Chronological feed of everything captured from Casey Newton.
This paper investigates weakly orthogonally invariant Finsler metrics, providing explicit expressions for their Berwald and Landsberg curvatures. It then establishes a system of partial differential equations that characterize generalized Finsler warped product metrics under conditions of vanishing Berwald and Landsberg curvatures. The work culminates in the construction of examples of non-regular Landsberg metrics that do not conform to the Berwald type.
This paper introduces a biharmonic interpolatory subdivision framework to enhance data fairness and smoothness on Riemannian manifolds. It extends the six-point Deslauriers-Dubuc stencil from the Euclidean setting to constant-curvature surfaces, establishing fourth-order smoothness. The methodology demonstrates improved performance over existing schemes in terms of fairness energy and curvature profiles while maintaining locality.
The HEALthy Brain and Childhood Development (HBCD) study utilizes dMRI to map infant brain maturation, but such large-scale multi-site studies face challenges with inter-scanner variability. This research demonstrates that ComBat-GAM harmonization effectively mitigates site-specific effects stemming from different scanner models within the HBCD 1.1 data release. This method significantly reduces statistical differences and effect sizes across diffusion tensor imaging metrics, underscoring the critical need for rigorous harmonization in large neuroimaging datasets.
This paper introduces a novel framework for analyzing the training dynamics of deep visual networks using principles from dynamical systems and signal analysis. It defines an integration score, metastability score, and a combined dynamical stability index from layer activations. This approach offers insights into model training beyond traditional loss and accuracy metrics, revealing internal representation changes and potential early convergence indicators.
Meta has launched an AI-powered support chatbot across its platforms, representing an incremental improvement over its previously inadequate support infrastructure. However, the bot notably lacks the capability to restore suspended or locked accounts — the single most demanded support function among Meta's user base. The gap between what the bot offers and what users actually need highlights a persistent structural failure in Meta's user support model.
Spotify launched a beta feature called Artist Profile Protection that lets artists review and approve or reject releases before they appear on their profile — a direct response to AI-generated impostor tracks accumulating millions of streams under real artists' names. The feature introduces a unique "artist key" system to authenticate legitimate catalog submissions. The problem has escalated because AI music tools like Suno and Udio now allow convincing fakes to be generated via text prompt, collapsing the effort barrier that previously limited this kind of fraud. This is part of a broader pattern where AI is commoditizing identity theft across music, voice cloning, and synthetic imagery — forcing individuals and platforms into a reactive posture to claw back control.
Court proceedings surfaced private texts between Mark Zuckerberg and Elon Musk that appear to contradict Zuckerberg's public positioning against being the "speech police" on Meta's platforms. The juxtaposition suggests Zuckerberg's rhetoric around content moderation rollbacks may be more strategically motivated than principled. The article, reported by Platformer's Casey Newton, frames this as a credibility gap between Meta's CEO's public statements and private behavior around platform governance.
Landmark 2026 jury verdicts against Meta and YouTube in Los Angeles and New Mexico found platform design features — infinite scroll, autoplay, push notifications, recommendation algorithms — to be actionable defects, circumventing Section 230's broad liability shield. The core legal theory: harms tied to *how* a platform is designed, not *what* users post, fall outside 230's protections. Casey Newton argues this content/design distinction, while imperfect at the margins, is legally meaningful and historically grounded — modern platforms are behaviorally optimizing engagement machines, not the passive content hosts Section 230 was written for in 1996. The practical implication is that restricting mechanical design features (autoplay, streak incentives, late-night notifications) offers a constitutionally viable path to harm reduction without triggering First Amendment concerns over content regulation.
Meta is actively defunding its independent Oversight Board, with full discontinuation of funding potentially occurring after 2028. The move aligns with a broader strategic pivot toward AI-driven trust and safety systems and aggressive cost-cutting to fund AI infrastructure. Negotiations are underway for a compromise, including a scenario where the Board separates from Meta entirely and pivots to serving other tech platforms. The trend of reduced case referrals to the Board suggests Meta has already been operationally distancing itself from the governance structure.
OpenAI is facing a convergence of destabilizing signals ahead of a potential IPO: an unusual acquisition, internal executive reshuffling, and a critical New Yorker investigation. These pressures collectively raise governance and credibility concerns at a pivotal moment for the company's transition from nonprofit to for-profit structure. The timing amplifies scrutiny on Sam Altman's leadership and organizational decision-making.
Anthropic's Claude Mythos Preview represents the first frontier AI model a top lab has deemed too dangerous for general release due to cybersecurity capabilities emerging from general reasoning improvements — not specialized training. Mythos has already discovered thousands of high-severity vulnerabilities across major OS, browser, and kernel codebases, including a 27-year-old OpenBSD flaw and Linux kernel exploits enabling full machine takeover. To leverage these capabilities defensively, Anthropic launched Project Glasswing, a coalition of 40+ tech companies with $100M in usage credits to find and patch vulnerabilities before adversaries can exploit equivalent models. The core tension: this approach centralizes unprecedented offensive power in a single private company, in a regulatory vacuum, while the open-weight model ecosystem is estimated to reach parity within six months.
Nine months after a costly internal overhaul, Meta is asserting it has re-entered competitive standing in the AI race. However, the pace of advancement from rivals means that even a successful turnaround may not translate into sustained leadership. The framing suggests Meta is reacting to the competitive landscape rather than defining it.
Recent bellwether cases in California and New Mexico have resulted in significant verdicts against Meta and YouTube, holding them liable for the harm caused by their platform designs, not just user-generated content. These cases are cracking Section 230 protections and signaling a potential shift towards product liability for social media companies. The legal strategy mirrors that used against the tobacco industry, focusing on the intentional design of addictive features despite known harms.
Recent legal judgments in California and New Mexico have found Meta Platforms Inc. negligent in product design, leading to user harm and the violation of consumer protection laws. These verdicts, which circumvent Section 230 liability shields by focusing on design flaws rather than content, could force significant changes in social media platform design, including features like infinite scroll and algorithmic recommendations. This marks a potential turning point in holding tech companies accountable for the addictive and harmful aspects of their products.
Recent bellwether trials are eroding the absolute immunity previously provided by Section 230 by reframing addictive platform mechanics—such as infinite scroll and algorithmic amplification—as defective product designs rather than protected speech distribution. This shift creates a new litigation front that may compel platforms to alter their core engagement architectures or face massive systemic liability. However, this trajectory creates a significant tension with First Amendment protections, as the boundary between product safety and editorial discretion remains legally ambiguous.