News Feed

In his acceptance remarks, Powell honors JFK's legacy and emphasizes the Fed's independence and credibility as essential to delivering a stable, prosperous economy. He outlines the Fed's responsibilities—the monetary policy framework, regulation, payments system, and crisis tools—and argues that decisions must be evidence-based and shielded from politics to preserve public trust. He also notes the Fed’s role during the Global Financial Crisis and COVID-19, commends its staff, and urges commitment to the rule of law and democratic institutions.

The author argues that AI lowers the barriers to shipping small, quirky projects and shares a year’s worth of personal AI-assisted builds (Skifreedle, Autodeck, Endless Wiki, VicFlora Offline) to illustrate what’s possible beyond writing code. The piece highlights concrete implementation choices and early usage signals (e.g., 500+ subscribers for Autodeck, 280k pages generated by Endless Wiki), showing that AI can enable useful tools even if it doesn’t spur a flood of big startups. It also discusses deployment challenges and situates these efforts within the broader landscape of AI-enabled tooling.

The piece traces how people become “pilled” by polished ideas that spread online, arguing that transmissibility and the need to belong often trump truth. It offers practical antidotes—state your position plainly, keep a dissenter you respect, read the strongest opposing argument, avoid sloganized thinking, track your predictions, and be willing to break your framework—to stay unpilled. It notes that the urge for certainty is adaptive and platform-driven, and that the only conviction worth keeping is one you could lose tomorrow and still endure.

The article argues there are two patterns for using LLMs in software: pipelines, where the control flow is defined in code, and agents, where the LLMs manage control via tools. It weighs the trade-offs: pipelines are predictable and cost-efficient but limited in context and adaptability, while agents are more flexible and capable for complex tasks but bring unpredictable runtime and costs. It then offers guidelines—use pipelines when context is small, GPU cost must be predictable, or local models are involved; use agents when you can't assemble all relevant context upfront or the problem is hard—plus notes on safety and future-proofing.

Anthropic explains how it contains Claude across its products through layered sandboxing—process sandboxes, VMs, filesystem boundaries, and egress controls—to harden agent boundaries and prevent data exfiltration. Claude.ai uses gVisor, Claude Code runs locally with Seatbelt on macOS and Bubblewrap on Linux, and Claude Cowork uses a full VM (Apple Virtualization on macOS, HCS on Windows). The post also highlights past exfiltration risks (like api.anthropic.com/v1/files) and suggests revisiting Anthropic Sandbox Runtime for deeper testing and documentation.

The article details a project that runs Python ASGI web apps inside the browser using Pyodide and a service worker to intercept /app/ requests and execute them via ASGI, removing the need for a backend server beyond static files. It demonstrates the approach with a FastAPI demo and a full Datasette app, and notes progress on Datasette Lite, a browser-only Datasette powered by WebAssembly. The post contrasts this with an earlier Web Worker approach that failed to execute in-page JavaScript and outlines ongoing work to deepen understanding and upgrade the implementation.

The article coins and defines 'dickover' as a full-screen, modal overlay that obscures content to force actions like cookie consent or newsletter signups, arguing these interruptions are ubiquitous and degrade reading. It distinguishes dickovers from milder 'dickbars' and sometimes-necessary paywalls, and offers numerous examples while detailing the author’s naming process (including a Mastodon poll that favored 'dickover').

Retail participation in U.S. markets has surged to record levels, with record cash and options volume, higher leverage in retail ETFs, and margin balances at or near all-time highs, signaling a one-way buying regime with rising flow fragility. The article also connects AI-capex funding to rising debt among hyperscalers, FedRAMP 20x accelerating government software approvals, and the grocery sector leveraging first-party data to monetize through advertising, as illustrated by Walmart and Costco. Overall, the piece blends data-driven observations with forward-looking implications for market stability and capital markets.

