Start with discovery

What we learn putting AI to work.

What we're reading on AI.

See all 525 posts

Tenobrus@tenobrushaiku 5.5 is an absolutely insane model release . i'd been starting to make frequent and good use of luna as a massive fanout machine, just having a reasonably smart agent look at and modify like every single one of ten thousand files for under $100 total. i really think peopleOct 10, 2026Open on X

Nathan Labenz@labenzThe other day I stopped to talk to an elderly woman who lives down the street. She had questions about something on her phone saying “you’ve been hacked” I handed her ChatGPT voice and it explained what was happening and what (not) to do, and she accepted the advice SheOct 10, 2026Open on X

Ryan@hellorymiOver the next few weeks, I'm going to document building a completely free, desktop version of Quickbooks from scratch, and proving that it is well maintained and worth using over the vibe coded slop all over Github.Oct 9, 2026Open on X

signüll@signulllone thing you start to realize as agents get insanely good is just how much of legacy software is basically workflow theater. there are so many legacy workflows that only exist because software historically couldn’t understand intent, gather context, make decisions, or takeOct 7, 2026Open on X

Mark Ajzenstadt@mardehaymHot take: forward-deployed engineering is overrated. Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineers, citing rising costs and the inability to evolve those systems independently. I’ve spent over a decadeOct 7, 2026Open on X

Erik Kokalj@erik_kokaljopenai just released Decisions API, Jev-like "system 1" multi-modal model based on GPT6 Luna! Each egg gets cropped and sent to OpenAI (~270ms latency) and returns P(clean, dirty, cracked). One image crop consumes ~280 input tokens (~$0.03/1K imgs). @roboflow RF-DETR forOct 7, 2026Open on X

Watch how we build.

We Built Our Own Business OS | Runpoint OS

Runpoint Partners · Sep 24, 2026

Runpoint Partners on YouTube Sam Gaddis on YouTube

Essays and field notes.

OpenAI DevDay, one week later: what worked, what didn’t, and what mattersField Note · Oct 6, 2026A week after OpenAI DevDay: which announcements are already useful, where the experience falls short, and what deserves your attention.

How hard is it to hire a senior engineer who wants to use AI?Field Note · Oct 1, 2026I want to hire experienced engineers who like working with AI. Here is how many there are, why they aren't moving, and what they cost.

Are markets betting on the end of the Great Stagnation?Field Note · Sep 29, 2026What stock prices, bond yields, prediction markets and professional forecasters say about AI and growth. Most expect a modest step up. The stock market expects more.

A new kind of model, and we're back on ClaudeField Note · Sep 25, 2026Jev reads text and makes a decision. Opus 5.5 and GPT-6 Sol sent our team back to Claude.

Keep your options open when buying AIField Note · Sep 17, 2026How to give capable operators access across AI providers, choose monthly plans, and keep company work portable.

The OpenAI usage debacleField Note · Sep 15, 2026Competing fixes for AI limits, our Decisions walkthrough, and what personal projects teach us.

What makes a model 'good'?Field Note · Aug 6, 2026Public benchmarks can narrow the field. A small test built from your own work tells you which model is good for your business.

Stop putting every AI project in the same budgetField Note · Jul 23, 2026AI is clearly making people faster and much less clearly making companies better. Three portfolios for deciding which AI bets to implement, underwrite, or fund like R&D.

Karp is right. Just not about you.Field Note · Jul 8, 2026What last week's Palantir moment actually means if you run a $20M to $200M company.

Field Note: Code with ClaudeField Note · May 8, 2026Two days at Anthropic's developer conference. The brain left the box, memory got a substrate, and the agent decomposition framework finally has a name.

Before the AI Project StartsField Note · Apr 25, 2026AI projects should be underwritten before they are built. The worksheet matters more than the demo.

AI Cuts Hours Before It Rebalances ThemField Note · Apr 9, 2026Alex Tabarrok's labor-hours framing is useful. The harder question is how those gains get distributed when firms are already using AI to shrink teams.

The Power of "Cloud" AgentsField Note · Apr 4, 2026The most useful agent in my life is not a SaaS product. It is Hermes on an otherwise idle Mac mini, reachable from Slack, with enough access to do real work.

What Will Your Kids Do?Field Note · Mar 16, 2026An interactive look at AI exposure across 342 US occupations, including the jobs that are growing, well-paid, and safer from automation.

The Orchestration ThesisField Note · Mar 3, 2026Perplexity just launched a $200/month product that coordinates 19 AI models. If the orchestration layer captures the value, what are the model makers actually selling?

Is Claude Code Safe? The Practical Answer.Field Note · Feb 16, 2026What Claude Code sends, whether Anthropic trains on your code, and the settings that matter for confidential business work.

The Best Tools Right Now Are CLIsField Note · Jan 28, 2026A practical guide for business users getting started with Claude Code, even without a programming background.

Ralph Wiggum vs. Claude Code TasksField Note · Jan 23, 2026Claude Code tasks changed the context-window problem. Smaller task packets make agent work easier to review, retry, and trust.

The AI Automation SpectrumField Note · Oct 15, 2025A plain-English map of deterministic automation, probabilistic agents, and how to match the approach to the problem.

Notes from the Front Lines of Sora 2Field Note · Oct 13, 2025A week of experimentation with Sora 2: cameo bias, meme slot machines, and creating an unhinged ad for Runpoint.

The AI Adoption Paradox: Why 95% of Companies Fail (And How to Be in the 5% That Succeed)Field Note · Sep 18, 2025Why AI pilots fail, why the rare winners matter, and how operators can move from experiments to production systems.

Reading Anthropic's Economic Index as an Adoption MapField Note · Sep 17, 2025Anthropic's economic index reveals AI adoption as a capability revolution, not a discount revolution, where clean data and clear workflows matter more than model choice.

If It Hurts to Delete, I Waited Too LongField Note · Sep 11, 2025Why clever scaffolding becomes tomorrow's technical debt, and how new model releases force us to confront our attachment to obsolete code.

Replit's New Agent 3Field Note · Sep 11, 2025A browser-driven test harness in a controlled environment, not a bigger model, is what lets the agent iterate autonomously for hours.

Outcomes Drive TrustField Note · Aug 28, 2025How deleting features saved a deal: when a VP called out our overengineered workflow, we learned the roadmap matters less than reality on the ground.

Learning Fast EnoughField Note · Aug 14, 2025From designing for "the average case" to sitting with nurses and dispatchers: how FDEs transform one-offs into product improvements that work everywhere.