Best Things I Watched
All-In: China Catches Up in Coding and the AI Memory Crunch
I listen to All-In every week, and this one was particularly good on open-weight models from China, memory constraints, and more.
The memory crunch part also matters. AI scaling is not just about more GPUs. It is about bandwidth, memory, power, and the cost of serving all this intelligence.
The next few years may be less about one lab running away and more about a messy stack of closed models, open weights, routers, and specialized systems.
Farhan Thawar: What Is Your Job Now?
Farhan’s line that stuck with me: “The code is not the collateral. The learning is the collateral.”
AI makes building faster, but it also makes it easier to produce a lot of output without understanding what actually worked.
News
Notion now works with Claude and Cursor
This is a big deal because it lets agents operate where a lot of nontechnical work already lives: docs, tasks, project plans, databases, and notes.
The important shift: cloud agents are moving from developer tools into the default operating systems of knowledge work.
Tweets Worth Saving
AI is helping more businesses get started and flourish
AI lowers the cost of starting: prototypes, landing pages, research, customer support, sales copy, ops, and internal tools all get cheaper.
That does not just mean more startups. It means more weird, specific, profitable little companies that previously would have been too expensive to try.
This might be one of the more underrated economic effects of AI: more shots on goal from more people.
I am working on a longer post on this.
My current view: useful agents are not magic chatbots. They are systems with a narrow job, the right context, the right tools, and a clear way to check their work.
The hard product work is deciding what the agent should know, what it should do, when it should stop, and how a human should review it.
Skills feel like one of the most underrated primitives for agents.
They turn repeated know-how into something reusable: instructions, preferences, examples, scripts, and domain-specific workflows.
This is where agents start to feel less like random prompting and more like compounding institutional memory.
Builds
Ramp’s Inspect is becoming a digital coworker
Rahul said 75%+ of code at Ramp now comes from Inspect.
That is a wild number, but the more important part is what Ramp invested in around the model: full dev environments, internal context, tooling, verification, and review.
This is the clearest pattern for serious coding agents: do not just give the model a repo. Give it the company operating system around the repo