Cursor takes on GitHub with Origin, Anthropic's leaked $65 billion run rate sets up the IPO valuation fight, and Stripe closes its $7 billion OpenRouter deal — then the main event: NLW's five-skill AI engineering map for knowledge workers, from capability mapping to rapid skill acquisition, all resting on a foundation of domain judgment.
We're moving from doing our work to managing agents that do our work — and that shift demands five new skills.
AI capability mapping. Context and harness management. Problem and product prototyping. New opportunity identification. Rapid new skill acquisition. Inspired by Andrew Ng's AI engineering skills map for developers, this is the knowledge-worker version — and it only works layered on a foundation of domain judgment, whether personal, borrowed from colleagues, or embedded in rubrics and context. None of it can simply be handed to you; you learn it by doing.
Cursor's new repo hosting platform is Git hosting plus better integrations for the coding agents you're already using — natural language queries across the code base, agents handling comments and commits, all without context switching or connectors. The launch is an early preview for existing customers.
Maybe Origin's biggest pitch is platform stability — GitHub's service has been widely perceived to be degrading over the past year, and right around the announcement it suffered a six-hour degradation. SpaceX AI's Matt Palmer quipped they'd have shipped earlier 'but GitHub was down.' At this point, complaining about GitHub is basically part of being a developer.
— Guillermo Rauch, Vercel CEO. The Vercel CEO piling on GitHub's outage in the middle of Origin's launch window — a rival platform boss twisting the knife on the incumbent's weakest point.
Cursor knows moving code is extremely hard, so Origin supports mirroring — leave your code on GitHub, let Origin sync it, and detach later if you like it. GitHub stays the system of record, theoretically minimizing the risk of breaking production workflows. Skeptics still want a longer security and infra track record before moving customer projects over.
A GitHub built from the ground up assuming agents push most of the code is what has people interested: triggering agents on code changes, running scheduled tasks from a single platform. If nothing else, the response shows the demand is there — but that doesn't make Origin's path easier, because ripping out core infrastructure like GitHub asks a huge amount of users.
Acquisitions stereotypically slow the acquired company's innovation; SpaceX AI seems determined for that not to be the case, with Origin landing just as the deal closes and Grok 4.6 billed as an early look at what the two will build together. The division of labor, per Cursor's blog: 'SpaceX is building the computing capacity needed to scale intelligence far beyond what exists today. Cursor will be one place where that intelligence becomes useful.'
Sources who've seen the investor update say Anthropic hit a $65 billion revenue run rate at the end of July — a sevenfold increase since the beginning of the year and roughly a 40% jump since the $47 billion figure disclosed in May.
The numbers matter because of the astronomical IPO valuation Anthropic is seeking: some investors expect $2 trillion, with some quoting an 800% growth rate to justify even more. The new figures put annualized growth at 550% — incredibly strong but slowing versus the first half — prompting author Tae Kim to remind everyone you can't extrapolate month-over-month growth acceleration to infinity. 'This is still insane growth.'
Combine Anthropic's $65 billion with the $40 billion annualized run rate OpenAI CFO Sarah Friar has been reporting and the two companies are at roughly $100 billion in revenue run rate — up from low single-digit billions at this time last year.
Below the rumored $10 billion, but a huge markup for OpenRouter, which last raised at $1.3 billion in May. The immediate discourse: did Stripe overpay for an AI gateway it could arguably build itself?
Against the 'surely Stripe can build its own gateway' chorus, one fintech engineer argued the crowd is anchoring on the wrong number: big acquirers pay relative to their own market cap, not the target's standalone worth — see Facebook/WhatsApp or SpaceX/Cursor. If buying OpenRouter improves Stripe's valuation prospects by more than five percent, the price pays for itself; competitive tension is how five billion becomes seven.
