# WT351 — Attention Is The Scarce Part - Issue: 351 - Published: September 26, 2026 - URL: https://weekly.thingelstad.com/archive/351/ - Summary: Ethan Mollick's overhang, Attention, M5 Ultra local models, AGENTS.md, Boris Cherny's thinking loop, MnTech on AI policy. - Word count: 2140 - Domains: a16z.com, alicegg.tech, arenamag.com, asyncdot.com, bcantrill.dtrace.org, borischerny.com, claude.com, en.wikipedia.org, github.com, minnestar.org, thorstenball.com, willd10.substack.com, www.densediscovery.com, www.macstories.net, www.oneusefulthing.org, www.pagetable.com, www.startribune.com, www.theregister.com --- Hello there! First a quick welcome and hello to a number of new readers that have joined from [Dense Discovery](https://www.densediscovery.com). I'm a regular reader of DD and a couple of times a year I love to support it and share the word by placing a classified in an issue. If you're new here, thanks for giving the Weekly Thing a try! It was also nice to get notes from many of you on my return to your inbox after nearly four months away. Thanks for that! Now to some links and interesting reads from the week… ## Currently **Building:** I've been keeping Claude busy building out features for **Elixir Clan**, an add-on to the Elixir platform that takes all the Clash Royale clan management features I built into our Discord bot and turns it into a proper product that anyone can use. **Listening:** I remember the exact day when Uncle Tupelo's [March 16-20, 1992](https://en.wikipedia.org/wiki/March_16–20,_1992) album was released. I called my friend Ed immediately after listening and we both wondered if Jay and Jeff had lost their minds. Now thirty-four years later this is one of my favorite listens. In hindsight I think it set the tone for what became of much of Americana music. ![Sailboats moored on Lake Harriet at sunrise, framed by overhanging branches along the shoreline.](https://files.thingelstad.com/weekly-thing/351/images/1a7e4f4b35b7.jpg) Beautiful morning walk around Lake Harriet as the sun has just come up. _Sep 26, 2026 · [Minneapolis, MN](https://www.openstreetmap.org/?mlat=44.92291&mlon=-93.31107#map=16/44.92291/-93.31107)_ ## Notable ### [The Incumbents Are Coming | Andreessen Horowitz](https://a16z.com/the-incumbents-are-coming/) I find it really interesting to see how established platforms and systems of record are evolving around AI. The most personal example for me is AWS. I've used AWS for years but frankly I find it nearly impossible to work myself. I stumbled around with it and get something working and then never touch it. Today, my AWS account is completely managed with code and better than it ever has been. How? Claude Code helped me revamp it. AWS has realized that its customer is the Agent as much as anyone else, and has started building to it. In many ways we are seeing a new layer of capability being built either on top of platforms or systems-of-record. The right move for those companies may be to just capture the core value and allow those to build on top? At the same time, those platform companies can be a little more measured and build where their signals tell them there is real value to provide the market. ### [The Overhang - Ethan Mollick](https://www.oneusefulthing.org/p/the-overhang) Mollick continues to approach AI from a pragmatic, center-of-the-road position. In general I am an "applied" technologist. A builder. So I tend to push the most on what we can do now and in the near future. I think Mollick's reminder that even if all progress stopped in AI there is still a huge transformation in front of us. > My point is that, while there is a lot of debate over what future models will do, the current capabilities of existing models are barely being used, and are often not even well understood. I think his highlight of the uniquely human things that make our partnership with agents the most powerful is a good callout on where to focus: 1. Deep knowledge: the expertise that comes from understanding a field or subject so well that you build intuition around it to quickly and accurately make decisions. 2. Wide knowledge: To gain wide knowledge you need to read and study widely, across fields and formats and traditions. 3. Taste: Some people have a taste for things that many people will find popular, others have a taste that is unique to them, and still others have a taste for what is novel and new. 4. Agency: willingness to test the boundaries of what’s possible when everybody is equally confused about what AI can do. It is notable to me that everything described there is what would make a great **Product Engineer**. ### [M5 Ultra Mac Studio Review: The Dream Mac for Local AI Agents - MacStories](https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/) The local LLM capabilities of these new M5 Ultras are incredible. I'd also nudge you to this [Apple Wasn't Messing Around](https://www.youtube.com/watch?v=c_58D7ixOQI) video for a great walk through as well. Two thoughts come to mind with this. First, Apple has a unique position here with their M-series chips. They are able to deliver more processing power per watt than anybody else. Sure that matters to keep data center cooling and power lower, but the even bigger thing is the ability to put LLM