Recently saved
See what's getting saved across QEY.
Texas paused new data center permits after its grid queue exploded from 63 GW to 474 GW in six months, mostly AI infrastructure. The freeze reveals real problems: developers spam speculative projects with no customers, the interconnection process can't scale, and the grid can't absorb this load without massive upgrades that cost billions and take years to build.
- Texas's grid operator went from handling 50 large loads at a time to 225 new requests by mid-November, forcing a shift to batch processing that's now stalled while regulators audit which projects are actually real.
- Data centers are trying to skip the grid entirely by building on-site power plants, but nearly every site still plans to connect to ERCOT eventually—they're just using private power as a bridge while waiting years for grid upgrades.
- The interconnection deposit ($50,000 per MW) is cheap enough that developers can walk away from failed projects by forfeiting only 20% of their deposit, leaving ratepayers stuck covering shared regional transmission costs that nobody else committed to paying for.
- ERCOT's reliability standards require new data centers to stay connected through voltage faults (after one dropped 3+ GW in Virginia), and regulators now want to charge large loads based on their peak capacity needs rather than actual usage—shifting costs toward data centers.
I appreciate the detailed instructions, but I need to be honest: I can't access the actual content of YouTube videos. The URL you've provided points to a video page, not written article content that I
David Perell interviews comedian Mark Normand about the craft of writing funny material—what makes jokes work, how to structure comedy, and the thinking behind his approach to humor.
- Normand breaks down the mechanics of joke construction: setup, misdirection, and payoff, with specific examples of how timing and word choice change whether something lands
- He discusses how comedians mine personal experience and observation for material, and the balance between authenticity and exaggeration that makes stories funny
- The difference between writing for stand-up versus other formats (podcasts, TV) and how each medium requires different comedic instincts
HackerRank is rolling out Chakra, an AI agent that conducts coding interviews while candidates have access to AI tools, measuring not just whether they get the right answer but how they think and work with AI. The company claims this approach cuts cheating flags by 70-80% and consolidates a three-round hiring process into one interview.
- Chakra has already conducted 500,000+ interviews in beta with companies like Snowflake and Capgemini, and HackerRank is positioning it as the future of its business, comparing the shift to Apple moving from iPod to iPhone.
- Candidates can use AI during interviews without penalty, which HackerRank found actually reduces cheating because it removes the incentive to secretly use outside tools—suspicious activity dropped 70-80% compared to traditional assessments.
- The AI measures "AI fluency" (how well candidates prompt AI and evaluate its output) and reasoning process rather than just final answers, while HackerRank says humans retain final hiring decisions to avoid over-delegating to algorithms.
- Regulatory scrutiny of AI hiring tools is already happening (NYC requires bias audits), and while Ravisankar argues properly tuned AI is less biased than humans, automated systems can inherit biases from training data and criteria.
A 19-year-old founder's startup Ghost just released Core, a $3,499 desktop device built specifically to run AI agents locally on your behalf. The machine comes with an Nvidia GPU, preloaded models, and encrypted local storage so Ghost and other third parties can't access your data.
- Core costs $3,499, includes an Nvidia RTX Pro 4000 SFF Blackwell GPU, and ships with three AI models preinstalled (Qwen and Gemma); preorders open Monday with delivery in late October.
- All code, model weights, and personal data run locally on the device—not in the cloud—and Ghost says it's designed to keep working even if the company shuts down.
- Javaid argues Core differs from a Mac Mini because it's purpose-built for AI agents with software designed to handle continuous task execution, memory management, and autonomous decision-making without constant user input.
- The device includes a firewall that monitors outgoing network requests and blocks agents from exfiltrating data or performing unauthorized actions.
Unable to fetch article content.
Unable to fetch article content.
Unable to fetch article content.
Eric Topol interviews Michael Lewis about his new book "Blockers," which examines civil servants resisting government dysfunction. Lewis's chapter on the NIH reveals a different account of events than previously reported, framed as an investigative mystery.
- Lewis's latest book focuses on "blockers"—career government employees across multiple agencies working against destructive policies, particularly from the DOGE team in the current administration
- The NIH chapter presents a previously unreported version of events, structured as a narrative mystery rather than straightforward reporting
- Lewis has shifted his recent output toward government criticism: 4 of his last 5 books examine US government failures and dysfunction
Tibo Sottiaux, head of ChatGPT and Codex at OpenAI, predicts that agents will handle most internet actions soon, the model picker will disappear, and people should build assuming models will be 10x better in a year. He also argues that complex agent architectures are a passing phase and that taste—understanding users and knowing what good looks like—will matter more than technical skills like typing speed.
- Agents will soon take most actions on the internet, forcing every product team to decide whether to build for them; Notion's MCP server launch caused system strain and forced economic rethinking.
- Models will improve at "incredible" rates in speed and cost over the next year, so builders should design as if today's capabilities are 10x worse than what's coming, not optimize for the present.
- Complex agent loops and graph architectures are a temporary solution—as models get better, larger single agents can handle what previously required teams, making intricate setups obsolete.
