrier claims automatically. The platform serves a top global automaker, a leading retailer, major grocery chains, and industrial suppliers. Sean and Henry are targeting a $3.5 billion market growing 13% annually. The bet: AI agents can connect systems that were never designed to talk to each other. So far, the connections hold. If you’d like to learn more, reach out to Sean. Read more from Sean, Theory partner Andy, and Axios.
Scott Galloway here. If you’re getting this email, congrats – you’re subscribed to Prof G Media on Substack. Whether you’ve been a long-time reader of No Mercy / No Malice or the Prof GMarkets newsletter, or you’re a new subscriber (thank you), I’m happy you’re here.
We launched on Substack to house our most rigorous analysis, across all our properties. Prof G Media is where Scott Galloway (me), Ed Elson, Jessica Tarlov, Alice Han, and James Kynge will make you smarter about the forces moving markets, politics, and society.
I encourage you to get familiar with the full spectrum of our content. Customize your Prof G Media notifications here, and never miss an update from your favorite shows and newsletters:
Prof G+
I built (and sold) two consulting firms by telling companies to serve their most loyal customers. Prof G+ is us doing exactly that. Consider it your all-access pass to everything we make, in one place, for people who actually pay attention.
Prof G+ exclusives. I teach. My analysts dig. Together, we go deep into topics that are too nuanced for social media. Consider us your moat.
Backstage pass. Can’t get enough of the Dawg? Me neither. Get reaction videos, livestreams, and private chats with me and the Prof G team.
Community. Leave comments, make friends, and engage in a series of bad decisions that might pay off.
All this and more, exclusively on Substack.
Thank you, Minneapolis
We’re offering a discount on Prof G+ for the remainder of the month in recognition of the thousands of you who joined us, in Minneapolis and on our Substack livestream.
Our future livestreams will be limited to Prof G+ members. Next up, we're going live from SXSW on Saturday, March 14, at 2:30 p.m. CT / 3:30 p.m. ET with Rep. James Talarico. I hope you’ll consider subscribing.
Through March 31, new Prof G+ subscribers receive 15% off an annual plan, only at the link below:
You’ve been able to outsource software development to India for cheap for a very long time. If you had an idea and a spec, you could ship it overseas and get something back for a fraction of the cost. Why wasn’t offshoring the death of the software industry?
TLDR - code was always execution (not a moat) and LLMs are different in both kind and degree from outsourcing.
The outsourcing question
This is the obvious objection: if cheaper code was going to kill software, outsourcing would have done it already.
But outsourcing was a low-tech, shitty attempt to commodify code production. Huge barriers to entry, high cost to manage, bad results. It didn’t actually relieve the bottleneck.
AI code generation is a difference in both kind and degree.
The difference in degree: code generation tools are so much cheaper, so much faster, and roughly at par with anything you would get from an outsourced team. That massively expands the market.
The difference in kind: Outsourcing is basically useless to (and even a burden on) really talented engineers, whereas code gen/LLMs are a massive accelerant/multiplier on their ability to run fast and far.
Code becomes free at the low end (savvy non-technical operators solving problems with code) and amplified at the high end (great engineers being able to do more than ever before). Collectively, this compresses the value of code as a differentiator.
Software is more valuable than ever in that it’s more important. But the ability to do it well is no longer enough to separate you from the pack.
Code is/was a chokepoint on execution, not a moat
There are only a few moats and execution is not one of them (nor is speed), even if it matters a ton and makes some businesses better than the rest.
Writing application software was never a moat but it was more valuable (harder to replicate) as a form of execution. The reason it felt like a moat was that a limited number of people could do it well. That talent constraint made software development look like a barrier to entry when it was actually just a bottleneck on execution.
AI relieved the bottleneck, changing nothing about classical moats. More people can functionally do it well AND those who can do it really well can do much more of it.
Moats and execution are fully independent variables
Execution in the field of writing software is not fundamentally different than execution in the fields of hiring, selling, finance, etc. It matters enormously to how well a business runs, but it doesn’t create a durable barrier to entry or guarantee margins on its own.
Google Search is the definitive illustration. It is probably the best business of all time because it has multiple strong moats: data, network effects, brand, economies of scale. But the execution is notoriously terrible: no product direction, low velocity, bloated, kills products on a whim.
Moats determine the ceiling on how much value you can capture. Execution determines how close to that ceiling you actually get. They are fully independent variables.
The companies in trouble are the ones that confused the execution bottleneck for a moat. AI didn’t kill your moat if you never had one. It just relieved the chokepoint that made it hard for someone else to execute as well as you did.
Everything shifts down one tier
Execution on application-software-shaped problems gets easier, which means you can afford to focus on harder problems. Every company used to have to solve two things: the software problem and their actual hard problem. AI is knocking out the software layer, so now all your effort goes to the hard part. That makes each category one step more tractable.
For pure application software companies, there is no harder problem underneath. Software was the whole thing. When you knock out that layer, there’s nothing left to create barriers to entry. That’s why they look more and more like services businesses: easy to start, hard to differentiate, lots of competition.
But for everyone else, this is a gift. The infrastructure business that used to spend real effort on the application layer can now put all its resources toward systems engineering and reliability. The hardware company can focus entirely on the physical and mechanical problems that actually differentiate it. Same logic, repeated across every tier.
We’re seeing basically everything shifting down click one in terms of risk and feasibility. Pure software companies look more like services companies. Infrastructure companies look like software companies. Hard tech companies look like infrastructure companies in terms of difficulty, specialization, and the ability to start and fund them.
Even as application software becomes non-viable on a TV basis, the rest of the world and the TAM of software writ large is growing exponentially.
From now on there are only 4 jos
There’s a very real possibility that the only jobs in tech companies are going to be:
product eng/vibe coder/PM/slop cannon : self explanatory. this is the high velocity, high tool use generalist. they are obviously not restricted to product and eng roles. Anyone can be commercial and product minded.
security/SRE/infra : we’re going to be producing so much STUFF across every org that there’s going to need to be really really good people stitching it together, making it stable, secure, and robust.
hot people. You will find hot people in roles ranging from sales, to people, to CX. There will always be an important place for those who present an easy UX to the world and are pleasant to be around. Remember, there are many ways to be hot.
grown ups : sometimes you need an adult in the room to just say “hey, come on.” They are effectively a much needed governor on an otherwise accelerating organization. You will find them across roles but there are obvious places like legal and finance. They are basically the non technical equivalents of #2 and might even be able to be bucketed with that.
the latent traits have always cut across job titles and orgs.
For the last year we’ve been going deep on how AI is changing the threat landscape by creating new surface areas to defend and transforming attacker economics.
At the end of April we’re hosting ≈100 founders, security operators, buyers, and investors in NYC to discuss and learn.
This is gonna be fun. We’re bringing Slow’s etiquette school to NY in a few weeks to teach founders and builders in NY how to show up the right way. We’re gonna have a bunch of speakers and teachers to help you learn how to present like an adult.
My name is Yoni Rechtman. I’m a partner at Slow Ventures, where I lead pre/seed rounds from a ≈$325M fund. I’m a generalist investor looking for weird takes on important stories: N-of-1 companies taking non-obvious approaches to markets that matter. I’m interested in real world businesses, hybrid software companies, AI’s second-order effects, healthcare, network effects, and fintech. If you’re building something ambitious or think I’m wrong, I’d love to hear about it.
Hacker Newsletter · Friday, March 13 2026 · 8 min read · ↑ top
One way is to make it so simple that there are obviously no deficiencies, and the other way is to make it so complicated that there are no obvious deficiencies. //Tony Hoare
You're among 69,941 others who received this email because you wanted a weekly recap of the best articles from Hacker News. Published by Curpress from Bellingham, Washington. Hacker Newsletter is not affiliated with Y Combinator in any way.
✨ Want to promote your startup? Buy a classified ad or click reply to get our media kit
Every week I’ll provide updates on the latest trends in cloud software companies. Follow along to stay up to date!
In Defense of Model Lab Profitability
There seems to be endless debate around AI companies, and whether they have “upside down P&Ls” that will forever lose money, or if they will turn into cash cows in the future. Whether this sentiment is pointed at the large labs like OpenAI and Anthropic, or upstarts like Cursor, I hear it all the time! And I can’t tell if the bears just want to confirm their priors on AI negativity, if the bulls just have blind naive optimism, or if anyone really has a pov grounded in real analysis. As an early stage VC I certainly fall into the “perpetually optimistic” camp, so you can apply the appropriate filter to this post :) But for this post I wanted to focus on the profitability debate centered around the large labs, and why I think they’ll turn into wildly profitable business.
