Micro

AI Is Already Finding Things in Auto Safety

Most of the AI talk I hear is about speed: faster reviews, fewer hours, cheaper reports. That’s fine, but it’s not the prize.

The real prize is AI that finds things, and I’m now seeing this work in practice, not just in theory. A cluster of owner complaints described in a dozen different ways. A warranty spike that quietly lines up with a new service bulletin. A trucking carrier whose inspection record is slowly getting worse. Signals like these sit across NHTSA, warranty and FMCSA data for months before anyone connects them. AI is helping build solutions to connect these signals, and the results have been great.

Catch them early and you change a recall decision, a case strategy, maybe a crash that never happens.

The catch: people have to see the reasoning. Otherwise AI just becomes a tool to demand lower rates.

🚗 📊

Personal Agents – Muse

Meta’s Muse is a personal agent that actually feels useful out of the box. Unlike the other agents floating around, it’s easy to use and comes with a solid set of capabilities right from the start. No long setup. No steep learning curve. It just works.

But this space is moving fast. What looks good today will feel basic in a few months. Personal agents are about to get much more capable — and much more common.

🤖

Crackdown on Subscriptions sharing

Spotify confirmed it’s rolling out QR-code household verification for Family Plan members — scan within 7 days or get locked out and downgraded to an individual plan. Spotify’s co-CEO calls it “a routine improvement” to 13-year-old verification tooling. Everyone else is calling it what it is: a crackdown on sharing with friends and non-cohabiting family, aimed squarely at Premium revenue.

Same playbook Netflix ran on password sharing in 2023, framed as protecting creators while users experienced it as paying more for the access they already had. Call it what it actually is: a crackdown wearing an “improvement” badge.

10x Engineer?

With Claude Code and other AI tools, the drudgery is out. No more keeping track of development environment quirks, package versions, API docs, broken devops, and incomplete/outdated project documentation. The stuff that the developers did not like doing and that ate hours but never felt like the actual work can be handled by the AI tools. There are many useful AI skills (prepare tickets, debugging, documentation), that engineers can adopt or develop on their own to take the drudgery out of the software development. Engineers are good at streamlining things and they have really embraced all these skills. I use them as well and love it all.

What’s left is the creative, productive part: deciding why & what to build, for whom, and why. Engineers, if wanted, can easily transition into the product management and marketing roles and attack the why, what, whom and why as well now. New ways and new models. What do we call this? Nx Engineer?!

New era, new slogan: drudgery out, focus in.

Personal AI Agent – Hermes

I have been running the Hermes agent on my Next laptop across multiple sessions. Its persistent memory is a cool feature which helps the agent store and retrieve from its memory between sessions. I can close a conversation and come back later without re-explaining the context. The agent remembers what we were doing.

Hermes is on a brand new installation of Omarchy Linux on a Lenovo laptop. Omarchy works great and I love it!

Hermes keeps getting smarter every day. It reaches for the skills we have built for different activities — the things I talk to Hermes about on a day-to-day basis — and turns them into real output, not just plans.

Hermes is like OpenClaw (which took the tech world by storm in early 2026). Hostinger gives a nice comparison between Hermes and Openclaw.

🤖

New model in town. What’s up Jev!

Been experimenting with Jev, Typesafe.ai’s “System 1” model — built for fast, intuitive responses rather than slow, deliberate reasoning. On its own it’s quick and cheap, but pairing it with a heavier LLM turns out to be the real win for certain workflows: let Jev make the fast first pass — triage, routing, a quick draft or gut-check — and only hand off to a deeper reasoning model when the task actually needs System 2-style deliberation.

For high-volume, low-stakes steps this cuts latency and cost without giving up quality where it matters. Right tool, right layer.

NeoClouds needed for the Frontier labs

Chamath’s deep dive on neoclouds gives some good info on NeoClouds. These AI-native clouds rent GPU hours 50–70% below hyperscaler list prices and get earlier access to scarce Nvidia chips. The key difference from traditional clouds: AI training occupies thousands of GPUs in lockstep for 1–4 months at a time, rack power jumps from 10–15 kW to 40–250 kW, and neoclouds strip out the software layers that get in the way. Frontier labs are bypassing AWS/Azure/GCP for compute. The open question: can the economics survive when customers are also building their own capacity and custom chips?

🏭 ☁️

AI Agents & Evals

Capability evals are about what you’re shooting for: “What can this agent do well?” You start low and hill-climb. Regression evals are about the baseline you’ve already reached: “Does it still do what it used to?” Once you’ve proven a capability, it becomes the floor — you don’t regress from there. (Anthropic on evals for AI agents) — a good post to understand AI agent evals. #goodread

✅ ❌

Wait… Are we are going to slow things down now?

Caught up on Pacing the Frontier — the open statement signed by 1,100+ employees across Anthropic, OpenAI, Google, and Meta asking the US government to back an international effort to deliberately pace frontier AI development. Notable that this isn’t outside critics; it’s people building the models themselves, including leadership at the labs, saying capability is outrunning the governance and safety tooling meant to keep up with it. Reuters coverage here.

Worth sitting with: the ask isn’t to stop progress, it’s to build the technical and governance muscle to pace it on purpose rather than by accident.