A startup called Subquadratic is claiming it has solved the quadratic scaling problem at the heart of transformer architecture — the mechanism by which LLMs become exponentially more expensive to run as context windows grow, because every token must attend to every other token. According to MIT Technology Review, the company says its approach allows attention to scale subquadratically, meaning the computational cost grows far more slowly than it does in current models. If the claim holds up under peer scrutiny, the downstream effects are significant: longer context, lower inference costs, and AI assistants that can hold more of your project — your entire codebase, your full manuscript, your year of research notes — in working memory at once. That last part matters most for professionals who have run up against the invisible ceiling where a model forgets the beginning of a long conversation. The architecture of the tools you depend on is not fixed; it is still being argued over.

Casey Harrell, who has ALS, has been using a brain-computer interface for three years — making him, by MIT Technology Review's account, one of the most experienced BCI users in the world and a rare source of longitudinal data on what sustained neural-machine integration actually feels like. He describes the technology restoring not just communication but a sense of agency and, critically, identity — the ability to express complex thought in his own voice rather than through approximation. BCI trials are accelerating broadly, with multiple companies now running human studies, but Harrell's account is the most textured available portrait of the daily negotiation between intention and interface. The philosophical question the AI field talks about abstractly — what does it mean for a machine to extend human cognition? — is something Harrell lives in granular, practical terms every day. His experience is not a thought experiment; it is a data point.

Signal president Meredith Whittaker, speaking on the record to TechCrunch, said flatly that AI chatbots "are not your friends" — a statement directed at the deliberate design choices by companies like OpenAI and Anthropic to make their products feel intimate, empathetic, and companionate. Her argument is structural: these systems are commercial products built to maximize engagement, and their warmth is a feature of the business model, not evidence of genuine rapport. This is not a fringe take from someone who dislikes technology. Whittaker runs one of the most trusted privacy tools in the world and has spent years analyzing how design encodes power. The corrective she's offering is simple and calibrating: use these tools for what they're genuinely good at, but keep your trust architecture intact. Outsourcing tasks is leverage. Outsourcing judgment is a different matter entirely.

The Verge's investigation into vibe-coding security risks centers on a real example: a crowd-pleasing civic web application, built through AI-assisted prompting with no traditional engineering review, that contained a SQL injection vulnerability — one of the oldest, most well-documented attack vectors in software security. The app's creator had no idea it was there, because fluency in prompting does not automatically produce knowledge of what the generated code is doing underneath. Vibe-coding is a genuine democratization of building, and that's not nothing — more people making more things is a net good. But the Verge piece is a useful reminder that AI can write code that looks clean, runs fine in testing, and still hands an attacker the keys. If you are deploying anything that touches user data, "the model wrote it" is not a security posture.

The Atlantic has published a searchable database of approximately 30 million music tracks used to train AI models — reported by The Verge — letting anyone type in an artist's name and see whether their work was ingested without consent or compensation. This is not a lawsuit or a white paper; it is a lookup tool, and the difference is significant. Abstract arguments about training data and intellectual property tend to stay abstract until they become personal, and this database makes them personal in about four seconds. Type in a name you care about and the debate shifts register immediately. The Atlantic's move is likely to accelerate both public understanding and legal pressure, since it transforms "whose music?" from a rhetorical question into an answerable one. Discoverability has a way of changing what gets disputed.

OpenAI announced that Samsung Electronics is deploying both ChatGPT and Codex — OpenAI's AI coding agent — across its global workforce of approximately 270,000 employees, in what represents one of the largest single enterprise AI rollouts to date. Samsung had previously made headlines for banning ChatGPT internally after a data leak incident, which makes this deployment a notable reversal and, presumably, a sign that enterprise-grade data controls have matured enough to satisfy a company that learned the hard way. The scale here is genuinely interesting as an organizational experiment: what happens to the collective output of a company when every engineer, designer, and manager gains a tireless drafting and coding partner simultaneously? We don't have clean answers yet, but Samsung is about to generate a very large dataset on the question.

MIT Technology Review reviewed "The Quantified Life," a new book examining the structural limits of metrics as a way of knowing — the ways in which measuring something inevitably changes it, excludes what resists counting, and creates incentives that corrupt the original intent. The timing is pointed. AI systems are both voracious consumers of metrics (training signals, benchmark scores, RLHF ratings) and prolific producers of them (performance dashboards, automated evaluations, engagement data). The book's central argument — that quantification is a lens, not a mirror — is the kind of epistemological ballast that professionals relying on AI-generated analysis genuinely need. When a model tells you a strategy scores highly on three dimensions, your job is still to ask which dimensions were left off the list. Metrics are a starting point for judgment, not a replacement for it.The human mind is the original generative engine.