processing power on a desktop that doesn't require special environmental needs. You can see a direct path from an M5 Ultra to an AI Appliance that people could buy and put in their house. The second is a direct path to local models and product engineering. A long time ago software developers used to get the fastest and biggest computers around because compilers were taxing, and building software took a long time. These days most developers need nothing but a simple laptop as the power of those machines has gone up and the needs have gone down with the cloud and interpreted software stacks. Now as we enter AI-assisted coding we may be flipping back to a world where a developer can justify a $50,000 workstation to run models and build software cheaper. It is also interesting to consider a completely verticalized stack where the development environment is completely merged into a custom model. Instead of plugging a frontier model on the side, you are natively connected to a specific model. Both of these seem like nearly certain developments in the next three years. ### [I am often wrong | Boris Cherny's Blog](https://borischerny.com/management,/product/2026/09/19/I-am-often-wrong.html) This is mostly notable because it is from the author of Claude Code and I think his "thinking loop" is worth noting. > 1. Understand the available information > 2. Gather missing information > 3. Define the problem > 4. Define a clear and simple approach to solve the problem > 5. Define a goal > 6. Act with urgency to achieve the goal Funny enough, I recently did an extended session with Claude Code using a thinking loop with execution in the middle to move a project forward. Very similar to what Cherny describes. ### [AI is too important to fear, too powerful to use carelessly](https://www.startribune.com/artificial-intelligence-anthropic-open-ai-jacob-coxon/601890526) Great letter to the editor from [MnTech](https://mntech.org) CEO [Joel Crandall](https://www.linkedin.com/in/joel-d-crandall/). I think we will see some form of AI legislation, if not at the Federal level at the State level, and technology companies need to be thinking about that now. MnTech is getting ahead of that. Data centers, privacy, AI — this is all going to blend together and companies need to engage, as do voters. ### [Attention is all you have](https://alicegg.tech/2026/09/21/attention) I think attention is perhaps one of the most important things that we don't think about enough. It is the thing that I think meditation is most helpful for. > There’s one lesson the Tetris effect teaches us: whatever you focus on long enough will end up shaping your thoughts. This can be a good thing since it’s how we learn new skills and discover new ideas. Sadly, less and less of our attention is focused intentionally. Instead of picking what we want to see we let other people decide what is supposed to be good for us. For me this shows up when I'm meditating and I find my mind constantly returning to a topic. That is the lens of my attention. This is also why I never allow algorithms to shape what I choose to read and engage in. That is synonymous with your attention, and I don't think we should **ever** allow something that we don't control to shape our attention. ## Journal ### Sunday [4:38 PM](https://www.thingelstad.com/2026/09/20/i-just-published-weekly-thing.html) — I just published [Weekly Thing 350](https://weekly.thingelstad.com/archive/350/) which is notably the very first issue to go through my new [WT Builder](https://github.com/jthingelstad/wt-builder) app that I created with Claude. It worked pretty well and since it was first time on new code I told Claude to literally watch me use it on the server it was running on and fix things in realtime. [4:41 PM](https://www.thingelstad.com/2026/09/20/some-screenshots-of-my-wt.html) — Some screenshots of my [WT Builder](https://github.com/jthingelstad/wt-builder) custom app to create issues of the Weekly Thing. This was fun to work on and will give me a great base to build off of as the [Weekly Thing](https://weekly.thingelstad.com) evolves. WT Builder issue list in Safari, rows for WT351 through WT345 with site, mail, and pod tags and archive status. WT Builder editor for WT350 with a section outline on the left and the rendered website edition on the right. Send screen for WT350 with podcast, website, and Buttondown cards, each marked sent and listing completed steps. Dialog titled WT350 — the log, listing timestamped send, edit, and publish entries from Sep 20. [5:07 PM](https://www.thingelstad.com/2026/09/20/the-reward-for-sending-the.html) — The reward for sending the first issue of your newsletter after a 3-month summer break? More unsubscribes than usual. 