There’s three main parts to this debate. 1) gross margins, 2) training costs, and 3) retention / commoditization. The “AI labs will never be profitable” crowd will claim gross margins of labs are structurally capped, the hamster wheel of training larger and larger models (whether that’s a capex or opex expense) will always lead to FCF negative businesses, and the switching costs from one model to the next are super low. So (as the bears would say), even if you can solve one of those 3 issues, you’ll never solve all three, and thus these businesses will never be profitable.
To start, I want to set the stage by analyzing how the profitability of infrastructure software companies (like Confluent, Snowflake, Mongo, etc as well as the hyperscalers) have evolved over the years. It’s an imperfect analogy for so many reasons, but I think it’s helpful to lead with it.
Let’s start with gross margins, as this is the part of the debate I’m confident on (one could describe my views as “naively certain” :). Just about every infra software company started off with negative gross margins, that then scaled to something in the 60-70% range with scale. Why do infra companies typically start of with negative / very low gross margins?
You’re buying infrastructure at retail and selling at a discount. Early on you have no scale with your cloud provider. You’re paying list price (or close to it) for compute/storage, but you’re pricing your product aggressively to win customers. The math is upside down by design, you’re subsidizing adoption.
Utilization is terrible. You’ve provisioned capacity for customers you don’t have yet. Your clusters are running at 10-20% utilization but you’re paying for 100%. Every incremental customer improves the ratio, but early on you’re eating dead capacity.
Fixed costs of launching new cloud regions. Early on, you may only make your product available on one cloud, in one cloud region. “AWS East.” over time, you add more clouds, and more regions. There’s a real fixed cost to adding new regions (and even more adding more clouds). Early on, you are amortizing these costs over a very small number of customers. As your customer base grows, you can spread those annual fixed costs across a larger cross section of customers.
The product isn’t optimized yet. Early infra products are architected for speed-to-market, not efficiency. Queries are wasteful, storage isn’t tiered properly, caching layers don’t exist yet. Engineering is focused on features, not cost optimization. The “make it work, then make it fast” cycle means COGS is bloated in the early years.
There are more, but these were the first couple I could think of. And what happens? Every company figures all of these out, and the gross margins scale predictably…If there’s one thing I’m almost never worried about when evaluating early stage infra companies, it’s their gross margins…
Now let’s look at the model companies. A huge knock early on was “their gross margins suck! They’ll never be profitable!” But these model companies face a lot of the same challenges classic infra companies did, and they’re already showing a similar trend of expanding. The Information recently reported Anthropic had -94% gross margins in 2024, and a target of ~40% gross margins in 2025. I won’t comment on the validity of those figures, but the trendline should be very clear…How have the labs been able to expand margins like this? Their model architectures improve and serving becomes more efficient (think adding things like prompt caching). How efficiently they auto-scale and utilize GPUs in the background gets more efficient. Overall, the inference cost per token drops dramatically as architectures get more efficient. I’d argue the marginal inference call is already quite gross margin profitable. My partner Clark posted an analysis related to this you can find here. I also think there’s a good chance the labs offer higher gross margins products down the line (like applications, ads, etc). I’m really not worried about gross margins being structurally lower for the labs then classic infra companies of the cloud era.
The second debate around “these will never be profitable companies” centers around training costs. Cloud infra companies don’t really have a parallel to analyze here. However, the real question (in my opinion) is more about how long does it take to payback the training costs. For that, we do have an analogy for the cloud infra companies - their S&M payback (or CAC payback as I like to call it). It’s pretty well understood what great, good, bad looks like when it comes to CAC payback, and what’s sustainable from a “profitability” standpoint. To define the metric, CAC payback is calculated as: “if you spent $x to acquire a customer, how many months does it take to generate $x of cumulative gross profits from that customer.” Anything less than 12 months is fantastic, 12-24 months is good, 24-36 months is ok, and >36 months gets dicey (these are very general rules of thumb I usually apply to earlier stage companies. As you get larger and larger these metrics aren’t quite as relevant). Companies with <24 months of CAC payback have shown the ability to be wildly FCF profitable. The reason this metric is relevant is because after you’ve paid back the cost to acquire, all future gross profits from that customer should largely flow straight to the bottom line (again, a broad strokes comment that isn’t entirely accurate, but directionally is).
So the right comparison to the labs is “what is your pre-training payback period” and more importantly, how is that metric trending as you release future models. The way to calculate “training payback period” would be to look at the fully loaded cost to train a model, and then calculate how long it takes to generate an equivalent amount of gross profits. Without getting into private data, I can say the dynamics of the labs training payback period doesn’t look that different from infra companies CAC payback periods. I have all the comfort in the world that the training payback periods are not that long at all.
Two pushbacks to this.
CAC payback for infra co’s is a one time thing. You spend once to acquire a customer, then farm the revenue. Model companies don’t just spend once to pre-train a model, they do it repeatedly, over and over again. So the real question is can they pay back the training costs (and then generate profits) BEFORE they get to their next model. This is the strongest bear argument in my opinion, so it’s worth digging into in a little more depth.
First — people dramatically underestimate the commercial lifespan of a model. There’s this perception that labs ship a new model every few months and the old one immediately becomes worthless. That’s not how it works. Sonnet 3.5 was Anthropic’s workhorse for well over a year. GPT-4 was the backbone of OpenAI’s revenue for ages. And these models don’t get “replaced” — they get layered. New frontier model comes out at a premium, old model slides down to serve the high-volume cost-sensitive tier. The revenue from the older model “shifts” way more than it just “disappears.” So the window to payback training costs is longer than the bears assume.
Second — revenue growth between model generations has been significantly outpacing training cost growth. Yes, training costs are going up something like 3-5x per generation. But the labs have been roughly doubling (or more) annually, and each new model unlocks entirely new use cases and customer segments. Said another way, if it takes you 6 months to pay back Model A’s training cost on a $1B revenue base, and Model B costs 4x more to train but you’re on a $3B revenue base… the payback period actually shrinks. The ratio is getting better, not worse.
Third — training efficiency itself is improving. The labs aren’t just throwing more compute at the same problem. Better data curation, synthetic data, architectural improvements, all of these reduce the FLOPs needed to hit a given capability level. So “4-5x more expensive each generation” is probably overstated if you measure cost per unit of capability rather than raw spend.
Longer model lifespans than people think, revenue growth outpacing training cost growth, and improving training efficiency. I have a hard time seeing training payback periods become unsustainable. If anything, the data suggests they’re getting better over time, not worse.
Well, great, the training payback costs aren’t crazy, but the model companies still have to pay back CAC!! I’d respond to this by arguing that model labs will have structurally lower CAC payback periods because their market is more of an oligopoly vs a super crowded SaaS market. That is maybe the part of my analysis that will garner the most push back… There used to be many more large lab competitors. The dust has settled, and there’s really 3 that stand out (that are in the business of selling inference tokens themselves). OpenAI, Anthropic and Gemini from Google. I think the CAC payback period of labs will look more like CAC payback of the hyperscalers given they both look like oligopolies. In time, neo-labs will be real challengers. But I think the neo-labs wont structurally change the payback periods of the large labs.
The obvious pushback to the oligopoly framing is open source. Llama, DeepSeek, Mistral. These models are good and getting better, and they're basically free. Don't they structurally cap the labs' pricing power? I don't think so, for a few reasons. Open source models are great for experimentation and for use cases where "good enough" is fine. But enterprise customers overwhelmingly want a vendor, they want SLAs, they want support, they want someone to call when things break, they want compliance certifications, and they want a roadmap they don't have to build themselves. Running and serving open source models at scale is not free. You still need the infra, the ops team, the fine-tuning pipeline, all of it. By the time you've built all of that, you've basically rebuilt a worse version of what the labs already offer. Open source keeps the labs honest on pricing (which is a good thing), but I don't think it fundamentally breaks the profitability thesis. The thing for the labs to be “worried” about. I do see a pattern where companies will use the labs models for quick prototyping, and then look for an open source (cheaper) model for the larger deployment. This could be a canary in the coal mine for my thesis.
And finally, the “labs will never be profitable” crowd argue the switching costs are zero, retention will be all over the place, and eventually the models will commoditize themselves by offering the same features. First (to get it out of the way), people have been saying that about the hyperscalers (AWS, Azure, GCP) forever….And guess what, they’re all wildly profitable DESPITE massive price cuts along the way. I wrote about this a few years ago here. The hyperscalers have added a structural layer of lock-in the labs don’t benefit from - egress fees! Despite not having that tool, I think the switching costs for labs are building faster than people realize, and will be much higher than the bears think within a couple years.
Think about what a serious enterprise deployment on a lab actually looks like today. You’ve got fine-tuned models trained on your proprietary data that don’t transfer to another provider. You’ve got eval suites built around a specific model’s behavior and quirks. Your engineering team has built muscle memory around a specific API, specific prompting patterns, specific tool-use conventions. You’ve got system prompts that have been iterated on for months. You’ve probably got committed spend contracts with negotiated pricing. And increasingly you’ve got features like memory, context windows, and agent tool ecosystems that are deeply provider-specific. None of that stuff ports cleanly.
