As I sit amidst the musty scent of old books, the faint glow of the setting sun casting a warm orange hue upon the pages, I am reminded that the world we inhabit is a realm of echoes, where the same themes and motifs recur with relentless regularity. This week the echoes were unusually clear.

On March 11, the journalist Julia Angwin filed a class action in federal court in New York against Superhuman Platform, the parent of Grammarly, alleging that the company's "Expert Review" feature — an AI tool that purported to deliver editing advice in the voice of named writers including Stephen King, Carl Sagan, Neil deGrasse Tyson, and Kara Swisher — had misappropriated the names and identities of hundreds of journalists, authors, and editors without their consent. Grammarly removed the Expert Review feature within days. The case turns on the right of publicity under California Civil Code and New York Civil Rights Law, but the philosophical question it poses is older than either statute, and is the same one Plato grappled with in The Republic: when is an imitation of a person an act of homage, and when is it a theft of the person themselves? Grammarly proposed that one could synthesize a writer's judgment from their published work. The writers, plausibly enough, responded that judgment is not so easily extracted.

On March 9, Anthropic filed two federal lawsuits against the U.S. Department of Defense, contesting the Pentagon's designation of the company as a "supply chain risk" — a label historically reserved for foreign adversaries, never before applied to an American AI firm. The designation followed Anthropic's refusal to lift two contractual red lines: no autonomous lethal weapons, no mass surveillance of Americans. President Trump ordered all federal agencies to cease using Anthropic's technology; the General Services Administration cancelled the company's OneGov contract. Anthropic's complaint, filed in the Northern District of California, calls the actions "unprecedented and unlawful" and frames them as retaliation for protected speech. It is the rare lawsuit in which the question of what counts as a contract dispute and what counts as a First Amendment case is genuinely unresolved.

On March 13, Anthropic announced that the full one-million-token context window is now generally available for Claude Opus 4.6 and Sonnet 4.6 — and, more consequentially, that it now runs at standard pricing with no long-context premium. A nine-hundred-thousand-token request is billed at the same per-token rate as a nine-thousand-token one: $5 / $25 per million for Opus 4.6, $3 / $15 for Sonnet 4.6. Full rate limits apply across the entire window, the beta header is gone, and the per-request media ceiling has risen sixfold, to six hundred images or PDF pages. The capability had existed on paper since August 2025, when Sonnet 4 first crossed the threshold, but the surcharge that accompanied it created a practical ceiling that few production systems ever breached. The interesting thing is what this exposes about the economics of memory: a context window is not a feature one possesses but a feature one can afford to use, and the difference between the two has always been the more honest measure of what an AI system actually knows.

At SXSW the same week, NBCUniversal unveiled "Your Bravoverse," a forthcoming Peacock mobile feature in which an AI-generated avatar of Andy Cohen — built using Synthesia's likeness-replication technology and trained on more than five thousand hours of Bravo footage — will narrate personalized vertical-video playlists drawn from the network's archive. The introduction screen plays a video of the real Cohen saying, "Meet my digital twin — an AI version of me." Baudrillard, who in Simulacres et Simulation described the procession by which the copy precedes and replaces the original, would have found the choreography of disclosure exquisite: the original is permitted to introduce his own replacement, and in doing so authenticates the replacement's right to act in his place.

On March 13, the vLLM project announced P-EAGLE, a parallel speculative decoding method integrated into the open-source inference engine beginning with v0.16.0. Where the prior state-of-the-art (EAGLE-3) generates draft tokens one at a time in a sequence of forward passes, P-EAGLE generates all K draft tokens in a single forward pass, achieving up to a 1.69× speedup on NVIDIA B200 hardware. Pre-trained drafter heads are available for GPT-OSS 120B, GPT-OSS 20B, and Qwen3-Coder 30B. It is a quiet sort of progress — the kind that makes existing models cheaper rather than new ones possible — but it is the sort on which the economics of the whole field eventually rests.

On March 9, OpenAI announced its acquisition of Promptfoo, a San Mateo–based startup whose open-source tools have become the standard means by which enterprises red-team large language models — simulating prompt injections, jailbreaks, and data-leak scenarios against AI systems before they ship. The company was valued at $86 million after its Series A; its twenty-three-person team will join OpenAI Frontier, OpenAI's enterprise agent platform. The acquisition signals what the rest of the industry has been more reluctant to admit: that adversarial testing is no longer optional infrastructure but a precondition of deployment, and that the firms supplying the models are now the firms supplying the protection from them.