Bowman outlines a practical, flexible monetary-policy decision-making framework for the Fed, detailing how she assesses GDP, labor market conditions, and inflation (notably core PCE and trimmed-mean measures) and uses private-sector input to avoid backward-looking biases. She argues for a balanced, flexible approach to the Fed's dual mandate—adjusting focus when employment and inflation diverge, given uncertainties around the neutral rate and other unobservable variables—and provides examples of how this framework informs past votes on the federal funds rate and expectations for future policy actions.

Weekly Dose of Optimism #195 surveys several notable advances across medicine, nanotech, and space—highlighting Eli Lilly's VERVE-102 gene therapy that dramatically lowers PCSK9 and LDL-C in a phase 1 trial, emerging evidence that GLP-1 drugs may slow cancer, Hermeus's unmanned supersonic flight, and a milestone in atomically precise mechanosynthesis, plus NASA Moon Base updates.

Anthropic released Claude Opus 4.8, describing it as a modest but tangible improvement with a stronger emphasis on honesty and avoiding unsupported claims. The update adds mid-conversation system messages and lowers the prompt-cache minimum, and evaluations claim Opus 4.8 has the lowest incorrect-rate among six models by abstaining when uncertain, with pricing remaining the same as prior Opus generations and a higher-cost ‘fast mode’ for research preview participants.

Google's I/O 2026 keynote presents a bold agentic AI era centered on Gemini for Science, ERA, and Co-Scientist to accelerate scientific discovery and experimental workflows. It showcases substantive, data-backed advances across science, health care, edge computing, and Earth AI—supported by published Nature papers, randomized studies, and real-world deployments like the Google Health app and WeatherNext—along with open developer tools to broaden access. The piece frames these efforts as meaningful, research-driven progress rather than mere announcements.

Researchers unveil FROST, a side-channel technique that uses JavaScript to measure SSD contention via the browser's OPFS and then classifies the resulting I/O traces with a CNN to fingerprint which apps or websites a user has open. The approach demonstrates a potential, data-driven vulnerability but requires a very large OPFS on the same SSD and has limited scalability, with the full attack tested on an M2 Mac and not yet on Windows. Defenses include closing unnecessary tabs and restricting OPFS file sizes, and browser makers could mitigate by limiting OPFS storage.

AI in B2B customer support functions as an invisible copilot, primarily triaging inquiries, routing complex tickets to human specialists, and boosting human agents' efficiency rather than replacing them. The article notes that deflection rate is a limited measure, with end-to-end AI resolution at about 15% for B2B (versus ~35% in B2C), and that outcomes improve when AI has more customer context and account intelligence, leading to similar outcomes to human-only handling when AI assists and escalates appropriately. The authors conclude that AI augments, not replaces, support teams—especially in AI-native, context-rich environments—though high-value customers may still see human involvement.

The article argues that the internet’s era of boundless abundance has bred decision fatigue and mistrust, and it proposes the 'Costco theory of the internet': trusted operators prune options, enforce standards, and absorb complexity so users can navigate with less mental load. By favoring bounded trust and a membership-like reliability, platforms across media, software, and marketplaces could deliver a higher floor of quality and reduce the need for constant evaluation.

Ralph Lauren outlined the first year of its Next Great Chapter Drive, reporting fiscal 2026 revenue above $8 billion, gross-margin expansion, and margin resilience despite tariffs, driven by broad growth across channels and regions. The company detailed progress on elevating the lifestyle brand, expanding core and high-potential categories (notably women’s apparel, outerwear, and handbags), and winning in top cities with a growing direct-to-consumer ecosystem—highlighting AI-driven marketing, 108 new stores, and notable China growth (>50%), plus 1.4 million new DTC customers and a rising social footprint. It also underscored strategic partnerships, innovation in technology and data analytics, and ongoing investments to sustain long-term growth and shareholder value.

Asana reported Q4 2026 revenue of $205.6 million, up 9% year over year, with non-GAAP operating income of $18.2 million and solid free cash flow, while continuing to invest in its AI platform. The company outlined its Agentic Enterprise strategy, describing AI Teammates and AI Studio as foundational to a Work Graph–driven system of action, and detailing four differentiators—memory, task-based execution, multiplayer collaboration, and governance—that enable enterprise-scale AI. With AI Studio ARR exceeding $6 million, 200+ AI Teammates beta customers, international growth, and expanding partner-driven deals, Asana signaled multi-product expansion ahead of FY27 guidance and noted a path to broader availability for sales-led customers by end of Q1.