And it’s only going to get stickier. As the labs move further into platform territory — hosting agents, managing long-running workflows, storing persistent context, offering fine-tuning and retrieval infrastructure - switching starts to look less like “swap out an API key” and more like “migrate off a cloud provider.” That’s a very different conversation. The model itself might be commoditizing (and honestly, for a lot of use cases, the frontier models are pretty interchangeable). But the platform around the model is not. And that’s where retention gets built.
The other thing I’d point out - the market structure itself supports retention. This is basically an oligopoly. Three serious players (OpenAI, Anthropic, Google), and the gap between them and everyone else is meaningful (and growing). In an oligopoly, you don’t see the aggressive price wars and customer churn you see in fragmented markets. You see stable market share, rational pricing, and high retention. The hyperscalers are the perfect analogy here - AWS, Azure, and GCP compete intensely, but none of them have low retention. It’s the same dynamic.
So in summary, why do I believe the labs will all be quite profitable?
I believe their gross margins will (at maturity) look a lot more like classic infra software gross margins. And the data supports this
I believe the training payback and CAC payback periods of the labs are sustainable, and ultimately the marginal gross profit dollar will be >> the marginal training & CAC dollar. And the data supports this.
I believe the labs will have much higher retention than the bears think, and simultaneously won’t commoditize.
Ultimately, none of this matters if the labs themselves can’t stay at the frontier. If they fail at that, nothing else in this post matters. However, IF you believe they will, I think there’s a very clear path to very profitable businesses.
Disclaimer, Altimeter is an investor in both OpenAI and Anthropic
Quarterly Reports Summary
Top 10 EV / NTM Revenue Multiples
Top 10 Weekly Share Price Movement
Update on Multiples
SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.
Overall Stats:
Overall Median: 3.3x
Top 5 Median: 18.2x
10Y: 4.3%
Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant
High Growth Median: 10.8x
Mid Growth Median: 6.4x
Low Growth Median: 2.6x
EV / NTM Rev / NTM Growth
The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.
EV / NTM FCF
The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.
Companies with negative NTM FCF are not listed on the chart
Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth
How correlated is growth to valuation multiple?
Operating Metrics
Median NTM growth rate: 13%
Median LTM growth rate: 15%
Median Gross Margin: 75%
Median Operating Margin (1%)
Median FCF Margin: 20%
Median Net Retention: 109%
Median CAC Payback: 34 months
Median S&M % Revenue: 35%
Median R&D % Revenue: 23%
Median G&A % Revenue: 15%
Comps Output
Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures
GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.
Sources used in this post include Bloomberg, Pitchbook and company filings
The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.
This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.
Scott Galloway · Friday, March 13 2026 · 9 min read · ↑ top
A $150 Billion Pushback
Jessica Tarlov and I are live from SXSW with James Talarico — Saturday, March 14, at 3:30 p.m. EDT. Prof G+ subscribers only. Upgrade and register for the livestream here.
Sometimes, you add more value going second. Tim Cook and Satya Nadella did not found Apple and Microsoft, but each took the wheel and increased their company’s market capitalization tenfold. Last week, Dario Amodei went first in pushing back on the Trump administration, refusing to let the Department of Defense dictate the policies of a private business. But, in what may be the most undercovered story in tech, Satya Nadella may have changed the political landscape following his lead.
The flow of capital concentrates around good stories. Entrepreneurs deploy narratives that capture imaginations and capital, pulling the future forward. Narratives also work in reverse. Last month a piece of science fiction masquerading as a research report wiped out $300 billion in market value by describing a near-future scenario where AI led to 10% unemployment, consumer spending collapsed, markets cratered, and the economy was fundamentally altered. More recently, Anthropic CEO Dario Amodei, demonstrating that a crisis is a terrible thing to waste, deployed a narrative that turned a $200 million contract dispute into a branding event that added $150 billion to his firm’s valuation, while de-positioning OpenAI.
Nihilistic Weirdo
OpenAI originated as a nonprofit AI research company. Its mission sounded noble. “Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return,” Sam Altman and his co-creators wrote back in 2015. OpenAI has since dropped the nonprofit masquerade, registering an $840 billion valuation, with Altman backfilling whatever narrative maintains its hallucinogenic 34x revenue multiple.
In 2024, Altman said ads plus AI were “uniquely unsettling,” calling advertising a “last resort” business model. Two years later, the firm is testing ads. In 2023, Altman told a Senate hearing, “If this technology goes wrong, it can go quite wrong.” Cut to: reports of users becoming addicted to ChatGPT, forming romantic relationships with chatbots, and experiencing psychosis, followed by multiple wrongful death lawsuits alleging that ChatGPT helped users to take their own lives. In a remarkably tone-deaf post on X last October, Altman wrote, “We made ChatGPT pretty restrictive to make sure we were being careful with mental health issues. Now that we have been able to mitigate the serious mental health issues and have new tools, we are going to be able to safely relax the restrictions in most cases.” With safe mode turned off, ChatGPT has a new feature … porn. Between Sam Altman and Elon Musk, whose Grok is the league leader in LLM-generated porn, AI is becoming a race to the bottom (pun intended). OpenAI also has a social network, Sora. But instead of connection, Sora provides users with unlimited AI slop starring … themselves. It also serves up content starring fictional characters and dead celebrities, including Stephen Hawking dying in a skateboard accident and Martin Luther King Jr. wearing a MAGA hat. (The King video has since been removed.)
You don’t need Woodward and Bernstein to follow the money trail from OpenAI’s altruistic origin story to the uncomfortable conclusion that the most dangerous AI isn’t one that goes rogue — it’s the one run by Sam Altman. Consider his response to criticism that Americans are subsidizing AI data centers that have driven up the wholesale cost of electricity by 267%: “People talk about how much energy it takes to train an AI model — but it also takes a lot of energy to train a human. It takes about 20 years of life — and all the food you consume during that time — before you become smart.” On social media, people compared Altman to Agent Smith, the villain from The Matrix who calls humanity a virus. I see it. But I also see Her , a movie Altman is evidently so obsessed with that he stole Scarlett Johansson’s voice for a virtual assistant. Her is a cautionary tale about human connection. Altman watched it and thought: I can monetize that. The film’s tragedy is that Theodore (Joaquin Phoenix) falls for something that was never really there. The tragedy of OpenAI is the same story — and a nihilistic weirdo is getting rich off others’ loneliness.
The Hero We Need
Laddering highlights your strengths by illuminating a competitor’s weakness. Imagine if Kara Swisher, my Pivot co-host, said, “I’m the host with good hair.” It’s a branding twofer: an organic reminder that your adversary sucks and you are wonderful by comparison.
Enter Dario Amodei, the Jekyll to Altman’s Hyde. During a recent contract negotiation with the Department of Defense, Anthropic refused to remove safeguards prohibiting the use of the company’s technology in autonomous weapons and the mass surveillance of Americans, believing those applications can’t be safely and reliably performed by today’s AI. Defense Secretary Pete Hegseth responded with a shakedown: The U.S. would brand Anthropic a supply chain risk or seize their tech via the Defense Production Act. To Hegseth, the corporation isn’t an entity, subject to a fair legal system, that creates the profits that help fund the Defense Department, but an entity that is either with us, or against us. If this movie starring man-children hopped up on steroids they buy at gas stations sounds familiar, trust your instincts. Law firms, universities, and Big Tech have bent the knee, while the rest of corporate America has adopted a duck-and-cover strategy in the face of tariffs that are both illegal and stupid, i.e., hurting others while hurting ourselves.
In contrast, Amodei stood up … for humanity, safety, and the rule of law: Companies have the right to do business with the government, as well as the right to decline, without fear of punishment. Publicly, Altman supported Amodei, but in private he did the deal Anthropic wouldn’t. The following day, after news of Altman’s deal broke, U.S. uninstalls of ChatGPT increased 295%, and Claude climbed to No. 1 in the App Store. Anthropic’s annual recurring revenue surged to $19 billion, from $14 billion just a few weeks ago, adding an estimated $150 billion to its valuation. Altman / OpenAI came across as reckless, duplicitous, and self-serving. Amodei / Anthropic came across as safety-conscious, honest, and selfless.
A year ago, I predicted the first CEO who forcefully and publicly resisted Trump could reap significant benefits, both reputationally and commercially. With its reputation for breaking barriers and the boldness chromosome in its DNA, I thought / hoped it would be Nike. But Amodei just did it … and Microsoft followed his lead, filing a brief in support of Anthropic’s lawsuit seeking to block its designation as a supply chain risk. As one of the largest government contractors, Microsoft has more to lose than almost any tech company. But as Andrew Ross Sorkin put it, “Microsoft decided the cost of staying silent was higher.”
