Project Mariner is dead. Turns out browsers were always the wrong layer.

article · openai · may 09

Why for you: The revenue race between the two leading frontier labs is now the clearest leading indicator of which company's infrastructure bets will compound. OpenAI hit $25B ARR this spring; Anthropic, still far behind in absolute dollars, is growing proportionally faster. The headline number is interesting but the closing-gap detail is the one to watch. In a market where compute contracts get signed against credible revenue trajectories, this gap will determine which lab gets the better terms from NVIDIA, AWS, and Google Cloud over the next 18 months.

article · ai agents · may 07

Why for you: Mariner operated at the browser screenshot layer — seeing the screen the way a human would, clicking and filling forms visually. Google killed it May 4, folding the technology into Gemini Agent and Chrome's auto-browse features. The reason is instructive: file- and code-level agents have become faster, cheaper, and more capable than vision-based browser agents, and the gap is widening. The clean architectural signal for anyone building agentic workflows: if your agent needs to interact with a UI, the question isn't "can I train it to read screenshots" — it's "can I expose this as an API instead."

article · ai policy · may 06

Why for you: Six months ago, the administration's posture was hands-off; today, Commerce is pushing a pre-launch evaluation framework and a White House AI oversight executive order is reportedly in draft. The proximate cause: Anthropic's Mythos model, capable of identifying and exploiting previously unknown security vulnerabilities, landed on national security desks in a way that made the default hands-off stance politically untenable. Regulatory direction, once set, compounds. This is the moment the needle moved — worth marking regardless of which party you think will own AI policy in 2027.

article · ai regulation · may 2026

Why for you: Connecticut's SB5 is the most substantive state AI legislation to pass in the US so far — not a principles statement, but an operational law with specific prohibitions on companion bots targeting minors, algorithmic accountability requirements, and transparency mandates for high-risk AI systems. Governor Lamont is expected to sign it. The compounding dynamic: when the strictest state sets the floor, multistate enterprises often find it cheaper to comply everywhere than to maintain different stacks per jurisdiction. If Connecticut's standard becomes the de facto national baseline, it changes how you architect AI features today.

article · ai business · may 10

Why for you: India's major IT services firms — TCS, Infosys, Wipro, HCL — built their model on labor arbitrage: sending skilled engineers to handle work too expensive to do in-house at Fortune 500 companies. OpenAI and Anthropic are now going directly to those same Fortune 500 companies with pre-built AI agents for consulting, financial services, and software development. The displacement risk isn't "AI replaces Indian engineers" — it's "AI lets enterprises skip the services layer entirely." This piece maps who is most exposed and how quickly the revenue pools are shifting. Good structural read on where the agentic enterprise stack lands.

event · ruby on rails · may 12

Why for you: General admission tickets ($699) go on sale May 12 at 10am CT — tomorrow — for Rails World at Palmer Event Center, Austin, September 23–24. This is the largest Rails World yet at 1,200 seats, and previous years sold out in minutes. The CFP also closes May 16 if you're considering submitting. Set a calendar reminder now if you want to go. The Austin venue and expanded footprint signal the Rails Foundation is treating this as a full-industry event, not just a community reunion — which means the talks will skew more production and less framework-internals.

article · ai ethics · may 11

Why for you: Representatives from the Hindu Temple Society of North America, the Sikh Coalition, the Greek Orthodox Archdiocese, and several other communities joined OpenAI and Anthropic executives in New York for the inaugural "Faith-AI Covenant" roundtable. The labs' framing: religious traditions have reasoned about ethics, consciousness, and harm for centuries — that accumulated thinking should inform values tuning. The cynical framing: this is a soft-power move to broaden the coalition arguing against hard regulation. Both framings can be true simultaneously, which is exactly why it matters for how Congress frames AI ethics debates this fall.

newsletter · simon willison · may 04

Why for you: Simon's monthly digest is the most consistently useful single-page orientation in the agentic engineering space. The April edition covers the month's most significant agent architecture developments, notable papers, and tools — useful if you've been heads-down in product and need to catch up on a month that included the Anthropic managed-agents announcements, several significant agent memory papers, and a few architectural debates worth knowing about. Dry, precise, no hype — exactly what you'd expect from the person who has been writing "things I learned this week" longer than most people have been using AI tools.

essay · noahpinion · may 06

Why for you: Noah Smith argues that development economics has spent too long on "small-bore empirics" — RCTs measuring whether cash transfers improve school attendance rates — and not enough on the big question the field was built to answer: why do some countries grow wealthy and others don't. The critique maps cleanly onto AI research: enormous quantities of benchmark results, relatively little sustained attention to which architectures and training approaches will matter in a decade. Worth reading as a frame for evaluating whether any research agenda is actually asking the right question at the right level of abstraction.