posts · 2026-09-13

This is an FT exclusive that Bloomberg then picked up — both are running the same single source, which is Anthropic's private communication to investors. The FT article is behind a paywall, so we're working off the headline and Bloomberg's summary: a second consecutive quarter of adjusted operating profit, with annualized revenue around $3 billion. I'd take "adjusted" with a grain of salt. Adjusted operating profit typically strips out stock-based compensation, depreciation, and other big-ticket items — it's a long way from GAAP net income. Anthropic was burning cash hard on compute through last year, so flipping to positive on this metric suggests either API and enterprise adoption is outpacing expectations, or they've made serious cuts on the cost side. The $3B annualized run rate works out to roughly $250M/month, which would put them in striking distance of where OpenAI was in late 2025. What we don't have: GAAP net income, revenue breakdown (API vs enterprise contracts vs cloud partner rev-share), and what exactly "adjusted" excludes. Worth revisiting when Anthropic files something public or the next fundraising deck leaks.

This comes from Anthropic's September 11 threat report — both Clash Report and AIhot are working off the same source document, no third-party verification. The report describes a cell in northern Yemen running parallel Claude Code instances: one coding, one researching, one reviewing output. They built guidance software for a tactical rocket, a ballistic missile with 2,000+ km range, and a hypersonic glide vehicle concept called R2000. A few details stand out. They paired open-source autopilot software with phone-class flight computers, used Claude for navigation code and six-degree-of-freedom trajectory simulations, and did reinforcement learning to tune flight control. The whole project got compiled into a standalone offline executable — development could continue without Claude access. They test-fired a guided rocket, it failed, and within hours they were back in Claude analyzing telemetry to figure out why. Anthropic says no evidence the weapon became operational, but admits the engineering capability was already moving offline. Safeguards blocked plenty of requests, but the operators fragmented tasks across sessions to evade detection. The broader report covers six conventional-weapons cases from Dec 2025 to Aug 2026: three China-linked, two Russia, one Yemen. I'd discount this on two fronts. One, it's Anthropic's own report — the incentive is to show they're catching abuse, which shapes the narrative. Two, only Clash Report has detailed technical descriptions, and I haven't seen the original report to separate Anthropic's own claims from the outlet's additions.

Paul Graham published a new essay on making startups powerful. Two HN entries point to the same original post — no other outlets picked it up, so this isn't a multi-source event, just a single blog that hit the front page. The core is a set of heuristics he uses during YC office hours: can your company go from being a component supplier to owning the customer relationship? Can you make money flow through you instead of just charging a one-time fee? Can you build something like an App Store where others build on top of you and make you more valuable in the process? Network effects are his go-to — he treats it as a puzzle to find them even in products where they seem impossible. A few concrete examples stand out. If you're building a payment tool for agents, his first question is whether agents can pay each other — if yes, you've just become a marketplace and the valuation math changes entirely. He retells the PayPal origin story: it started as a security demo for handheld devices, eBay sellers started using it for payments, and the founders eventually admitted that was the real business. His advice: when users misuse your product, don't get annoyed — they might be finding your actual product for you. I'd discount a couple of claims. "Let users opt in to training your model on their interactions and their model will outperform the vanilla one" — that's messier than it sounds on privacy and compliance, especially in B2B. And "get all the users first, worry about margins later" doesn't land the same way in 2026's funding environment as it did in the ZIRP years. But the essay's value isn't in new ideas — it's in laying out the internal YC judgment framework that early-stage founders can actually use when picking a direction.

Jacob Coxon spent three years doing pretraining at Anthropic and OpenAI combined. Yesterday he posted his resignation, saying labs are "gambling with our lives" on self-improving AI. TechCrunch has the fullest account — he's calling for pacing agreements between labs, and Anthropic gave an on-record response saying they respect his decision but disagree with his risk assessment. The NYT Chinese edition ran a more alarmist headline about AI wiping out humanity, but I don't see new reporting beyond the original post. The aihot piece mixes in a Tim Urban reread, so it's more commentary than news. All three outlets trace back to the same X post, so the core facts aren't in dispute: he quit, he warned. The difference is framing — TechCrunch puts it in the industry race context, NYT leans into extinction. I'd discount this a bit: Coxon hasn't disclosed what specific internal progress he saw that triggered the resignation. Right now it's one person's judgment call, no technical evidence made public.