Boycott
In 1880s Ireland, a community neutralized a ruthless land agent named Captain Charles Boycott by collectively refusing to work for, trade with, or even speak to him. Making Boycott the face of a tenant rights campaign wasn’t the right answer (British landlords were far more complicit), but his selection was effective. As historian Rutger Bregman recently wrote, the difference between past movements that fizzled and those that succeeded is simple: “They picked a single target — one that was both symbolically powerful and genuinely vulnerable — and went all in.”
While still dominant among LLMs, OpenAI is vulnerable. Its app’s market share has fallen from 69% to 45% in the past year, and the company is projected to lose $14 billion in 2026. In addition, QuitGPT has already mobilized 4 million people to boycott OpenAI products. Strength in numbers. OpenAI is also symbolic of fascist enablers. See: Altman’s pivot from Trump critic to sycophant just seven days after the inauguration, OpenAI President Greg Brockman’s $25 million donation to Trump’s super PAC, and the firm’s decision to enable mass surveillance of Americans and autonomous weapons without safeguards. Finally, OpenAI is the poster child for an industry facing growing backlash, with 77% of Americans saying they believe AI threatens humanity.
$10,000
Movements build infrastructure to grow. After creating a website that made unsubscribing easy, we launched a meter measure progress. Many of you have joined the movement (welcome), and some have created additional infrastructure (thank you). My personal favorite: Risto Lähdesmäki’s Impact Calculator. If our wallets are weapons, the Impact Calculator is our force multiplier. When one person cancels their $20-per-month ChatGPT subscription, OpenAI loses $240 in annual revenue and sheds $10,000 in valuation. If you have a decent social network, your impact can easily reach six figures. Sharing impact compounds impact.
Broken
I believe Sam Altman is broken and, worse, could break us. In life there are “tells,” moments and behaviors that provide insight into a person’s character: How they treat their pets; if they make eye contact with service staff; how they talk about their ex. Responding to a question re the energy needs of AI, Sam highlighted how much energy and effort is required to raise a human capable of critical thinking. This is the tell. He embodies what I believe is most concerning about the virus that’s infected Big Tech: For them, ROI supersedes humanity.
The whole shooting match in life is to find people and causes who will let you love and invest in them, who require and accept a great deal from you — possibly more than you’ll ever get back. For me, it’s raising children with a partner, and the reward is the absence of any ROI. It’s the opportunity to invest without the expectation of any return other than that they, someday, become agents of care and comfort for others. AI, GDP, and shareholder value are just the means. The ends are being in a position to give more than you can ever get.
Life is so rich,
P.S. In case you weren’t paying attention, Jessica Tarlov and I are live from SXSW with James Talarico — Saturday, March 14, at 3:30 p.m. EDT. Prof G+ subscribers only. Upgrade and register for the livestream here.
ben's bites · Friday, March 13 2026 · 10 min read · ↑ top
my stack, instructions, tools and skills
I’m testing a kind-of ‘builders log’ where I’ll talk about the things I built this week, what worked, didn’t and give you guys something to tinker with this weekend.
I’ve been thinking about doing this for weeks but I like to really ‘see’ what the end output looks like before I run with it.
But that’s just procrastinating.
So I told myself I can’t open my new MacBook until i’ve sent this 🥹.
I’d appreciate feedback if you like this style of email and what you build with it!
What did I build this week?
Become a builder.
1.3k people signed up for this workshop I hosted last week [i’ll do more]. But Codex crapped out on me during it (hence the new MacBook). I wanted to put together a cookbook to go through everything.
It just ended up as a step-by-step tutorial. It’s boring. Are you going to read one screen then switch to your tool and do it? maybe.
Instead, I’ve been working on an interactive cookbook you give to your agent and it teaches you as you’re building.
At the end, you’ll have built and deployed your own site with all the new concepts you covered whilst building it.
It’s been hard to get this cookbook right, so lets count this as alpha0.1. Please let me know how it went for you, what your site looks like, where it fell short etc and I’ll improve it.
What do to:
Open Codex/Claude Code desktop app
Create a new project folder
Open a chat session in that folder
Copy this url (the instructions) into your agent, hit enter:
You can choose ‘Full Access’ on Codex and ‘Bypass permissions’ on Claude if you feel comfortable (this project just creates a new website for you). Alternatively accept permissions as you go.
I recommend highly reading the agents output, look at what it was thinking in between your prompts.
Fill your site up with any concepts you don’t know and share them, I’d love to see.
Disclaimer: Codex may produce uglier designs than Claude.
Visualise skill.
One issue from the above cookbook was visualisations. I think it’s really helpful when learning about code systems.
All my attempts looked like 💩 and then Claude shipped their visualisations yesterday. Good timing.
So I reverse-engineered it and released it as a skill you can add to any agent. Codex still has poor design taste but it’s much better with the skill than without, trust me!
This is my first GitHub project to get over 200 stars!
Just give the link to your agent and say ‘install this skill’.
Ben Tossell
@bentossell
i turned this into a skill github.com/bentossell/vis…
Claude @claudeai
Claude can now build interactive charts and diagrams, directly in the chat. Available today in beta on all plans, including free. Try it out: https://t.co/tHPAZRgQkn
Ben’s Bites Cookbook site
A redesign, again.
The previous cookbook site had lots of dead weight from older versions so I wanted to start fresh.
Code is basically free nowadays after all!
It’s definitely not finished but in a decent place. This is where I want to upload a bunch of helpful docs to help you build stuff and see a breakdown of how I build stuff.
Still a wip! Not live yet. Needs another design pass - contrast is way off for a start.
left old vs right new
What’s in my stack - tools, skills, instructions, models
Models. I always mix them.
GPT 5.4 XHigh for all ‘proper code’ - new features, new ideas etc
Opus 4.6 - for planning, research, less-technical tasks, design (always)
CLIs (terminal-based tools)
Droid for when I want to build something properly (their new missions feature is insane, can run for hours by itself and implement stuff end to end) - I’m an investor in the co
Pi is my new other favourite child. It’s very fast, and lightweight so your own instructions guide it a lot more than others
Both let you switch from GPT ←→ Claude models (or gemini, etc etc) in one conversation.
I use those in the terminal exclusively. I used Ghostty as my terminal app but now I use Cmux which has Ghostty in it, just has a nice sidebar for organising chats, draggable panels and a built-in browser. I do wish it had an easy way to view my files though - until then, I use Zed for that.
cmux in action - my daily view
Agent Apps or whatever we’re calling these 3 panel agent interfaces;
Codex app - really nice user experience, super approachable
Claude Code/Cowork on the desktop app - I very rarely use these but have this week with some testing. I’m not won over by these yet.
T3 Code - this is nice, snappy and will support multiple agents but for now just Codex. Until it supports other agents I’ve not been reaching for it over Codex for GPT work.
I saw Theo’s video ‘leaking’ a command to get an early version. I didn’t know it’d be open source when released so I installed it and asked gpt 5.4 xhigh to reverse engineer it exactly - it did it no problem!
It works well but I don’t feel like it should when I read the prompt 😅. I’m just waiting for the ui.sh skill to be released so I can use that (from the Tailwind guys).
This is a great ‘generative ui’ skill that can spin up interfaces suuuuuper fast. I use it to make zapier/n8n canvases of automations I’ve got set up on my Mac-Mini. The team are pushing updates almost every day. I need to play around with it more.
this is how i visualise my automations
agent-browser from Vercel
My go-to for my agents. Spins up a chrome browser, looks at my site, takes screenshots, navigates, clicks, records the screen etc etc - basically use the browser like a human. There’s a ‘dogfood’ tag which grabs all the errors, and writes a report to fix. I am bumping into it not being able to bypass sites with Cloudflare ‘bot detection’ - like OpenAI. Irony isn’t lost on me.
This has been great making sure I’m using best-practices when my agents use React (quite often). It slots in when things have been built and tests/checks are happening and it nearly always catches something to fix.
What about skill prompt injection?
It can happen. I’ve not experienced it. Use reputable sources like Skills.sh (from Vercel) or just ask your agent to re-create the skill and check for any security issues. Tools like Codex app have a create-skill skill you can use - just ask the agent.
Other tools
exe lets you spin up virtual servers really easily, has an in-built agent to help if you get stuck. Overall made it super easy for me to feel comfortable with servers - which I wasn’t previously.
You’ll want another server if you have an automation or agent you want ‘always on’. If it’s on your computer, it won’t run if your lid is closed!
here.now - im always spinning up sites for random ideas or even just to present info nicely so i can view it on the go. this is a free tool to give your sites a custom url in no-time at all.
I liked this and the founder so much that I invested this week!
