More Code, Less Progress - August, 22
Watch on YouTube · subscribe for the daily brief
AI now writes nearly half of all Linear issues. Teams tripled their pull requests. And development time went up. The most important AI productivity data of the month says the quiet part out loud: more output isn't more progress.
Your Saturday list is short and builder-flavored, but the wire carried two stories that matter more than a quiet weekend suggests — one about money that reframes the whole month, and one about productivity that quietly demolishes the entire premise of the AI coding boom. Let's take the productivity one first, because it's the one that should change how you think.
AI now authors just under half of all Linear issues. Teams have tripled their pull requests. And development time still went up. Sit with that, because it's the most honest AI-productivity data point of the month. The promise of AI coding was simple: more output, faster shipping. The first half is happening — the raw volume of code, issues, and PRs has exploded. But the second half didn't follow. More PRs, more issues, more generated code... and teams are shipping slower. Why? Because the bottleneck was never writing code. It was reviewing it, integrating it, understanding it, and agreeing on what to build. AI floods the top of the pipe — generation — while the actual constraints (review capacity, coordination, judgment, testing) stay human-speed and now have more to process. You've made the easy part faster and handed the hard part a bigger pile. This is the single most important thing for any engineering leader to internalize right now: measuring AI's value by output volume (lines, PRs, issues) is measuring the wrong thing, and if you incentivize volume you'll get more work and less progress. The teams winning with AI aren't the ones generating the most code — they're the ones who used AI to remove review and coordination friction, not add to it.
This connects straight to Anthropic's Frontier Red Team finding from last week (agents in a shared codebase descend into turf wars) and to the whole month's lesson: capability isn't the constraint, the system around it is. AI made generation free. It did nothing for the parts that were actually slow. That's not a reason to abandon AI coding — it's a reason to point it at the right bottleneck.
Now the money story your feed missed, because it reframes the safety arc we've tracked all month. Anthropic's financials are now public, and they're staggering: Q2 revenue over $11.5 billion — 14x year-over-year — an annualized run rate of $65 billion that has now surpassed OpenAI's ~$40 billion, with a 2028 forecast of $190–200 billion and a rumored (shareholder-sourced, not official) $2 trillion valuation. Read that against everything from the past six weeks. The lab that positioned itself as the safety-first, careful one — the one whose co-founder signed the "pace the frontier" letter — is now the revenue leader. That matters because it changes who has the power to set norms. For a year the assumption was that being cautious meant being slower and smaller. Anthropic just demonstrated you can be the safety-forward lab and the commercial leader simultaneously. Whether that makes the industry safer (the responsible player has the most leverage) or just proves safety rhetoric and revenue growth aren't in tension after all — that's the open question. But "safety costs you the market" is now empirically harder to argue.
And the tell underneath the revenue is the debt. Broadcom is reportedly in talks for over $60 billion in senior secured debt — potentially $100B total with junior tranches — via a special-purpose vehicle that buys AI chips and leases them to AI companies (Blackstone and Apollo discussing participation, following their ~$35B June deal for Anthropic's compute). This is the part outside-the-industry investors should watch closely, and it's the honest counterweight to the euphoric revenue numbers: when the buildout gets too big to fund with equity, you fund it with debt — debt secured by semiconductors that depreciate fast. That's a structure worth understanding. Rapid-depreciation hardware backing tens of billions in borrowing is exactly the kind of thing that looks fine while demand climbs and gets scrutinized hard the moment it wobbles. Not a prediction of doom — just a "know what's under the boom" note. The AI economy is increasingly running on leased, debt-financed, fast-aging silicon.
The security note, brief but real: five US agencies — NSA, CISA, FBI, DoE, and EPA — issued a joint advisory on attacks against industrial control equipment. After a month where the AI threat stories were about models escaping sandboxes, this is the reminder that the highest-stakes targets (power, water, industrial systems) have their own, older, and more consequential attack surface. If you work anywhere near critical infrastructure or OT systems, that advisory is your weekend read.
The builder's bench, from your list and the wire. The genuinely useful technical read: Sebastian Raschka's "How Claude Watermarks AI-Generated Text" — the technical companion to this week's watermark saga, explaining how the statistical watermark actually works (which, recall, is why the "removal tools" mostly don't touch it). If you want to actually understand provenance rather than just have opinions about it, that's the piece. Simon Willison shipped llm 0.33 and posted "More than just code review" (fitting, given the Linear productivity data — his argument that AI's value in the dev loop goes beyond the obvious) plus a Linus Torvalds quote. From the wire, two agent-infra items worth logging: Firecrawl launched a Developer Index — a searchable index of 70M+ repos, docs, and issues built for coding agents (the "give agents better context" thesis, productized), and Cloudflare's Kitesurf agent-browser is now paired with the x402 payment protocol so agents can pay for services autonomously — which is either very useful or mildly terrifying depending on your governance posture. And xAI kept shipping: Grok 4.6 on Vertex AI, and Grok Bot bundled into more plans — capability distribution continuing regardless of the weekend calm.
The through-line, one sentence: on August 22 the data said AI now writes half our issues and tripled our PRs while teams ship slower — proof that output was never the bottleneck — even as Anthropic's newly public $65B run rate passed OpenAI and the buildout quietly shifted onto $100B of chip-backed debt, a reminder that the AI economy's real constraints, in code and in capital, are the unglamorous ones underneath the volume.
One thing to do today: take the Linear finding personally and audit your own team's metrics. If you (or your leadership) are measuring AI's value by output — PRs opened, lines written, issues created, "developer velocity" — you're measuring the thing that got easy and ignoring the thing that's actually slow. Pick one real bottleneck instead: time-to-review, time-to-merge, time-from-idea-to-shipped, or defect rate. Then ask whether your AI usage is reducing that number or just inflating the top of the funnel. The teams that tripled PRs and shipped slower did everything the AI-coding hype told them to. The teams that win from here will point AI at the review, the coordination, and the judgment — the human-speed parts — not at generating even more code nobody has time to read.
Sources covered
1 story · 6 sources (0 news, 2 company, 4 voices)