News Feed

The post discusses tagging a large blog by using a “hallucinate” approach: have an LLM generate novel tag concepts without revealing the full vocabulary, then map those concepts to the closest real tags in the existing corpus using vector embeddings. It cites Doug Turnbull’s method and provides a sample prompt structure to guide the classification, such as for the query “brown coffee table.” The goal is to reconcile free-form tag ideas with a fixed, extensive taxonomy through a hybrid LLM-plus-embedding workflow.

The article clarifies that the visible back circles on iPhones are sapphire-covered lens shields, not the actual camera lenses, and they’re highly scratch-resistant. It argues that even minor scratches rarely affect image quality since light is focused on the sensor inside the camera, supported by demonstrations from Zack Nelson and DPReview showing scratches are hard to create and typically inconsequential. It concludes that you don’t need to worry about scratching the lenses.

Charts of the Week traces 'neoclouds'—firms repurposing legacy physical networks to meet AI compute demand—showing rapid revenue growth but heavy capital costs, depreciation, and debt that threaten long‑term profitability relative to hyperscalers. It also discusses price and usage dynamics through token spending efficiency (e.g., Databricks' Smart Router, Ramp data) and surveys implications for horizontal SaaS, cyber security demand, and AI-lab talent flows, illustrating the growth-versus-capital-intensity tension in the AI era.

The article argues that in AI coding, the fastest-iterating teams win, tracing Cursor’s journey from a product-first code editor to developing its own coding models and then agreeing to a $60 billion acquisition by SpaceX via SpaceXAI to gain massive compute. It emphasizes Cursor’s edge—a developer-centric product, multi-model flexibility, and rapid Composer model iterations built on high-quality coding data—allowing it to stay competitive as frontier labs scale. The piece contends that the combination of iteration speed, data advantage, and now huge compute from SpaceX could redefine the competitive landscape for AI coding tools, even as the frontier model race intensifies.

Weekly Dose of Optimism #206 is a curated Friday briefing of notable tech, defense, and science news. It spotlights Avidrone’s DARPA Heavy Lift win, Recast’s weather-startup emergence, and Fuse Energy’s FAETON-X breakthrough in neutron yield, along with other updates like Matic’s robot voice control and Thrive’s capital moves. The piece weaves in analysis about the implications for emerging technologies and markets, including why higher current efficiency could speed fusion to the grid and why weather tech and 'scarce assets' are gaining credibility.

sqlite-utils 4.2.1 fixes a crash caused by an undeclared typing_extensions dependency, which surfaced when users install sqlite-utils directly without its dev dependencies. The update describes a smoke-test method to verify the CLI works without those dependencies, using uv run --isolated --no-default-groups sqlite-utils --help to test in a checkout. This illustrates a practical approach to dependency management and testing in isolated environments.

sqlite-utils 4.2 introduces major enhancements to table.transform(), enabling complex alter-table workflows by rebuilding the table, copying data, and replacing the old one. It now preserves a broader range of schema edge cases (including check constraints, unique constraints, and column comments) and adds new introspection properties for constraints, plus several smaller improvements. The release acknowledges contributors and notes that a crash was fixed in 4.2.1.

The article examines the social and philosophical implications of AI-generated writing, arguing that labeling text as AI-generated functions as a form of authorship that shapes how we judge meaning. Drawing on Barthes and Foucault, it analyzes how AI prose signals authorship and reveals the token-prediction nature of language, while noting the due diligence around watermarking and detection. The piece concludes with a personal note that it is largely human-written with about 10% authorship, illustrating the evolving negotiation of authorship in the AI era.

Traditional AI benchmarks rely on academic tests that don’t always reflect real-world work, and frontier models are beating these tests while still faltering on long, domain-specific tasks. Vals addresses this by evaluating models on actual work products—like legal research, complex financial document analysis, and software development—using domain experts, private test sets, and automated grading to deliver fast, scalable, and adaptable evaluations. By serving as an independent scorekeeper that continuously updates tasks as the frontier shifts, Vals aims to give buyers a trustworthy way to assess AI capability amid rapid advancement.

Google Design highlights its approach in the Achieve goal-tracking app by introducing a custom completion animation built with Styles. Instead of the standard Material ripple, a brand-colored ring expands and fades on tap, paired with a subtle pressed-state gradient to celebrate micro-wins. The team emphasizes polishing these custom interactions while Material handles the underlying foundation, illustrating a design strategy that prioritizes tailored UX and a clear separation of concerns between Material Design and custom styling.

Google introduces Credentio, an open-source C++ library for working with C2PA Content Credentials (versions 2.2 and 2.4), designed to run entirely locally to validate media provenance without sending files to cloud services. It emphasizes a high-performance engine with a small memory footprint, enabling instant, privacy-preserving validation of large assets, and supports configurable trust lists, detailed parsing of credentials, and actionable verdict reports. The project also aims to eventually generate and embed Content Credentials, with contributions welcome at mediaprovenance.googlesource.com.

JetBrains outlines its rigorous evaluation pipeline for frontier AI, using private repositories and leaderboards to decide when Claude Fable 5 is appropriate. In tests, Claude Fable 5 shows higher accuracy and efficiency than Opus 4.8—achieving a 44.3% Python pass rate (vs. 28.2%), solving more tasks with about 22% fewer steps and better reliability. The company pairs this with safety and data-retention considerations and a future roadmap toward a collaborative, agent-assisted software development cockpit.