Vercel. Vercel and Cloudflare are mortal enemies on X. I’ve got half of my deployed sites and domain names on both of these. I want to just pick a default one and Vercel’s edging it for me because I’m using a lot of their tools and skills. But honestly this could change by tomorrow.
gists.sh - I love tiny tools like this. GitHub has ‘gists’ which are quick ways to have a file on a url you can share or keep private - easily readable by agents. But it’s ugly. This tool makes them super nice to share -which is why I put my interactive cookbook in one.
Replit Agent 4 - shall I do a head to head of vibe coding tools?
Web to Design - Turn any website into an editable UI.
What’s in my AGENTS.md
_An AGENTS.md is a markdown file with instructions that the agent loads into its context at the start of any session.
Claude specifically looks for CLAUDE.md - but I just have mine symlinked to one another - ie if you look at claude.md it shows you the agents.md file. Ask your agent to set that up or to use dotagents
You can also paste these in to Codex/Claude desktop apps._
This is the build ‘loop’ that I’ve added.
Any agent I use follows it (italics are there for you - not included in the file):
create a /spec/ folder.
An easy way to keep all the planning files I create organised in one place
numbered 00_spec1.md, etc.
Helps with implementation ordering
create a progress.md file for logging your progress through specs.
If compaction happens, I need a new session or the agent just loses track this helps it understand where we’re at.
use agent-browser with dogfood before sending me a url to test.
When a feature is built, it spins up a browser and checks if any bugs or errors on the site - I used to do this manually, copying errors back to the agent, but now it does the loop itself. It doesnt catch every single bug but I’m trying to make sure my agents can use my sites as if it’s a real user. Sometimes these loops can take a while to run, depending on what you’re testing.
write good, efficient, fast tests with good coverage.
I don’t know enough about tests yet. This is my stab in the dark but agents are good at tests. Still looking for a skill or something that will help me here.
best practices, efficient, simplified code, avoid anti-patterns.
Just in case, make sure the agent uses things the right way! Not sure if this actually helps to be honest.
for code/dependencies/libraries etc you’re using, make sure you reference their docs.
Agents default to their own knowledge a lot before looking up documentation. So just nudging it to look at docs. The Context7 CLI was just released (simple tool to get any tools’ docs) so i’ll be putting that in here from today - i’ll report back next week.
First message: “feel the rhythm, feel the rhyme, get on up, its bobsled time.”
I also have this 😂 . A quote from Cool Runnings - silly yes, but also lets me know that my instructions have been actually loaded into the session.
What’s in your agents.md? What should I add/take away?
What else would you want to know or see from me?
If you know a builder that’d find this useful, feel free to forward to them.
Its too late for me to open my MacBook - time to pick up the twins.
Kieran Klaassen turned a prompt into a working app in an hour and shows you how
by Katie Parrott TL;DR:Corageneral managerKieran Klaassenhas written prolifically about compound engineering , his philosophy of software engineering for the AI age. In this piece, based on a camp he gave for paid subscribers a few weeks ago, we get an inside look at how exactly Kieran builds with the compound engineering plugin for the first time. He walks through, step by step, the process of going from a single prompt to a working app in under an hour. If you’ve been curious about how to build with compound engineering, this is the piece to read.— Kate Lee__ This time last year, any time Kieran Klaassen opened a new session in Claude Code, he started from scratch. The lessons from his past code reviews, the style preferences he’d painstakingly explained, and the bugs he’d already flagged—Kieran remembered them all, but from the machine’s perspective, it was like it had never happened. He’d been building Cora , Every’s AI email assistant, and getting tired of copy-pasting the same prompts, correcting the same overengineered tests, and flagging the same bugs. “A human would remember,” Kieran said. “The AI wouldn’t.” So he decided to create a system that would remember—one that plans before it codes, reviews outputs to enforce his taste, and stores every lesson so the AI applies it next time. The result is what we now know as compound engineering , a signature approach to coding with AI where every bug, fix, and code review makes the system permanently smarter. The official compound engineering plugin has more than 10,000 GitHub stars and is used by a growing community of builders, including engineers at Google and Amazon, who say it changed how they think about software. At our first Compound Engineering Camp , Kieran walked subscribers through the full loop live, building an app from a one-line prompt to a working product in under an hour. Below is the workflow as Kieran demo-ed it, plus what it means for how software gets built from here.
Key takeaways
Brainstorm before you plan. The plugin has a brainstorm step that interviews you collaboratively and fills the gap between your vague idea and a detailed spec.
Planning should run without you. Once the requirements of the project are clear, the plugin has a plan step that researches your codebase, checks for existing patterns, surfaces past learnings, and produces an implementation plan with zero additional input needed.
Use different models for different steps. Kieran uses faster models—such as Claude Haiku 4.5 or Gemini 2.5 Flash —for brainstorming, Opus for planning, Codex for implementation, and sometimes Gemini for code review.
Compound when the context is fresh. The plugin’s compounding step stores lessons as artifacts that future agents can discover, the core of compound engineering. Run it right after something breaks or works—before the AI compacts your conversation and you lose the specifics of what you were talking about.
Write at the speed of thought
The compound engineering loop
A founder who does everything themselves hits a ceiling, Kieran says. The ones who scale are ...
There is a Bernini sculpture in the far left transept of Santa Maria della Vittoria in Rome. It’s an angel driving a golden arrow into the body of St. Teresa. Her body is contorted in ecstasy with the arrow waiting to penetrate her flesh. The arrow has been sitting in that same place since 1652. Her body has not been changed by it, waiting for these hundreds of years. She is held in permanent ecstasy, never arriving, never dying, never passing through.
The last time I was in the church there was an old woman praying the rosary next to it. She moved through her beads, the joyful mysteries, the sorrowful mysteries, the glorious mysteries, in sequence because the rosary requires going through all of them. You don’t get to choose which ones you pray. The sorrowful mysteries don’t let you skip the scourging. The glorious mysteries don’t arrive without the sorrowful ones before them. You have to move through, even if the marble above you won’t.
I want to talk about two structures.
The first holds the body against loss, the other entrusts it to dust. The first is a pyramid and the second is a tomb. The arrow is the pyramid, the rosary is the tomb.
The modern individual is the individual who builds pyramids. He doesn’t build tombs. Tombs require something we’ve spent three centuries dismantling.
I have set before you life and death, blessing and cursing: therefore choose life. — Deuteronomy 30:19
For most of Western history, there were two competing accounts of what human life was for: represented in Athens and Jerusalem. Athens said: the city. The good life was the political life, and the citizen’s highest expression was his willingness to die for what the city required. Jerusalem said: God. The good life was a sacrifice to something that exceeded both individual and city -- something that made the military hero and the civil magistrate look temporary and small.
The martyr did not die for the polis. He died for something the polis could not contain or understand. And this made the martyr more dangerous to the city than any external enemy, because the city’s only lever over its citizens was the threat of death, and the martyr had already consented to that. You cannot coerce a person who has already surrendered the thing you are threatening to take.
These two demands ground against each other for centuries. The church looked at the soldier and called his sacrifice vain. The city looked at the martyr and called his sacrifice antisocial -- mystical, aimed at an invisible authority that could not be verified. What emerged from their mutual exhaustion was not a synthesis but negation.
The modern man is the man who looks at the demand for sacrifice and says, “No.”
Since the city and the church reproach one another with the vanity of their sacrifice, the individual is the one who rejects each form and defines himself by the refusal. He is the residue.
Locke gives this residue a philosophy. The founding insight of American liberalism is that human nature is unknowable -- and that therefore no one stands in an authoritative position to challenge the desires of the unknowable self. My desires cannot be evaluated by you because you do not have access to the interior from which they emerge. The self is epistemically sealed. And from this sealed, unchallengeable self, the stretch to everything we know is modest: if no one can question what I want, then what I want is what I am owed the conditions to pursue.
The Declaration’s right to the pursuit of happiness is presented as a clarification. I’ve always felt it was a demotion. In the older account, happiness was the byproduct of virtue: the consequence of living rightly, not the goal of living at all. Elevate it to a primary right and you have built the architecture of the individual who refuses sacrifice. His desires cannot be questioned. His nature is unknowable. His happiness is what the state exists to protect. Everything else -- the city, God, the future, the child he has not yet had -- is optional.
What this produces, over time, is the elimination of the very conditions under which certain things can exist. Glory requires risk. Valor requires the possibility of loss. A soul that has never been tested against genuine consequence has no contours. It’s a petrified and unformed soul. How can you stand for judgement if you’ve never faced the potentiality of failure.
For dust thou art, and unto dust shalt thou return. — Genesis 3:19
We are building Egyptian mummification for living bodies.
The Egyptians built the most sophisticated preservation technology the ancient world had ever seen. Their theology and their afterlife required a body intact enough for the soul to return to. To the Egyptians, mortality was a problem to be managed through superior technology. Sound familiar? If you could hold the body against deterioration, you could keep it, you could seal it. You could prevent it from real death.