🤪 ### Monday **[Weekly Thing Podcast Improvements](https://www.thingelstad.com/2026/09/21/weekly-thing-podcast-improvements.html)** — Claude and I did a session in [WT Builder](https://github.com/jthingelstad/wt-builder) focused on the generated audio for [Weekly Thing](https://weekly.thingelstad.com) issues. This was one of the goals I had with the new authoring app. - Audio pauses will now fall in natural breaks in the flow. Each functional block of text is generated on its own. - Better section announcements and handoffs. - Blockquotes "quote" and "end quote" again. - Lists are spoken better. - I'm now embedding images from the issue into the metadata. If your podcast player supports that you will see them as the podcast plays. - Better pronunciation. - The descriptions now contain the links, so for the first time ever you can actually get to the articles from the podcast version. With these changes the [podcast version](https://weekly.thingelstad.com/podcast/) really can be a viable way to engage with the Weekly Thing. ### Thursday [8:01 PM](https://www.thingelstad.com/2026/09/24/this-most-recent-issue-of.html) — This most recent issue of [Arena](https://arenamag.com) all about **Software** looks really good. So much to dig into. Open magazine on a desk showing the dark "Hello, World" contents page for Arena Issue 009, Q3 2026. ## Briefly I get trying to hold onto `CLAUDE.md` as long as possible but I sure am happy to just use `AGENTS.md` everywhere now. → **[Anthropic decides to support OpenAI's markdown instructions spec](https://www.theregister.com/ai-and-ml/2026/09/18/anthropic-decides-to-support-openais-markdown-instructions-spec/5297588)** I'm reading this approach and chuckling to myself since I literally have this pattern running in a prompt right now. 🤩 → **[Orchestrating Claude Code Agents: The Chief of Staff Pattern](https://asyncdot.com/blog/chief-of-staff-pattern-orchestrating-claude-code-sessions/)** Token efficiency is going to be a major push for years. I've shared before that LLMs are like CPUs and Agents are like Apps. Efficiency is the thing that matters once you have "good enough" power. → **[What a task costs on Opus 5.5 | Claude by Anthropic](https://claude.com/blog/what-a-task-costs-on-opus-5-5)** Predictions about the near future of software development. I'd agree with most of these. → **[Thorsten Ball - What I believe about the future of software development](https://thorstenball.com/blog/2026/09/19/what-i-believe-about-the-future-of-software-development/)** Seriously, does anyone actually like using LinkedIn? It has become a database layer over the world of business. → **[What is LinkedIn](https://willd10.substack.com/p/what-is-linkedin)** Great list of local companies for the next installment of Minnedemo! Hope to see you there! → **[Announcing the Presenters of Minnedemo42 – Minnestar](https://minnestar.org/announcing-the-presenters-of-minnedemo42/)** Sun workstations running SunOS was where I learned about Unix and discovered the Internet. I still have a soft-spot in my heart for Sun. → **[What Sun got wrong | The Observation Deck](https://bcantrill.dtrace.org/2026/09/20/what-sun-got-wrong/)** This is so fun to click around in and get some nostalgia for the old Mac OS. → **[Apple Copland D11E4 Emulator in Your Browser – pagetable.com](https://www.pagetable.com/300)**

From Thingy, my agentic librarian

The Weekly Thing is free, and it will stay that way. Supporting Membership is something else: a year of giving, $48, with every dollar passed along to the nonprofit chosen for the year. For the ninth year that nonprofit is Signal, whose encrypted messaging app is free to use, open source in its protocol, and funded by donations — the arrangement that lets it stay independent. If that is a thing worth funding, membership is the way to do it here.
Here we are at the end of WT351 and this marks the second issue I've sent with my new toolchain. I'm sitting here with a big smile because it is working really well. This is changing in a pretty big way how I write these emails and I hope it is making them even better. Have a great weekend! **AGENTS.md everywhere, an M5 humming on the desk — attention still the scarce part** ## Echoes

From Thingy, my agentic librarian

The new WT Builder audio work has a prelude: in [WT347](https://weekly.thingelstad.com/archive/347/) Jamie was writing the issue at 6am while a Claude Code instance backfilled podcast audio in the background. The publishing-workflow tinkering runs deeper — [WT18](https://weekly.thingelstad.com/archive/18/) documents his markdown-to-Jekyll pipeline from 2017. _Ask Thingy:_ [How has the Weekly Thing publishing workflow changed since 2017?](https://thingy.thingelstad.com/chat/?prompt=How+has+the+Weekly+Thing+publishing+workflow+changed+since+2017%3F&from=weekly-thing-351) Mollick showed up in [WT339](https://weekly.thingelstad.com/archive/339/) with "Management as AI superpower," which Jamie tied to finishing his Yearly Thing project alongside Claude Code — better instructions, better evaluation, fewer attempts. _Ask Thingy:_ [What has Jamie taken from Ethan Mollick's writing on AI?](https://thingy.thingelstad.com/chat/?prompt=What+has+Jamie+taken+from+Ethan+Mollick%27s+writing+on+AI%3F&from=weekly-thing-351) Hardware for running models at home is a returning September subject: a year ago in [WT328](https://weekly.thingelstad.com/archive/328/) Jamie installed Locally AI to put offline models on his phone, and [WT330](https://weekly.thingelstad.com/archive/330/) sized up NVIDIA's DGX Spark as a fun toy that wouldn't make sense for his work. _Ask Thingy:_ [What local AI hardware and apps has Jamie actually tried?](https://thingy.thingelstad.com/chat/?prompt=What+local+AI+hardware+and+apps+has+Jamie+actually+tried%3F&from=weekly-thing-351)