Google Pixel 11 adds Camera Looks—a sensor‑level, style‑driven processing suite that lets users apply presets like Digi, Classic, and Velvet and tweak controls such as contrast, warmth, tint, and saturation during capture. Instead of pushing all photos toward a single 'perfect' look after the fact, Looks modifies processing early in HDR Plus and the camera pipeline, enabling minimal processing or more pronounced stylization with fewer irreversible artifacts. This shift toward style‑first computational photography marks a move away from a single optimal image and toward user‑selected aesthetics, with the Original look remaining the default and Digi expected to be popular.

Cisco Systems, Inc. earnings transcript from August 12, 2026.

The article describes building alchemy-utils 0.1a0, a database-agnostic version of sqlite-utils powered by SQLAlchemy to work across multiple engines. It covers rapid prototype work guided by AI tools, a TDD workflow, and testing against PostgreSQL, SQLite, and DuckDB, plus practical examples (a one-liner CLI for listing rows and a CSV-to-DuckDB insert) and performance optimizations. The narrative includes concrete commands, repo setup, and notes on reducing a slow initial import from about an hour to roughly 35 seconds.

An in-depth field-guide style evaluation of Grok 4.6, praising its speed and information-dense, collaborative communication that enables a more synchronous, end-to-end workflow across coding and knowledge-work tasks. It presents experiments—from browser automation and spreadsheet apps to 3D and video workflows—highlighting how short prompts with clear completion criteria and a robust verification loop yield reliable results, while noting limitations in non-UI domains. The piece concludes that 4.6 is a dependable, multi-domain assistant worth trying and guiding readers to provide feedback.

The article analyzes why B2B payments lag: despite tidy net terms, cash arrival is governed by a small set of large overdue invoices rather than random delinquencies, with $7.2 trillion in US trade receivables and each extra day of delay costing roughly $150 billion. It reveals that the overdue dollars are highly concentrated—the biggest tenth of past-due invoices hold the majority—so effective collections target a few invoices with automated outreach (81.7% of emails) while humans handle disputes and high-value relationships, typically resolving with a median of about two asks. The piece argues this shifts the focus from chasing volume to strategic, relationship-aware targeting and automation.

The article analyzes why getting paid in B2B hinges on targeting a small set of large, overdue receivables rather than chasing volume, using data from Stuut’s AI-powered collections platform. It highlights that the top tenth of past-due invoices holds about 65.7% of overdue dollars, DSO can mask risk, and bigger invoices tend to stall more, creating concentration risk. It also shows that automated outreach dominates collections while humans handle disputes and key relationships, making the 'ask' function a lightweight, scalable process.

Base Power Company just closed a $1 billion Series D at a $13 billion post-money valuation, fueling its ambition to become a top global energy player by solving the grid bottleneck with near-home batteries (Base Core) and a Gen 3, vertically integrated battery program. The piece outlines its 'Capability Cannon'—a growing mix of deployments, manufacturing, software, and policy know-how—and a 'Make, Move, Store, Sell' framework to scale from Texas to Illinois and beyond, with potential to extend into AI infrastructure. It argues that owning the grid’s high-ground will create a defensible position against entrants and incumbents and could enable Base to capture the largest share of the energy ecosystem over time.

WhatsApp is piloting an optional Scam Alert feature that runs a small on-device ML model to warn users about potential scam messages while keeping end-to-end encryption intact. The system is designed with on-device processing, user control, and privacy-preserving telemetry; only aggregated, differentially private counts are sent to WhatsApp via confidential federated analytics, and content never leaves the device. The post also outlines safeguards and transparency measures—verifiable model behavior, public model weights, OHTTP anonymity, and a plan to publish code and expand bug bounty—to enable independent verification and security testing.

Neel Somani's Power 2026 primer argues that the real bottleneck for AI expansion is electricity price and reliability, since data centers' power demand grows rapidly and is shaped by grid constraints. It explains that power prices are determined by marginal cost through a system-balanced market overseen by ISOs, with the marginal unit typically a natural gas generator and prices set by the cheapest additional unit needed to meet demand (merit order). The article also outlines how data-center developers and investors hedge energy costs and how to read power-market data such as the EIA-860.

The article tells the story of a data-entry job where the author creates a streamlined workflow—validation on data entry, standardized templates, and regular reminders—that yields a dramatic productivity gain (over 10x). Yet management resets expectations and the author is let go after training a replacement, illustrating the central point that productivity gains often become the new baseline and aren’t sustainable without broader organizational changes, a pattern echoed in AI-driven tool use.

The article introduces knowledge profiling and WikiProfile to diagnose LLM factuality in terms of encoding (stored knowledge) versus recall (retrieval), evaluating 13 frontier models on 2,150 facts. It finds encoding is near saturation while recall remains a bottleneck: 95–98% of facts are encoded, yet 26–34% cannot be directly recalled, and 11–12% remain unretrieved even with thinking; thinking substantially improves recall for encoded facts, especially for rare facts and reverse questions. The takeaway is that future gains may come more from better recall utilization than from further scaling.

This Economic Research review analyzes whether worker retraining programs can mitigate AI-driven labor market disruption, drawing on 56 randomized US studies and European experiments. It finds that retraining yields positive but modest gains on average (employment up 2-3 percentage points, earnings about $1,000/year, with ~ $13,000 cost), though sector-focused programs tied to high-demand jobs can perform much better though they are hard to replicate; the authors recommend immediate investment in demonstrating, evaluating, and scaling the most promising programs, including expanding a leading program and measuring results via a dedicated fund.