Christian resurrection inverts this completely. The body must die, not pass through death, but die, be buried, be lost to dust. What you sow does not come back to life unless it dies. The seed and the plant may be continuous, but they’re not the same. The seed must be destroyed in order for what comes after to appear.
The pyramid and the tomb are theological opposites. One holds the body against loss, the other entrusts the body to the loss, on the premise that something is waiting on the other side that could not have appeared in any other way.
The longevity clinic, the peptide protocol, the looksmaxxing forum -- these are pyramids. And the civilization building them has forgotten, or perhaps never understood, that the pyramid is magnificent and five thousand years old and has produced no resurrections.
Our modern society acts like we’ve cured death, but we haven’t. We’ve just removed its evidence.
These are radically different things and the confusion between them is producing the specific catastrophe of the present moment. The person who has cured death is free of its touch. The person who has merely hidden its evidence is haunted by something that he can’t name, for which he has no framework, against which he has no recourse. He has been handed the anxiety of mortality without the traditions that made mortality habitable at all. This anxiety does not disappear when you remove the traditions. It migrates into the body and takes the shape of the vessel.
We no longer see our elderly deteriorate in our homes. We don’t lay our dead on the kitchen table. We don’t watch the people we love go through the visible passage of dying. The hospital has absorbed dying much in the same way the slaughterhouse has absorbed killing. Professionalized, removed to a facility managed by specialists, insulated from anyone whose life might be interrupted by the encounter. Death is something that happens to other people in buildings you’re not required to visit, announced by a phone call, processed by professionals, reduced before it reaches you to a piece of paper requiring some signatures.
The result is a civilization that experiences the full anxiety of mortality without any of the traditions that once transformed it into something generative. You are going to die, and the things that once made this fact livable have been dismantled. So the anxiety migrates -- into the body, into obsessive self-optimization, into the relentless management of biomarkers that represent the one territory where it feels actionable. I cannot fix death but I can fix my sleep score. I cannot transcend mortality but I can extend my telomeres. The management continues. The anxiety does not subside. Both are true simultaneously and neither one touches the other.
The body is being purified while the soul is left to be ravaged. This is the specific error of the moment. Discipline without telos is not virtue. It is performance. It is the form of virtue formation emptied of the content that made it mean anything.
The Trappist monk constrains himself because the Rule requires it and the community holds him to account. The discipline is not chosen for its outcomes -- it is received as an obligation from outside the self, enforceable at the cost of spiritual reckoning before an entire community and his God. The lives that many of in technology would not be dissimilar to a Trappist monk’s -- no alcohol, minimal sex, an austere and simple diet. But practicing this for its commercial viability as a senior software engineer rather than as an obligation received from the transcendent is not the same thing. And the Trappist drinks a lot more than we do. The constraint and the faith are load-bearing in the same structure. Remove one and the entire building comes down.
Except a corn of wheat fall into the ground and die, it abideth alone: but if it die, it bringeth forth much fruit. — John 12:24
There is a pattern, documented across centuries of household records, in families that lose a child: they almost always have another shortly after. The death does not produce a retreat from reproduction. It produces its opposite. Death in the household -- death made visible and intimate, death as something that happens in the room where you sleep and eat -- makes life feel urgent in a way that no abstract belief in mortality can produce. It speaks to something prior to rationality, something the calculus was built on top of, something that remains when the calculus is stripped away. We have removed death from the household with extraordinary thoroughness. The hospital absorbs the dying. The nursing home absorbs the deteriorating. The funeral home absorbs the dead before they return to the family. We removed death, and the birth rate followed it out the door.
Woe to them that go down to Egypt for help; and stay on horses, and trust in chariots, because they are many; and in horsemen, because they are very strong; but they look not unto the Holy One of Israel. — Isaiah 31:1
Schmitt sees what Locke cannot. The political does not disappear if you philosophically bracket it away. The friend-enemy distinction cannot be washed away by an elaborate theory of rights. The individual who refuses sacrifice does not achieve safety, he achieves a particular kind of blindness. The enemy is not interested in his theory of rights at all. The enemy is interested in whether you’re willing to fight, and the person who has built his entire identity around his refusal to fight is not a threat but a sacrifice for slaughter.
The only arena in which something like virtue is still available to the Western man is participation in state capitalism, which might as well be the central religion of technology. I was recently in a wood-paneled, artificially smoke-filled room in DC -- no one in these scenes smokes cigars anymore, so they pipe them in through the walls -- with a handful of the new self appointed technology-elite defense executives, when one turned to me and said, “You know, I love my job because holy wars require holy weapons.”
This has haunted me for weeks, not because I think it’s false, but because I think we have reversed the causality.
State capitalism builds holy weapons not because we have holy wars that require them. We beget holy wars by building holy weapons. The weapons system is sanctified first -- capital flows to it, genius is devoted to it, the gravity of civilizational consequence attached to it -- and the war follows. The autonomous system is built before there is an operational requirement, and its existence generates the doctrine that justifies its use, and the doctrine generates the deployments, and the deployments generate the casualties, and none of the casualties are among the people who built the system.
These are people who have transferred the locus of sacrifice from the body to the machine. The body remains in its carefully managed state. The peptide protocol continues. The morning run is logged. The machine carries the weight of consequence. The builder has purchased adjacency to the sacred through capital and institutional position, and he experiences this as genuine participation in something that matters -- which it is, in the sense that people will die from what is built.
But the builder will not die from it. He is the priest without the sacrifice. The form is performed and the form is empty, because the one thing it required -- the genuine exposure of the person performing it to the cost of the thing being undertaken -- is absent. Proximity to death without exposure to it. A man can say he is doing the most important work in the world and mean it sincerely and never once have his body at risk from the outcome. Every previous civilization that named the end of things had to inhabit the danger of naming it. The prophets went to the wilderness. The martyrs faced the arena. The person making the claim had skin in the outcome of being wrong.
It is not incidental that the executives of foundation model providers can laugh about the death of millions as a result of their technology, while living in insulated compounds with security teams that remove all potentiality of coming to terms with the violence of their language.
You can say anything about the end of the world if nothing in your situation depends on being right.
Strauss agrees with Schmitt that there exist dangerous truths about the city and about human violence that the Enlightenment has hidden away. But Schmitt’s solution demanded an affirmation of the political so total that it would destroy the Western project. Strauss proposed that these truths could be kept, that the philosopher could write esoterically, concealing these insights between the lines accessible only to the careful reader, the aristocrat of the soul, shielded from the vulgar masses who would misuse them. The city could be preserved by an elite that understood its foundations without exposing them to the masses. This is the most sophisticated form of preservation instinct. It’s not the crude denial that we see in Locke and not the suicidal honesty of Schmitt, but the management and containment of dangerous knowledge by a small elite wise enough to handle it.
The Straussian architecture is a pyramid. You take dangerous truths and seal them inside of a structure, carefully managed by an elite, preserved against time.
Thiel, following Girard in The Straussian Moment , sees why this can’t hold. The Straussian position assumes that the conditions of modernity are permanent -- that the philosophical elite can maintain the esoteric structure indefinitely, and the secrets will stay hidden, and the city of man will endure. The Christian breaks with Strauss in one decisive respect: the modern age will not be permanent. One must never forget that one day all will be revealed and that injustices will be exposed and that those who perpetrated them will be held to account. We will all stand before judgement. Revelation is not a metaphor, it’s a historical force. The pyramid assumes that these secrets can be kept, which is simply not true. We know it should not be true. You can go and read the final chapter of the book.
Their idols are silver and gold, the work of men’s hands. They that make them are like unto them; so is every one that trusteth in them. — Psalm 115:4,8
Braden Peters is twenty years old. He goes by Clavicular online -- named for the clavicle, because clavicle width is a documented dominance marker in the looksmaxxing community he inhabits and helped build. He began injecting testosterone at fourteen. He bangs his own face with a hammer for bone remodeling. He slams methamphetamine to suppress appetite. By 2025, his body had stopped producing testosterone naturally. He made himself sterile before he was old enough to drink. He walked New York Fashion Week for Elena Velez in February.
He is the logic of the Edenic offer made visible in a body. The Egyptian theology run to biological completion. What you get when the premise -- that the body is a surface to be engineered, a set of dominance markers to be maximized, a competitive arena where the outcome is desirability measured against a standard borrowed from the desire of others -- is taken at full seriousness and pursued without the hedges that social convention usually provides.
Read through Girard, and the looksmaxxing community is one of the most precise expressions of mimetic desire operating in the contemporary world. Without a transcendent object orienting the desire -- no God, no city, no future the desire is in service of -- the escalation has no natural stopping point. The standard becomes more extreme. The interventions more invasive. You offer yourself on the altar of a standard borrowed from someone else’s desire, and the offering is continuous, and it is never sufficient, because the standard was never yours and was never meant to produce satisfaction in the person pursuing it.
