David Boyle

This is what it looks like when a senior leader works with AI on real tasks. Hours of work in 25 minutes. This walkthrough covers four tasks. Meeting prep, a board recommendation, a sponsorship deck review, a pricing model rerun. AI read 15 emails, eight documents, two transcripts and a revenue model. I steered, checked and decided.

@gregisenberg Bought the other day

in a world of INFINITE content, infinite choice, infinite scroll, people are starting to want things that END - finite formats - physical products - no ai, no internet - clear boundaries there’s a real shift here and it’s going to create massive companies here we go

@gaughen Had a similar issue! I bought a chisel and it wasn’t able to make the statue of David from a block of marble.

I didn't realize how hilariously bad artificial intelligence was until tonight, when I asked it about something I'm an expert on. I asked it about zoning laws in Pennsylvania and the AI hallucinated case law that doesn't exist. Silicon Valley wants us to rely on this slop? 😬

@randal_olson @ManusAI Did you ask for permission or offer a royalty?

This week, I encoded Edward Tufte's data visualization principles into an API. Then I let an AI agent try to pass it. I gave @ManusAI a CSV of women's bachelor's degree percentages across STEM fields (1970-2011) and one prompt: visualize this data. It produced a standard chart. Correct data, readable axes, nothing wrong. But a legend box instead of direct labels. No annotations calling out the rise and fall of women in Computer Science. Default colors. This is what every AI agent produces right now. So I pointed it at the Tufte Test, a quality standard I built in Truesight that checks charts against seven of Tufte's core principles. The API came back: fail on direct labeling and integrated annotations. Five other criteria passed. A quality standard gives an agent something a vague prompt never can: a precise list of exactly what to fix. Manus revised on its own. Legend box became direct endpoint labels. A subtitle surfaced the key insight. An annotation marked the Computer Science peak at 37.1% in 1983. Two prompts total from me. Everything else was autonomous. Any AI agent that can call an API could do this. What matters is the pattern: encode expert judgment once, deploy it as an API, and every AI agent in your stack builds against it. Your taste becomes infrastructure at scale instead of manual review. The Tufte Test is available as a template in Truesight if you want to try it on your own charts. Full writeup + demo video: goodeyelabs.com/insights/the-t…

Read / subscribe here: steadman.ai/newsletters/da…

Four weeks of Saturday AI Thoughts. This week: the growing gap between AI's power and our ability to use it safely. A client told me that every person they show Claude Code to says the same two things, in the same order. First: I'm so excited. Then: I'm completely terrified. I feel both. Every day. The top builders across Sequoia's portfolio are three to five times more productive than a year ago. The median? Up ten to twenty percent. The gap is widening, not closing. Amazon learned what happens when power outpaces care: an AI coding agent autonomously deleted and rebuilt a live production environment. Thirteen hours of downtime. Then two more major outages in March. The organisations I'd bet on aren't just the ones moving fastest. They're the ones building in care as they go. Also in this edition: why more AI tools make you less productive (the answer peaks at three), what ATMs and iPhones tell us about job displacement, and three practical things to try on Monday.

In Feb 2020, it was obvious the world was about to shut down. It is just as obvious now to anyone with eyes to see. The same lesswrongers who first warned the world about Covid are the ones screaming from the rooftops about imminent AI takeover

Great writeup. Curious about the execution model — are your overnight automations running headless (launchd/cron spawning claude -p) or interactive sessions? And the AM Sweep subagents — are those within a single Claude Code session or separate processes? The operational differences are significant for reliability and TCC permissions on macOS.

Would have been a better position to hold BEFORE signing the contract

I finally got around do doing a weekly email of the things I’m seeing that feel important. It’s online here if you’re curious: steadman.ai/newsletters/da…

@_simonsmith Cannot believe they still can’t live link back to a shared drive, like Claude has had forever

ChatGPT projects got an update for sources. Not sure when this happened. Cleaner layout, and also you can add text sources. Was that always possible? Still missing the ability to connect live documents from Google Drive. Hope that comes soon.

Here's ten hours of work done in 25 mins. Meeting prep, a board recommendation, a sponsorship deck review, a pricing model rerun. AI read 15 emails, eight documents, two transcripts and a revenue model. You should be working this way!

Full video (with four - very different - tasks) here: youtu.be/_Ap3LZ58Au8?si…

This is what it looks like when a senior leader works with AI on real tasks. Hours of work in 25 minutes. This walkthrough covers four tasks. Meeting prep, a board recommendation, a sponsorship deck review, a pricing model rerun. AI read 15 emails, eight documents, two transcripts and a revenue model. I steered, checked and decided.

@EricLevitz @JasonSilva

I wrote about why Silicon Valley is convinced we're in a "February 2020" moment (and how their dire prophesies could prove mistaken) vox.com/politics/47879…

@jianxliao @openclaw This works for me. Super simple. github.com/davidbeglenboy…

why hasn't anyone built a channels + heartbeat + tmux cli wrapper with claude code to replace @openclaw? honestly openclaw is great but it's just too bloated and unstable

@jianxliao @openclaw Here’s my attempt - a massively simplified method of remote access by telegram that reduces security issues. github.com/davidbeglenboy…

Introducing TinyClaw 🦞 OpenClaw in 400 LoC @openclaw is great, but it breaks all the time. So I recreated @openclaw with just a shell script in ~400 lines of code using Claude Code and tmux. Everything works! WhatsApp channels, heartbeat system, cron jobs, and it uses your existing Claude Code plugins and setup. It’s super stable and extremely easy to deploy compared to openclaw, just install Claude Code! github.com/jlia0/tinyclaw

@EleanorKonik I adapted some of it for my non-coding workflows. Here’s my rough setup github.com/davidbeglenboy…

Really need to adapt this kind of thing to non-technical work. Who is working on that?

And this

Time for this again

In the last month, something changed. Executives freed from corporate procurement constraints have started working in ways their employers wouldn't recognise. My own work has changed more in four weeks than in the previous eighteen months. Meanwhile, the average knowledge worker's workflow hasn't changed since their organisation first gave them ChatGPT access. Same tools. Same prompts. Same output. COVID proved we can't see doublings until they overwhelm us. AI capabilities are doubling every four to seven months. Most organisations haven't noticed. They will very, very soon. An open letter to CEOs about what that means - and what to do about it.