Published on September 02, 2026 at 06:00 CEST (UTC+2)
Claude Fable 5.1 and Claude Mythos 5.1 (1002 points by denysvitali)
Anthropic introduces Claude Fable 5.1 and Claude Mythos 5.1, which are the same underlying model with different safety safeguards. Fable 5.1 is generally available, while Mythos 5.1 is restricted to trusted access programs for cybersecurity and life sciences work. The models feature improved research capabilities and a significant price reduction—estimated 25% lower for typical workloads and up to 45% savings for highly agentic usage, primarily via cheaper cache reads. Anthropic also announces Enterprise Frontier Safeguards (EFS), which allow enterprise customers to maintain complete data privacy by storing data in customer-controlled cloud infrastructure.
How accurate have Ed Zitron's AI skeptic predictions been? (505 points by jatins)
Dan Luu examines the accuracy of prominent AI skeptic Ed Zitron's predictions, comparing them against real-world outcomes. Luu discloses his own neutral position on AI progress, noting that while many futurist predictions have historically been wrong, he has also consistently argued that AI job displacement is underestimated. The article appears to systematically evaluate Zitron's specific claims to see whether his skepticism has been validated by events. It reflects a broader attempt to move beyond polarized AI discourse and demand evidence-based assessment of both boosters and skeptics.
True Rate of Unemployment (16 points by ptrhvns)
The Ludwig Institute for Shared Economic Prosperity (LISEP) publishes its True Rate of Unemployment, which measures the percentage of the U.S. labor force that is functionally unemployed—those without a full-time job wanting one, the jobless, or workers earning below a living wage ($26,000 in 2025 dollars). As of July 2026, the true rate stands at 24.9%, compared to the headline unemployment rate of 4.1%. The rate increased 0.2 percentage points in July, marking the fourth consecutive monthly rise, suggesting a growing gap between official statistics and Americans' actual economic well-being.
The efficient frontier of LLM inference (62 points by philipkiely)
This article explains the concept of an "efficient frontier" in LLM inference, borrowing from economics to describe tradeoffs between latency, throughput, cost, and model quality. Inference engineers can use two types of techniques: those that move a deployment along the existing frontier (e.g., swapping latency for throughput) and those that push the entire frontier outward, creating more efficiency overall. Examples include quantization, distillation, pruning, and adjustable reasoning levels. The post positions inference optimization as a strategic lever for making AI deployments more cost-effective while maintaining desired performance.
Show HN: Weedout – Safari extension that hides YouTube AI-labeled videos (81 points by masteranza)
Weedout is a Safari extension that automatically hides YouTube videos labeled "Made with AI" from home feeds, search results, related videos, playlists, and Shorts shelves. It relies solely on YouTube's own disclosure badge, avoiding heuristics or false accusations, and can optionally auto-skip AI-labeled Shorts. The extension runs entirely on the user's Mac with no accounts, tracking, or data collection, and costs $1.99 as a one-time purchase. It reflects a user-driven demand for curating AI-generated content out of social media feeds.
My local model setup on an M4 Pro Mac Mini (98 points by raybb)
Kevin Lewis details his local LLM setup on an M4 Pro Mac mini with 48GB RAM, running a server that handles everything from agent backends to quick phone queries. His stack includes Qwen3.6-35B-A3B-OptiQ-4bit for reasoning tasks, Gemma-4-E4B-it for lightweight chats, and the oMLX inference server, with Tailscale networking connecting his devices. He cites two main motivations: cloud APIs are "rented land" with unpredictable pricing, limits, and model swaps, and data privacy concerns about what cloud providers do with user data. The setup takes about 30 minutes and demonstrates that capable local AI is becoming practical on consumer hardware.
FBI Probes Service Selling 153M+ Drivers Licenses (69 points by tatersolid)
KrebsOnSecurity reports on a new dark web identity theft service called Nexus, which is selling digital scans of more than 153 million driver's licenses from the U.S. and Canada, plus over 10 million ID cards. Evidence suggests the images were siphoned from a widely-used identity verification company based in Louisiana. The FBI's New Orleans field office has launched an official inquiry into the source, and the stolen data includes licenses of high-ranking government officials like Defense Secretary Pete Hegseth. The incident highlights growing security risks around centralized collection of sensitive identity documents.
Introducing Ad Blocker for Firefox on iOS (350 points by HieronymusBosch)
Mozilla introduces a built-in Ad Blocker for Firefox on iOS, using Apple's WebKit Content Blocker technology and the EasyList filter list. It's off by default and can be enabled in Settings, blocking many third-party ads and ad-related trackers before they load. The feature does not block ads served directly by the site, search result ads, or sponsored content in new tabs. It complements Firefox's existing Enhanced Tracking Protection to reduce clutter and give users more control over their browsing experience.
Sonic Pi (62 points by Bluestein)
Sonic Pi is a free, open-source live coding music synth designed for creating and performing music with code. It supports multi-channel audio, well-timed MIDI and OSC (Open Sound Control), and Ableton Link for networked performances. The tool is aimed at a wide audience: professional musicians and DJs, composers, blind and partially sighted people, and students learning computing through music. It includes a friendly tutorial and is available on Windows, macOS, and Linux, with support available via Patreon and GitHub Sponsors.
The ChatGPT/Codex app bundles a full copy of LibreOffice (287 points by timpera)
Simon Willison discovers that the OpenAI Codex desktop app (now rebranded as ChatGPT) bundles a full copy of LibreOffice, along with complete Python and Node.js installations, and native binaries for Poppler, git, and other tools. These are stored in a ~/.cache/codex-runtimes/ folder taking about 1.7GB, with skills that instruct Codex on how to find and use those binaries. The observation reveals that AI coding agents are increasingly shipping as self-contained runtime environments, pulling in significant open-source software to enable document processing and other tasks without relying on cloud services.
Implication: Expect more "dual-use" model strategies where capability is held constant but alignment and governance are customized per customer segment; enterprises should assess which tier fits their risk profile and compliance needs.
The AI discourse is maturing from hype versus panic to empirical evaluation of predictions.
Implication: Practitioners and investors should develop internal tracking of AI forecasts—from industry analysts, researchers, and internal teams—to inform strategy rather than relying on charismatic pundits.
Official economic metrics are diverging from the lived reality of AI-driven labor disruption.
Implication: AI/ML leaders should monitor broader labor-wellness indicators (e.g., wage stagnation, involuntary part-time work) to anticipate societal pressure for regulation, tax policy, and corporate responsibility around automation.
Inference efficiency is becoming a core competitive battlefield, with "efficient frontier" thinking as the new framework.
Implication: AI teams should benchmark deployments on the latency/throughput/cost frontier and actively adopt push-the-frontier techniques (e.g., speculative decoding, attention sparsity) to stay competitive; the winners will be those who treat inference as a strategic engineering discipline, not a fixed cost.
Consumers are starting to demand tools to filter out AI-generated content.
Implication: For creators and platforms, this implies that AI content must be high-quality and clearly additive, not just volume filler; product teams should plan for native AI-labeling and user controls to avoid backlash.
Local LLM inference on consumer-grade hardware is crossing a threshold of practicality.
Implication: The ecosystem of local inference servers (oMLX, Ollama, llama.cpp) and efficient quantized models is becoming a credible alternative to cloud APIs. Businesses should evaluate hybrid deployments where sensitive or high-volume tasks run locally while occasional heavy tasks use cloud.
AI agents and coding tools are becoming self-contained, bundling entire open-source software stacks.
Analysis generated by deepseek-reasoner