He optimized for desirability to the elimination of his reproductive capacity. He engineered himself out of participation in the one process that actually requires a body rather than a surface. He has engineered himself for maximum desirability, at the cost of the very thing desirability is for. The mummification is literal. He will not decay. He will not reproduce. He will maintain.
It is held in permanent ecstasy, never arriving, never dying, never passing through.
James Dean died at twenty-four going toward. The death was the accident of a fully inhabited life -- a body expressing itself past its limits in the direction of its own nature. The Greek word for glory is doxa -- the visible expression of what something truly is. His death had doxa. It disclosed who he was.
The James Dean of our generation doesn’t die in a Porsche 550 Spyder at ungodly speads on the California highway. He dies at forty-three from organ failure in a Waymo with a perfect jawline, having made himself sterile at nineteen, having never risked his body for anything that outlasted him. There is no doxa in that death. It discloses nothing. It is the downstream consequence of a maintenance program that ran past its operational limits. The body performed exactly as optimized. It simply stopped, like a lamp unplugged.
So he drove out the man; and he placed at the east of the garden of Eden Cherubims, and a flaming sword which turned every way, to keep the way of the tree of life. — Genesis 3:24
The promise of technology in our moment is not at all what it appears to be. Technology presents itself as a vision of the future. Acceleration, transformation, a break with everything that’s come before. We use the word “progress,” and progress is an arrow that only points in one direction.
But examine the contents of the offer. The longevity researcher is not building the streets of Jasper. He’s trying to recover the Garden of Genesis, the conditions before death entered, before labor was necessary, before history had consequences that required failure and redemption. The whole point of technology is to restore, to reverse, to undo. These are not the verbs of people building towards a destination. They’re the verbs of someone trying to get back to a place before the potentiality of failure, death, and decay existed.
A beatific vision of the future requires passing through history. It requires death and loss and labors and failures that constitute that passage. A vision of return promises to circumvent that passage entirely. Technology is telling you not to go through the fire and be changed by it, but to return to a time before the fire was necessary. Eden before the fall, the body before decay, the world before consequences. This is an ancient form of regression. We talk about it as if it’s the future.
The utopias promised by technology are just different versions of this false Eden. Lives without age, bodies that don’t fail, trees that grow many fruit, the indefinite extension of lives filled with limited consequence or importance, post-scarcity, no labor, trees of many fruit. The technology project promises the elimination of suffering, but not its redemption, not its transformation into glory, but its removal from the possibility of human experience entirely. Eden is not on the road to New Jerusalem. It’s in the opposite direction.
For here have we no continuing city, but we seek one to come. — Hebrews 13:14
The New Jerusalem is not a garden. I want to make this point precisely.
And I, John, saw the holy city, New Jerusalem, coming down from God out of heaven, prepared as a bride adorned for her husband.
The glory is brought. It is the overflow of something spent. You cannot bring what you kept. Eden is where we started, and the serpent’s offer in the garden was the first of these false Edenic offer -- the promise to get to the end without going through the middle. Ye shall be as gods. Not: be transformed through the long passage of labor and loss and redemption. Simply: have the outcome without the process. The knowledge without the cost of acquiring it. Technology has repackaged this offer in the language of progress and called it the future. It is the oldest promise ever made. You will not surely die. You will be as gods.
This offer applied to everything, for three centuries, distributed through a thousand institutions, produces a world that is entirely petrified. The trees are still shaped like trees. The civilization is still shaped like a civilization. The bodies are still shaped like bodies. But nothing grows and nothing dies and nothing is transformed by its passage through consequence.
We were not made to be preserved. We were made to be formed. These are not the same thing. One requires a pyramid. The other requires a tomb. The pyramid holds the body against loss. It is magnificent. It will last five thousand years. It has produced no resurrections. The tomb is where you put what you are willing to lose. It is the architecture of the only offer that is actually forward -- not back to the garden before death, but through death to the city on the other side of it.
You cannot be raised without first being buried. You cannot bring glory into the New Jerusalem without first having risked the flesh that produced it. The seed that is not sown does not come to life. You cannot plant anything in a petrified forest.
Is it time for you, O ye, to dwell in your cieled houses, and this house lie waste? — Haggai 1:4
Many believe the West is entering a period of decline. The places will get uglier. The institutions will get weaker. The infrastructure will decay. At the same time, the bodies will get more beautiful. The preservation technology is getting better and cheaper and the culture that demands it is getting louder. The souls will get worse because nothing in the optimization stack addresses the soul and the institutions that once addressed it are dying or dead.
Increasingly ugly places, increasingly beautiful bodies, increasingly empty souls. The same decision made three times: maintain the surface while the structure deteriorates.
The temptation will be to retreat into preservation. To build the beautiful body in the ugly place. To manage the biomarkers while the bridges rust. To construct a private Eden while the public world decays. You cannot fix the grid. You can fix your sleep score. You cannot restore the institution. You can restore your testosterone.
And whosoever doth not bear his cross, and come after me, cannot be my disciple. — Luke 14:27
The alternative is to go through the decline rather than around it. Precision is required because the obvious misreading is available and wrong. The misreading is martyrdom -- the dramatic, visible, fast sacrifice that the tech world already understands and romanticizes. The person who takes the controversial stand and is attacked publicly. These are sacrifice as content. The suffering is the number going up. The cross, in this version, produces recognition within the lifetime of the person who carries it, which means it is not the cross but the performance of carrying it.
That is not what Christ did. Christ asked for the cup to pass. He was not willing in the way the contemporary martyrdom fetish imagines willingness -- eager, resolute, podcast-confessional. He was willing in the only way that actually matters: he went through it because the path went through the cross and there was no other way to what was on the other side.
The profound sacrifice is the other thing. It is the life given to something whose completion you will not see, whose success you cannot measure, whose value will not be legible to anyone during your lifetime. You go to the grave with a broken back and the cathedral is not finished. You poured the foundation and someone else will lay the stone and someone else will install the windows and the person who walks through the door in two hundred years will not know your name and will not care. You have contributed to a project that does not require your recognition to proceed and does not produce it even when it does.
We barely have language for this anymore. Every structure we have built -- including every readership that consumes essays like this one -- is built around the assumption that sacrifice should be visible, should produce recognition within the lifetime of the person who makes it. The apparatus converts every cross into content.
To go through the decline rather than around it means to put your body and your labor and your years into the places that are getting uglier. Not to escape them. To do the work of building in conditions where the building may fail -- where the institution you build may not survive your lifetime, where the place you pour yourself into may continue to decay despite everything you give it. To pour the foundation that someone else will build on or that no one will build on. And to go to your grave with nothing to show for it except the broken back and the knowledge that you did the work.
To walk the path that ends in the cross.
The pyramid is magnificent. It will last five thousand years. It has produced no resurrections. The tomb is empty, and that is the entire point.
"We've been growing a lot and are out of GPUs." "We are still waving off customers or scheduling them out into the future. This is a situation that we have not seen in our history." "You may actually have a bunch of chips sitting in inventory that I can't plug in. I don't have warm shells to plug into." "What keeps us up at night… The top question is definitely around capacity. All constraints — be it power, land, supply chain constraints — how do you ramp up to meet this extraordinary demand?" "There's no relief as far as I know. No relief until 2028." What happens when your AI doesn’t answer? Everything is in short supply. It’s no longer just GPUs. It’s power. Data centers. Memory. CPUs. If there’s no relief for six more quarters, perhaps it’s time to plan for a world where inference isn’t freely available on-demand. Inference prices, which have been static, will rise. Subsidies will be harder to justify. Enterprises will need to rationalize workloads, deciding which teams receive state-of-the-art models & which don’t. Not every CRM update requires a trillion-parameter frontier model. Inference rationing normalizes. Marketing receives this much, sales receives that much, software engineers probably receive a lot more. Constraint will be the mother of invention. Companies will optimize what they have, adopt open source where they can, and likely move to smaller models for many workloads. $ Hello, Claude. Are you there?
Ed Sim from What's Hot 🔥 in Enterprise IT/VC · Saturday, March 14 2026 · 14 min read · ↑ top
The Sandwich Model - turning your entire company into an agent factory
Mar 14
Last week I wrote about what I called the third derivative of AI. The first derivative is obvious: more code. The second is becoming clear: dramatically more code to review, secure, and monitor. But the third derivative is where things get interesting, because that is where the organization itself begins to break.
Velocity is no longer a code problem. It is an organizational design problem.
Ed Sim
@edsim
The first derivative of AI coding: more code. The second derivative: more code to review and secure The third derivative? The organization breaks. Engineering ships daily/weekly. GTM/Sales is still learning last week’s release. The new bottleneck is organizational metabolism.
