We’ve automated every single thing we can @every with AI agents. And yet there’s way more human work to do than ever. We’ve gone from 4 -> 30 human employees since GPT-3. I wrote a report on the structural reasons: how AI makes expert competence cheap, why that drives up demand for experts, and why the dynamic only intensifies as we approach AGI. After Automation: every.to/p/after-automation

Killer rec from @danshipper 🙏

“@OpenAI just killed your startup” is a recurring reaction to new AI releases. @sama argues that it misses how much there is left to build. His reasoning: OpenAI can imagine only a small fraction of what developers will create. Giving builders access to its models and tools lets them explore possibilities the company wouldn’t think of itself. Watch the full episode: every.to/podcast/how-sam-alt…

.@danshipper and @craigmod on the necessity of setting a boundary with AI. AI can be so compelling, it can be hard to look away, Dan admits, describing how he can’t stop checking on his agents once he’s set them on a task—which can sometimes run for 20 hours at a stretch. Craig relates: the idea of locking himself away with 10 people and “mainlining AI” ten hours a day for six months is exciting—but he deliberately holds back. If he doesn’t, he says, he’ll lose touch with the human part of himself capable of writing the “weird books” only he could write. Read the full episode transcript: every.to/podcast/transcript-…

For those who want to move from one-off chats to delegating whole projects to teams of agents but are unsure where to start: every.to/p/codex-graded-my-a… Thanks @every

The legal metaphor for AI engineering (where you write laws, then fine tune them as they meet the real world with case law and precedent) came up several times this week, and I’m liking it more and more. every.to/thesis-statements/d…

Evals are showing up in job listings. @Lennysan shared 25 product management openings, and nearly half asked for experience writing tests of how well an AI system performs a specific task. At Every, @Nityeshaga is helping people build benchmarks based on their own work and standards. When a new model launches, they run those tests and decide whether to switch. When several models can do the job, the decision comes down to speed, cost, and how closely the output matches what you'd have written. Start with tasks you already do and a few examples you’d approve. A leaderboard won’t tell you whether a model is worth switching to. Read the full piece: every.to/context-window/why-…

What changes for an AI app when users can bring their own ChatGPT subscription? @danshipper explains the developer’s tradeoff: Limit usage or pick a cheaper model to keep costs manageable, potentially making the app worse. A portable subscription could let users access the models in their plan instead. @sama says he wants to take his subscription between apps, without signing up for a new one each time. Watch the full episode: every.to/podcast/how-sam-alt…

Dan clearly in the pocket of OpenAI because he notably doesn’t even mention the meh pepperoni pizza.

Vibe checking the food at @OpenAI Dev Day:

here is my official @OpenAI DevDay FOOD vibe check (I know you’ve been waiting)

oldie but goodie

Every business will have its own agent, the way every business has a website—and customers will expect to reach it from inside @OpenAI’s ChatGPT or @claudeai, not by visiting the company's site. Companies are bringing their agents into conversational platforms like @SlackHQ, @Microsoft Teams, and @WhatsApp. @tryramp, for example, is issuing people corporate cards via Slack DMs. At Every, we use our company agent daily. Our Every Agent is in beta: every.to/agent?utm_source=x&… Read the full piece: every.to/context-window/why-…

.@danshipper and @nataliazarina on why modern work is like gardening. Before gen AI, knowledge workers were like sculptors—doing everything with their hands. Now, Dan says, they’re more like gardeners: they don’t make the plants, but create the conditions for them to grow. Natalia frames this as an evolution: knowledge workers used to be individual contributors using systems to do tasks themselves. Now, they’re managers building systems to support a bigger team—one that includes AI. Having good people management skills—creating conditions for success—can also make someone a good “model manager” of AI. Link to the full episode on how Natalia, Every’s head of consulting, uses Codex every day: every.to/podcast/transcript-…

.@danshipper on how AI helps him manage his company better. Fable 5.1 is proving stronger at what Dan calls “discernment”—discerning whether information is relevant enough to the situation at hand, or a person’s interests, not just what’s technically related. He tested it by having the model build him a feed of every meeting at Every, complete with short summaries. It also flagged meetings with 50/50 decisions where it would be helpful for him to weigh in. Dan says it’s the first model he’s used that has pulled out things worth his attention, giving him a new lens into his own company. Feeds used to be the domain of social media companies with millions of data points. Now, Dan says, you can build your own in natural language. Read the full vibe check: every.to/vibe-check/fable-5-…

dev day vlog is LIVE!!!

A guide to getting started with open models: Step 1: Choose your model. Larger models like @Zai_org’s GLM-5.3, @Kimi_Moonshot’s Kimi K3, and @deepseek_ai’s V4.1 Flash handle harder work, but a tiny model like @googlegemma’s Gemma 4 E4B may be enough to sort support tickets or extract invoice numbers. Test your pick on real examples of your work—you may need to tune the prompts. Step 2: Choose your hosting. Use a service like @OpenRouter to try models without managing the hardware. Or run a model locally with @lmstudio for more control over your data, if your machine has enough memory. Step 3: Connect your tools. Add your provider’s API key and endpoint to a compatible coding or writing tool, or use LM Studio’s chat interface. Start with one recurring task whose results you can check. Explore @kai_zau’s full guide: every.to/guides/getting-star…

i agree with this sentiment but it’s a category error. there’s no such thing as what machines produce, only what humans produce with machines

we published a sick guide to getting started with open models on @every: every.to/guides/getting-star…

New AI models like Sol and Sonnet 5.5 are faster and more token efficient. We’re not always better off for it. @jackcheng used to exhaust his weekly allocation in about a day. As the models became more efficient, that same allowance stretched to three or four days. He could act on more ideas immediately. He’d think of a feature and send AI to build it. But he was spending less time considering whether it belonged. He even started questioning the project itself: Why keep building a personal AI tool when entire teams at frontier labs were developing alternatives? The model was getting better at executing his ideas. He still needed time to decide which ones were worth pursuing. Jack wants to try letting new features sit overnight. Before bringing a difficult question to AI, he’ll write down why it matters and what he thinks the answer might be. Read Jack’s full piece: every.to/p/what-i-learn-when…

We finally found someone that doesn't contain multitudes