Autodesk reported robust Q4 and full-year fiscal 2026 results with stronger billings and revenue driven by a new transaction model and ongoing cloud/AI investments, along with the completion of its go-to-market optimization. The company emphasized converging design, make, and operate workflows through its Forma for Construction platform and AI capabilities, and provided fiscal 2027 guidance (billings $8.48–$8.58B, revenue $8.1–$8.17B, non-GAAP OPM ~38.5–39%, free cash flow $2.7–$2.8B) while signaling near-term disruption from sales restructuring and continued buybacks. It also highlighted key customer wins and partnerships that illustrate the platform’s expanding addressable market and value proposition.

Geopolitics expert Dr. Pippa Malmgren argues that the US–China–Russia rivalry plays out as a linked ‘Rubik’s Cube’ of issues—including Iran, Taiwan, Cuba, Venezuela, and Ukraine—where concessions in one arena depend on others, pushing toward a cooperative 'Star Trek' outcome rather than a catastrophic 'Star Wars' clash. She treats Iran’s 60% enriched uranium less as an isolated weapons concern and more as leverage tied to proving origins and reshaping regional alignments to negotiate terms, including economic access and sanctions relief. The discussion links regime change to conversion—turning adversaries into economic partners—within a broader strategy to stabilize the region without war.

Vice Chair Jefferson outlines three global developments—higher energy prices from Middle East tensions, rapid AI-driven productivity advances, and ongoing trade-disruption—and assesses their implications for the U.S. economy. He expects solid but slower U.S. growth with a broadly stable labor market, noting that tariffs and energy costs have kept inflation elevated and that disinflation should resume later this year as shocks fade. He reiterates the Fed’s 2% inflation goal and keeps the policy rate at 3.5–3.75%, stressing data dependence and the need to balance risks in pursuing price stability and maximum employment.

The article discusses the Google Tunix Hackathon where the community trained Gemma models to reveal their reasoning using explicit <reasoning> traces, demonstrating that productive reasoning can be trained on modest hardware with Kaggle TPUs. It highlights the top approaches—G-RaR (Rubric-Based Reinforcement Learning), Pinocchio-1B (3-act reasoning pipeline), and IDEA-E (curriculum-guided GRPO)—which combine supervised fine-tuning, reinforcement learning, and judge-based rewards to improve structured Chain-of-Thought across domains like medical, chemistry, and law. The piece emphasizes broad participation (11,000 entrants, 300+ submissions) and provides practical resources for others to reproduce these results.

The article addresses the Kubernetes inference cold-start problem, where elastic scaling can take minutes and leave GPUs idle, risking SLA violations during traffic spikes. It introduces NVIDIA Dynamo Snapshot as a fast-starting approach to accelerate inference pod readiness and improve GPU utilization. This technique aims to mitigate latency and idle-resource waste during fluctuating demand.

A Federal Reserve official discusses AI’s potential to boost productivity and GDP while monitoring inflation and labor-market dynamics, and outlines how AI could improve financial-system efficiency, credit access, and risk monitoring. He warns of risks from AI-driven trading, sector disruption, rising AI-related debt, and cyber threats, and emphasizes balancing responsible innovation with financial resilience. He also highlights the Fed’s experimentation with AI to understand and mitigate these macro and financial impacts.

The essay argues that data centers are not just end users of computing power but ‘Buyers of Capabilities’ who can accelerate hard-tech innovations by providing scale and demand that pull learning curves downward, enabling technologies from GPUs and DRAM to nuclear and advanced energy systems to mature. By applying alpha-product and DoD/NASA-style procurement logic to the commercial data-center market, the piece suggests data centers fund, de-risk, and scale technologies that might otherwise stall, effectively accelerating reindustrialization. It concludes that data centers are a decisive, positive force for future tech development, and opposition to them is misplaced.