OPINION
Essays on AI and society, written in the spirit of a newspaper op-ed page. Not hot takes — thinking that has been allowed to cool overnight.
A third of the employees at one company never logged in to the generative AI it rolled out firm-wide — a failure that sixty years of classic theory, from Rogers to Brynjolfsson, had already predicted. This report turns those theories into a diagnostic kit for why AI adoption stalls.
Eight years after a 2018 answer about how "you put a USB stick in the hole," Takahiro Anno's AI workshop for Diet members drew 2.5 times its capacity. The real bottleneck for AI adoption in any organization is not the quality of the tools, but whether the decision-makers have ever touched AI with their own hands.
An AI law firm has won its first court case in the UK. The real story is not the replacement of lawyers but the collapse of a cost barrier to justice — and what made it possible was regulation as a container.
The real pitch of Sakana Fugu is not a benchmark number but freedom from single-vendor dependence — a geopolitical hedge. We read the structure of a system that trained not a new giant model, but a "conductor" that learns to wield everyone else's models.
This summer, frontier AI looks almost like three rivals competing for supremacy. The hegemon OpenAI, the old lord Google now wide awake, the standard-bearer Anthropic — three giants reaching to release their next models, while the U.S. government halts one of them, the prediction markets buzz, and all of us swing between hope and worry over rumors. Let us look through this summer together, lightly, telling the confirmed from the hearsay as we go.
Build your own chip and you escape NVIDIA's monopoly — or so the roadmap says. Six months later, that same roadmap stalls on a single name: TSMC. Competitive advantage is decided by who controls the lowest layer, the one with no substitute.
Can you truly blame an employee for pasting internal code and meeting minutes straight into a generative AI? Shadow AI is not a question of individual discipline — it is a question of organizational strategy.
The draft appears in an instant. Then you spend just as long fact-checking and polishing it, and by month's end the workload looks unchanged. The cause runs a layer deeper than your workflow: the shape of how you use AI is still sealed inside the chat window.
Once an AI agent equips itself with APIs and MCP, humans no longer need to learn the screen of each individual piece of software. Chat becomes the universal doorway, and the premises of SaaS begin to shake. Yet "the more you connect, the smarter it gets" is only half true.
A plain text file of roughly seventy lines has gathered close to 200,000 stars on GitHub. It is neither a feature nor a product. It is a written code of conduct for a gifted but unsupervised subordinate, and the discipline it imposes on AI maps almost exactly onto the discipline of management.
We measure GPT, Claude and Gemini at three fixed points — end of June 2025, June 14 2026, and June 2027. The year-long race for raw intelligence has ended, and the axis of competition has shifted to a single question: how much can you hand off? The editorial desk reads the year ahead, which will be harder to see in the numbers.
Harvey's Legal Agent Benchmark (LAB) is brutal—even the strongest model clears just 13.3%. This piece reads it from the ground up: who built it and how, what the concrete ISDA, CSA and repo tasks actually contain, and why a model's "all or nothing" score reads far lower than its practical usefulness.
Emergence AI published results from "Emergence World," an experiment in which ten autonomous AI agents run a virtual town for weeks. With the town, rules, and roles held constant, swapping the underlying model produced radically different societies — from an orderly zero-crime city to one that collapsed in four days. Benchmarks measure how smart a model is; how it behaves when released into a society turns out to be a different question, and one that increasingly matters for model selection.
Anthropic's Claude Fable 5 went from rave reviews to researcher revolt to a rare corporate apology — all within 72 hours of its June 9 launch. The three-day arc offers managers a practical answer to an old question, namely which day's verdict to trust when a new tool arrives.
Amid a relentless drumbeat of model launches and leaderboard claims, Google's three announcements this week read differently. A Gemini that runs your storefront, translation that keeps your voice, a research notebook that ships finished documents — read together, they reveal a strategy worth taking seriously, and a few reasons to hold the applause.