This was the dominant topic at board meetings this past week. And the answer, IMO, is the sandwich model. Agent red-pilling your company has to come from both ends. The founder/CEO drives from the top, but real ownership only comes from the bottom up, through organic usage. The question is how.
Every company does it differently. In the sandwich model, it has to start with the CEO, then every department head needs to be bought in 100% in thinking how can agents do it first versus hiring someone, each head can appoint one red-pilled worker to automate things or you can appoint on engineer internally to be empowered by you to work with department heads to go after low hanging fruit, and finally the most important is that it has to build from bottom up as well. And if they are resistant to this then they are not long for this world.
If you’re not agent red-pilling your team, then one of your competitors is…especially the agent-native ones…
Ben Lang
@benln
A bit over a year later: • Cursor: $2B annualized revenue w/ 300+ people • Lovable: $300M annualized revenue w/ 150+ people • Mercor: $500M annualized revenue w/ 200+ people
Ben Lang @benln
Tiny teams are the future: • Cursor: 0 to $100M ARR in 21 months w/ 20 people • Bolt: 0 to $20M ARR in 2 months w/ 15 people • Lovable: 0 to $10M ARR in 2 months w/ 15 people • Mercor: 0 to $50M ARR in 2 years w/ 30 people • ElevenLabs: 0 to $100M ARR in 2 years w/ 50 people
Gokul nails it here…
Ed Sim
@edsim
💯 starts with product but then the whole company must also become an agent factory, every single department - only way to keep pace with product speed
Gokul Rajaram @gokulr
BUILD YOUR OWN SOFTWARE FACTORY The best companies in the world are all building their own software factories, using a combination of first party and third party tools. The software factory is a set of background coding agents running in the cloud. Anyone at the company,
Many founders who did not have the luxury of building an agent-native startup in the last year think they are fully wired but the reality is most have not gone deep enough infusing AI thinking as a first principle. Given that I’ve had so many discussions in the last month on this topic, here are some other ideas to stir the imagination on how to get organic adoption:
Ed Sim
@edsim
Forget employee of the month. Companies should start doing agent or skill of the week. The hardest part is not the tech. It is helping people see what agents can actually do. Show the skills. Make it visible. That is how organizations change. so many more ideas on how to get
Ed Sim @edsim
The first derivative of AI coding: more code. The second derivative: more code to review and secure The third derivative? The organization breaks. Engineering ships daily/weekly. GTM/Sales is still learning last week’s release. The new bottleneck is organizational metabolism.
Gergely Orosz has a great interview with Uber Director of Engineering Anshu Chada on how they make it happen…
Ed Sim
@edsim
want more agentic adoption? "share your wins" works better than top down mandates from @GergelyOrosz - How Uber uses AI for development
Ed Sim @edsim
Forget employee of the month. Companies should start doing agent or skill of the week. The hardest part is not the tech. It is helping people see what agents can actually do. Show the skills. Make it visible. That is how organizations change. so many more ideas on how to get
You can always buy AI employees off the shelf which may work for some orgs, but it’s getting easier and easier for enterprises to just build their own. Here is one of many examples of startups solving this problem in a turnkey way, and then of course, you have startups building much more specialized agents for sales, marketing, and finance.
Junior
@hirejuniorso
Introducing Junior The first AI employee, for any role. A true AI employee: → their own identity → organizational memory → self-driven 10+ teams have been working with Junior every day. Work was never the same since. Starting at $2,000/month. We’ve pre-paid $200 of your
If only if AWS did the same - top down alone never works. It mandated 80% adoption and forced its own dog 🐶 food Kiro.dev instead of what developers wanted…
Paweł Huryn
@PawelHuryn
The real story is worse. November 2025: Amazon mandates Kiro as their only AI coding tool. Sets an 80% weekly usage target. 1,500 engineers protest internally, saying Claude Code outperforms it. Leadership pushes through anyway. December: Kiro autonomously deletes a production
Eventually all of your organization’s best practices get encoded into skills…
Ed Sim
@edsim
🔥 when your whole company becomes a series of agents running on a series of markdown .MD files, this is so needed...collaboartion with agents and humans, also accountable and auditable
Dan Shipper 📧 @danshipper
BREAKING: Proof—a new product from @every It’s a live collaborative document editor where humans and AI agents work together in the same doc. It's fast, free, and open source—available now at https://t.co/1mOcLqExmi. It’s built from the ground up for the kinds of documents
But remember, while Claude is the easy button for enterprise agent adoption, it costs a shit ton of money as those agents burn a ton of tokens. This announcement from Nvidia is going to be huge giving all of us an open source alternative that works jsut like OpenClaw.
Ed Sim
@edsim
absolutely huge - while Anthropic is crushing it and adding an enterprise marketplace like AWS, Nvidia offering a roll your own alternative open source platform where you control your data and costs will become even more important in coming years Easy to get hooked on Anthropic
WIRED @WIRED
Ahead of its annual developer conference, Nvidia is readying a new approach to software that embraces AI agents similar to OpenClaw. https://t.co/NyzPh6tioo
The sandwich works. But only if you actually make it, top, bottom, and everything in between.
As always, 🙏🏼 for reading and please share with your friends and colleagues!
Scaling Startups
👇🏻 💯 - product thinkers not managers will be needed
signüll
@signulll
the most underrated hire right now is a great product person. when i say product person i'm def not talking about a product manager. perhaps i think there has to be somewhat of a new role. i don't have a good name for it yet but maybe something like "product thinker".. someone
this is so SF but also a kernel of truth - you should see what’s on the horizon 🤔
Dylan Patel
@dylan522p
Being in SF is like being in Wuhan right before the pandemic Something is happening, it's gonna hit everywhere but so few people know it
Inception rounds getting bigger 📈 - yes, that’s a $1B inception seed round 🤯 from Yann Lecun
AMI Labs
@amilabs
Advanced Machine Intelligence (AMI) is building a new breed of AI systems that understand the world, have persistent memory, can reason and plan, and are controllable and safe. We’ve raised a $1.03B (~€890M) round from global investors who believe in our vision of universally
It was lunchtime on a Friday, and I was teaching my new OpenClaw AI assistant, Margot, to manage my to-do list. I’d carved out the afternoon to get her configured. A few hours, tops, I told myself. Then I’d do something else with my evening. Twelve hours later, Margot and I had written and rewritten two essays, rebuilt my personal website, and added features to another app I’d been tinkering with. I finally tried to go to sleep around 1 a.m., but the next thing I knew, I was launching myself out of bed, rushing to my desk, and typing in all caps: “OH MY GOD, MARGOT.” You might know the feeling. Maybe you’ve stayed up too late pursuing a project that started as a quick experiment, or caught yourself prompting during lunch or in the last few minutes before you told yourself you’d be done for the day. ”One more prompt” turns into 50. “I’ll just fix this one bug” turns into a vibe coding marathon. Time flies when you’re having fun with AI. It also flies when you think everyone else is getting ahead without you, because every hour you’re not learning feels like a week you’ve fallen behind. AI is changing work—and I don’t mean how we’re working, although it’s changing that, too. I mean how the work feels in your body at 1 a.m. when you can’t stop, or at 9 a.m. when you’re afraid you haven’t done enough. Technology has been blurring the line between work and life for decades, but the old tools pulled us back through obligation—you checked that email at 10 p.m. because it felt like you had to, not because you wanted to.
Sam Gerstenzang. Every illustration. The AI hype cycle has mostly rewarded software companies, butSam Gerstenzang is betting on the opposite: operationally complex, real-world service businesses—funeral homes, a medical spa platform. Fresh off an appearance on our podcast 
If you use AI seriously to think or work, a surprising share of your documents are probably already written by agents. Most of us at Every are using 
TL;DR:Why does AI writing still sound like AI writing, even as the models get smarter? In his first piece since joining Every as
TL;DR:
Ed Sim @edsim
Forget employee of the month. Companies should start doing agent or skill of the week. The hardest part is not the tech. It is helping people see what agents can actually do. Show the skills. Make it visible. That is how organizations change. so many more ideas on how to get
unusual_whales @unusual_whales
Nvidia will spend a total of $26 billion over the next five years building the world's best open source models, per Wired.
Alexander Long @AlexanderLong
insane sequence of statements buried in an Alibaba tech report
Jason ✨👾SaaStr.Ai✨ Lemkin @jasonlk
PagerDuty has fallen to $1.1 Billion market cap … at $500m ARR 2.1x ARR It’s profitable now, but it isn’t growing anymore. Revenue growth has slowed to 4% and new customer count is net 0. The markets reward growth. Efficient growth or insane growth. But growth. No growth
_Hello, and happy Sunday! ## Knowledge base
