I am trying this as a weekly newsletter: a quick roundup of the AI things I try, watch, read, and save each week, which is a lot! I got the idea from Drew Bredvick, thanks, Drew!
Some themes:
The models are getting good enough.
The bottleneck is moving to context, workflow, tools, taste, and review.
The question is no longer “which model is smartest?” It is: who can make agents useful inside real work?
Videos + Podcasts
Dan Shipper on Lenny’s Podcast
Core idea: more automation does not mean less work. It means more work becomes possible.
Agents create more drafts, code, options, and analysis. That makes human review, judgment, PM taste, and coordination more valuable.
Benedict Evans on Lenny’s Podcast
Best framing: AI is like the internet in 1997.
Obviously important, still early, and most of the winning apps and business models probably do not exist yet.
Naval’s AI Industrial Revolution podcast
Excellent panel (Guillermo Rauch, Max Hodak, Blake Scholl, and of course the great Naval)
How to Build a Self-Improving Company with AI
Agents do not just need documents. They need a working map of how the company operates.
Decisions, workflows, exceptions, preferences, and tribal knowledge all need to become executable.
Charlie Holtz building Conductor
Developers are becoming managers of small AI teams.
Conductor feels like an early glimpse of the “AI manager” workspace.
Tyler Cowen on agents, consciousness, and the future of AI
Cowen’s useful point: AI probably will not create instant mass unemployment.
But it will create a brutal adjustment period. The winners are the people who take initiative, learn the tools, and move before institutions do.
Dan Shipper on Figma and the SaaS resurgence
The “SaaS is dead” narrative may be backwards.
If agents become power users of software, great SaaS products may get more usage, not less.
The next moat may be building software that agents can easily understand, use, and return to.
News
NVIDIA released Nemotron 3 Ultra
550B total parameters, 55B active. Built for long-running agent workflows. Another sign that agent orchestration is now a first-class model target.
From X
The new devtools moat is being easy for agents to find, understand, and wire up.
Vercel and Shopify prompt-to-store
Prompting a real Shopify storefront is a nice example of AI moving from demo to deployable workflow.
The best product a company builds may still be its culture.
Every company’s first AI strategy should be a skill library
Do not start with a giant AI strategy deck.
Start by capturing how your best people do the work.
Products I Tested
Lets Claude orchestrate many subagents through a reusable script.
Best for audits, migrations, research, and tasks where one chat thread is too small.
Turns Cursor’s agent into programmable infrastructure.
Useful for CI, internal tools, backend automation, and scripted coding workflows.
Reads your Claude, Codex, and Cursor sessions.
Builds a profile of how you work with AI: part mirror, part benchmark, part future hiring signal.