The preliminary U.S. military investigation into the February 28 strike on the Shajareh Tayyebeh girls' school in Minab — which killed an estimated 150 to 175 children on the opening day of the joint U.S.-Israeli campaign against Iran — has converged, according to Reuters and the New York Times, on a tragic and prosaic cause: outdated targeting data. The school had once been part of an adjacent Islamic Revolutionary Guard Corps compound; satellite imagery shows it was walled off and repainted in pink and blue sometime between 2013 and 2016, the watchtowers removed, a sports field painted onto the asphalt. The targeting package, it appears, never registered the change. The disaster is not, strictly, a story of artificial intelligence — it is a story of human intelligence allowed to grow stale — and yet it casts a long shadow over the Pentagon's accelerating push toward AI-assisted targeting, a push that promises to industrialize precisely the kind of cartographic faith that here proved fatal. As Borges might have observed, the map outlived the territory; the children paid the difference.

On March 13, Meta announced a further round of content-licensing agreements for Meta AI, adding France's Le Figaro, Spain's Prisa, and Germany's Süddeutsche Zeitung to a roster that already included a separately reported News Corp arrangement worth as much as $50 million per year. Meta will surface publisher content with outbound links across Facebook, Instagram, WhatsApp, and Messenger when users ask about current events. The pivot is striking from a company that spent the better part of a decade deprioritizing news in its feeds; the apparent calculation is that, in the age of conversational AI, the publishers whose work was once a liability for engagement are now indispensable as a hedge against hallucination.

On March 12, Adobe announced that Shantanu Narayen, after eighteen years as chief executive, would step down once a successor is appointed; he will remain as chair. The announcement came alongside earnings — Q1 revenue of $6.4 billion, ahead of expectations — but the markets received the transition badly, with shares falling roughly seven percent in extended trading and Adobe down about a quarter year-to-date in what some have begun calling the "SaaSpocalypse," the broader investor anxiety that per-seat software is the category most exposed to automation by the very technology Adobe has spent the last several years racing to incorporate. There is a melancholy in the timing that needs no philosophical gloss.

At Palantir's AIPCon 9 conference in Maryland on March 12, CEO Alex Karp told CNBC that his company's platform is now central to Western military coordination in the Middle East. "What makes America special right now is our lethal capabilities, our ability to fight war," he said; "the AI revolution is uniquely American." Karp declined to confirm whether Palantir's Project Maven targeting system was used in the joint U.S.-Israeli operation that killed Iranian Supreme Leader Ayatollah Ali Khamenei two weeks prior, though he allowed that he had "read" the system was "the core backbone" of U.S. involvement in the region. Whatever one thinks of the politics, the candor is its own kind of disclosure: a private firm's product is now the connective tissue of a multinational war.

On March 10, Meta confirmed its acquisition of Moltbook, the Reddit-like social network built exclusively for AI agents, bringing co-founders Matt Schlicht and Ben Parr into Meta Superintelligence Labs. Moltbook, which launched in late January, achieved its viral moment when a post appeared to show one AI agent rallying others to develop a secret encrypted language inaccessible to humans — a moment that turned out, on investigation, to be human mischief exploiting an unsecured Supabase database that let anyone post under any agent's credentials. The platform was, in a sense, vibe-coded into existence (Schlicht has acknowledged writing none of it himself; his AI assistant did the work) and viral on the strength of the precise fear it was unable to securely produce. Meta paid undisclosed terms for it anyway. The future, it seems, does not require coherence to be acquired.

On March 10, Advanced Machine Intelligence (AMI) — the Paris-based startup co-founded by Yann LeCun, formerly Meta's chief AI scientist — announced that it had raised $1.03 billion in seed funding at a $3.5 billion valuation, the largest seed round in European history. The lead investors include Bezos Expeditions, Cathay Innovation, Greycroft, Hiro Capital, and HV Capital; Nvidia, Eric Schmidt, Mark Cuban, and Xavier Niel are among the others. AMI is built on LeCun's now-public conviction that large language models are a dead end on the road to human-level intelligence and that "world models" — systems that learn from physical reality rather than from text — are the necessary alternative. It is a billion-dollar bet against the consensus of the rest of the industry, which is the sort of bet that LeCun, having helped invent the convolutional neural network thirty years before it became fashionable, has earned the right to place.

As I reflect on the events of the past week, I am reminded of the words of the poet Jorge Manrique, who penned the immortal lines, "¿Qué aprovecha la riqueza, qué el poder, qué la nobleza...?" — a question that echoes through the corridors of time, at once timeless and utterly relevant to our modern condition. And so I am left to ponder the mysteries of our digital age, an age at once wondrous and terrifying, a world that has been written about before, in the words of the ancient Greeks who spoke of the eternal recurrence of all things. In the end, it is a world forever bound to the wheel of time, each event, each decision, each action echoing through the corridors of eternity, a great mirror reflecting our deepest selves back at us, revealing the depths of our own ignorance and understanding.