How to read this
📰 Quick hits
The headlines worth knowing even if you read nothing else this week.
via Hacker Newsletter
Meta releases Muse Wonder and Muse Spark Open models
via Hacker Newsletter
via Stratechery
SpaceX buys more spectrum, moving toward primary cellular coverage
via Stratechery
Meta and Google add an AI-assisted coding round to their interview loops
via The Pragmatic Engineer
ChatGPT launches virtual try-on
via Retail Dive: Tech
🔍 Translation vs. understanding
The questionIs AI genuinely understanding, or just translating context into plausible output, and where does real human comprehension still earn its keep?
Grieving The Loss Of Details
purplesyringa.moe
First-person account of a low-level programmer whose identity rests on understanding how things work, as LLMs make explanation cheap.
"If an LLM can analyze the code and explain pointer provenance to you and babysit you until you get it, there is little value in expertise."
How to Be More Ambitious: A Tutorial for Humans With Models
Felix Rieseberg · felixrieseberg.com
Concrete prompting practice for pushing models up, down and across; the Depth section argues models switch vocabulary without switching methods and need an outside judge.
"When prompted to act like a professional, models often switch vocabulary without actually switching methods."
💰 Value concentration when creation costs collapse
The questionWhen building something gets cheap, where does the value (and the money) actually pool up?
38 Things We've Learned About Building AI Products
Ian Vanagas · PostHog
Retrospective from 10 shipped PostHog AI products with concrete numbers and admitted dead ends; names context, not the harness, as the durable advantage.
"A missing tool is a feature request written by a robot."
Systems Thinking Needs a Scorecard
Ben Yoskovitz · Focused Chaos
Practitioner building a usage-based AI product (Candor) shows how to measure whole workflows and unit economics, noting that cheap AI output shifts work onto readers, with the OMTM and open doubts about viability included.
"Silence is cheap to host. It can be expensive to misunderstand."
🪵 Thick engagement vs. thin optimization
The questionWhen is the slow, effortful, deep version of the work worth it, versus the fast and frictionless one?
I Know Why the Mathematicians Weep
Derek Thompson · derekthompson.org
Thompson and Strogatz on AI winning the craft of mathematics while interrupting the human conversation that gives it meaning; likely partly paywalled.
"it is succeeding brilliantly along the craft axis of science, but it is interrupting the thousand-year conversation."
🚀 Small teams, disproportionate output
The questionHow do tiny teams punch so far above their weight?
Redesigning source control at Anthropic
Anthropic
Seven people built a git replacement and review tool in five months; includes where Claude needed steering and why merge queues and CI strain at agent scale.
"Very often, we found that the role of the engineers was to tell Claude 'no.'"
🗺️ Planning artifacts shape the work
The questionHow do the documents you write (specs, decision records, the agent's workspace) steer what actually gets built?
What Customer and Partner Portals Look Like When Agents Show Up
Britton Russell · Rangle
Britton Russell from our team on portals as delegation layers, where an agent is one more link in the chain of who acts for whom. Every portal action becomes a permissioned interface an agent can call, the rules that live in people's heads get written down where agents can read them, and the company owns the whole thing so it can change as the traffic does.
"The identity and permission model is the product. Everything else in the portal sits on top of it."
How to Design a Good MCP Server
Refactoring
Practical design guidance for agent-called tools (jobs not endpoints, descriptions as runtime, production lifecycle) with Notion and Block examples of replacing 1:1 API mappings.
Are Harnesses Worth the Cost?
Naresh Jain · martinfowler.com
Paired-run experiment where a behavioral harness added about 10x tokens with no gain for the frontier model but lifted a cheap model from 3/10 to 10/10; the author revises his expectation.
"A requirement can be explicit in the brief and still lose influence during implementation."
🧬 Transmission of capability
The questionHow does knowledge and skill actually move between people, and from people to AI?
Building a Memory Pipeline (ai that works #77)
ai that works · Boundary
Practitioner retrospective on mining coding-agent sessions into CLAUDE.md changes, with a v1 that was bad, a weak/strong model cascade at about 2 cents per session, and the finding that the best fix was one line pointing at an existing architecture.md.
"The best memory isn't always a new fact."
👀 On the radar
Lower-confidence picks worth a skim if the topic grabs you.
- Why every company now needs to think and operate like a lab team | Josh Woodward Podcast with the Google Labs head on running experiments, killing ideas, and which skills rise in value; only show notes were available, adjacent topic.
- Software Factories Are a Local Maximum Short opinion piece arguing agents remove the friction that made teams ask why they build something; close to consensus but a useful framing.
- How To Effectively Debug When You No Longer Understand The Code Cindy Sridharan on debugging as an emergent skill when code is not understood; fetch was blocked by Cloudflare so judged from title and blurb only.
- How to tell if a startup has a moat Structural framework (switching costs, network, scale, data effects) for what stays defensible when building is cheap; framed for job-seekers and fairly standard strategy, with data effects the AI-relevant part.
- Building resilient AI-powered software, and planning for when things go wrong Sam Newman book excerpt on disaster recovery plus a section on resilience of AI-powered software (model provider changes, multi-model, observability); skim only the AI section.
- Uber's AI infra, from the inside Distinguished engineer tour of Uber's model gateway, harness wrapper, token-cost controls and code review as the new bottleneck; real agentic dev practice but at enterprise scale.
- New trend of building internal vibe-coding apps at tech companies Ramp's four-person team and Stripe's one engineer built internal Lovable-style tools for non-engineers in weeks; short, full text in the email.
- AI Is a Tool. Using It Well Takes Skill. Rebuttal to Rich Hickey's 'AI is Not a Tool'; stays at definitions and analogy, though it restates that guiding requires expertise.
- I paid people to try and follow my README Short practitioner retrospective: watching real people follow a README aloud surfaced assumptions his own testing missed.
- The Brain Sandwich Brain, then AI, then brain workflow to avoid skill atrophy; a simple heuristic with no structural analysis.
- Put a Name on It Named single-owner projects with a weekly written-update rhythm as orgs flatten; familiar DRI advice with a concrete cadence.
- What on earth is going on with AI pricing and packaging? Lovable's growth lead on why AI pricing is opaque when the cost of the work is unpredictable, unlike seat-based SaaS.
- How Meta designed Muse: an AI agent people actually want to talk to (Alex Cornell) Design director on progressive disclosure and a single persistent chat for an agent; adjacent to the interests, and visible chain-of-thought as a latency workaround is a fresh angle. Based on the newsletter summary only; the podcast was not fetched.
- The People Who Outsource Emotional Labor to AI Interview-driven reporting on people running difficult texts through chatbots; names the cost of delegating inner work but mostly settles on consensus, so a quick read.