Stripe views itself as the core financial plumbing of the internet economy, and tokens are a new essential currency in that economy. Not all tokens are equal, users need to move easily between categories of them, and OpenRouter is the leading company doing that. To get the customers, infrastructure, and brand online immediately, you pay the price it takes to get the deal done — and my guess is it pays off for Stripe in big ways.
Knowledge work is moving from doing our work to managing agents that do our work. The five skills that matter in that transition: AI capability mapping, context and harness management, problem and product prototyping, new opportunity identification, and rapid new skill acquisition. Combined with a foundation of domain judgment, you get a knowledge worker more capable and more powerful than ever before.
Ng's four AI engineering skills — building and deploying AI applications, software engineering fundamentals, using coding agents, and shaping the build — are his answer to a noisy, hype-filled information environment. The key insight: even if agents write the code, deeply understanding software (cost, scalability, reliability, speed trade-offs) still drives better stack, architecture, and testing decisions. Foundations still matter; agentic coding is now its own skill on top.
The foundation under all five skills is the ability to define quality, recognize trade-offs, understand consequences, and take responsibility for decisions. Just because ChatGPT can write the copy and make the ad assets doesn't mean someone with no marketing experience can plan and execute a campaign well — and it's not marketing in general that matters, it's marketing in the context of your organization, the stuff that never got encoded into training data.
Domain judgment doesn't exclusively come from 10 or 20 years in a field. There's personal judgment — standards and pattern recognition you've internalized; borrowed judgment — expertise supplied through collaboration, review, and shared decision-making; and embedded judgment — expertise captured in examples, rubrics, policies, and evaluations you feed the AI.
AI can blow you away one minute and make a mistake the next that you wouldn't expect from the least capable intern — Ethan Mollick's 'jagged frontier.' Capability mapping means knowing what AI is natively good at, when a task needs an assisted approach versus workflow automation versus a truly agentic solution, which models and effort levels suffice, and how much human oversight is needed. Nobody can hand it to you; you learn it by trial and error, because your particular tasks may break the conventional wisdom.
Context management is making sure the AI has the information it needs — past campaign performance, analytics, customer feedback — rather than a bare prompt and a prayer. The harness is everything else surrounding the model: instructions, documents, tool access, permissions, memory. Some of this is governed by your organization, but a lot is customized per worker — and today's best practices will almost inevitably have changed six months from now.
Not marketers becoming software engineers — but when everyone can push code, a lot of previous work gets done differently. Instead of manually pulling campaign data from every platform into Excel, a marketer can build an internal dashboard plus the engine underneath it, ingesting via APIs with AI surfacing the first layer of insights. Domain judgment still supplies the translation layer to real actionable insight — but a huge amount of time is freed up for exactly that judgment-style work.
Knowledge workers should not think of themselves as turning into full product managers and software engineers — but being able to build things to solve problems and do parts of our jobs is perhaps the most significant shift in how knowledge work happens that we've ever experienced. If you've been on the fence, go messily clunk your way through Codex and Claude Code and figure out which parts of your work could be transformed.
If skill three is using code to solve existing problems, skill four asks what's newly possible: every knowledge worker has an endless list of things they'd do if time and resources were no limit, and agents make a much bigger portion of it viable. One useful trick: imagine your organization handed you a team of software engineers to do whatever you want with. Before long you hit totally net-new ideas — think small-company marketing teams building and releasing games as top of funnel.
The meta-skill that cuts across the rest: recognizing which adjacent skills have suddenly become valuable, creating space to experiment and learn by doing in real-world environments, and assessing whether the output is worth integrating into how you work. Skill acquisition becomes not the work of semi-regular upskilling seminars but a continuous, ongoing process — and there's far less inertia in how you work than in how your organization does.
If experienced workers use AI to do what younger workers previously did, how does the next generation ever develop domain judgment? One answer worth exploring: shift from thinking about AI as single-player to AI as multiplayer, with the core unit of AI moving outside the individual to the small team. A subject for a future show.