- The OpenAI plugin ecosystem could be a major opportunity: plugins get revenue share when users spend tokens, and ChatGPT recommends them to 1.2 billion users based on retention and quality.
Liquid AI released d1 with image support, a decision model designed for yes/no and classification tasks. It matches or beats GPT-6.1 Sol on most benchmarks while costing 19-200x less and running 200-300ms per decision.
- d1 processes images and text to output probabilities for binary, choice, or scoring questions in a single forward pass without token generation
- Benchmarked against GPT-6.1 Sol and Claude Opus 5.5 on six real applications (support ticket filtering, circuit board inspection, etc.), matching or beating GPT-6.1 Sol on four of them
- Costs 19-200x less than comparable frontier models with significantly faster inference speed
This is a landing page for Hailuo 02, an AI video generation model that lets you create videos from still frames with text descriptions. You can set start and end frames, describe what you want, adjust camera movement and sound, and output videos up to 6 seconds at 768p resolution.
- Frame-to-video generation: input start and end frames plus a text description to generate video content between them
- Customizable output: control camera movement, add sound effects, and render at 768p up to 6 seconds long
- Free tier available: sign up to create videos without payment
Norway is drafting legislation to ban camera-enabled smart glasses in parks, beaches, schools, shops, and other public venues, citing privacy concerns about covert recording. The move reflects growing global pushback against the devices, with restrictions already appearing in UK and US venues, and reported incidents of women being filmed without consent.
- Meta's Ray-Ban smart glasses have sold at least 9 million pairs since early 2023, but incidents in France show they're being used to film women without knowledge, with clips posted on social media as "pranks."
- Norway's minister of digital affairs wants to prevent a society where people worry about being secretly recorded in places they expect privacy, like gyms and healthcare facilities.
- California's governor vetoed legislation that would have criminalized using smart glasses to record in private spaces like changing rooms and doctors' offices, despite the state having existing unlawful recording laws.
- Meta argues its glasses have a tamper-proof LED that blinks when recording and can't be disabled, but this hasn't stopped reported misuse.
Unable to fetch article content.
Bank of Japan official Uchida warned that while AI adoption has boosted demand and economic activity, a gap between inflated asset valuations and actual profits from AI investments could trigger a market pullback. The central bank is monitoring whether AI will push inflation above its 2% target and affect Japan's natural interest rate.
- AI has acted as a major demand shock supporting economic activity, but elevated stock valuations may not match the actual profits generated by AI-related investments, creating correction risk.
- Stock price gains from AI have eased financial conditions, but heavy bond issuance by AI companies has simultaneously pushed up long-term interest rates, creating conflicting pressures.
- The BOJ has identified AI-driven demand as a potential factor pushing inflation above its 2% target, which could require further monetary tightening.
Google froze its open source bug bounty program starting October 1 because AI-generated submissions overwhelmed the system with invalid reports and hallucinations. The company won't reopen it until Q1 2027.
- Google engineers and open source maintainers were flooded with automated submissions that were mostly invalid or contained AI hallucinations, making the program unmanageable.
- The pause confirms cybersecurity experts' earlier warnings that AI-generated content poses a real threat to bug bounty programs by degrading their signal-to-noise ratio.
- Google is redirecting participants to its other bug bounty programs while it figures out how to handle the AI submission problem.
Unable to fetch article content.
Unable to fetch article content.
Living Models paired Gemma 4 with BOTANIC-1, a plant genome foundation model, to identify causal mutations in crop breeding. The system solved a decades-old genetics problem—distinguishing which mutation actually drives a trait from dozens of hitchhiking variants—in hours instead of months of field trials.
- Classical bioinformatics tools hit a wall on the melon test case, narrowing 47,492 variants to 3,061 candidates but couldn't break linkage disequilibrium to identify the true causal CmEIN3 mutation; BOTANIC-1 ranked it #1 out of 2,494 SNPs using evolutionary constraint scoring.
- BOTANIC-1 placed true causal mutations in the top 0.1% of candidates in 15.9% of cases across 500+ validated mutations, versus 4.6% for the best non-GLM baseline, and in the top 1% in 48.8% of cases versus 33.9% for classical pipelines.
- The modular architecture runs Gemma 4 E4B locally on a single NVIDIA L4 GPU via Ollama, keeping proprietary genomes on-premises while slashing costs and making domain-specialist AI accessible on standard hardware.
General Catalyst acquired Ohio nonprofit Summa Health through a shell company (HATCo) structured to guarantee profits for the VC firm regardless of patient outcomes or hospital performance. The author argues the deal's asymmetric payoff math and General Catalyst's pivot away from failed digital health investments expose this as a private equity play dressed up as healthcare transformation.
- The deal structure is mathematically rigged: General Catalyst profits no matter what happens to Summa Health or its patients, a heads-I-win-tails-you-lose setup the author says regulators and analysts failed to stop.
- General Catalyst created HATCo (named after CEO Hemant Taneja's initials) to finance a hospital buyout it couldn't afford through normal VC channels, then claimed it's neither VC nor private equity despite following that exact playbook.
- The acquisition signals General Catalyst is fleeing a dismal decade of digital health failures and poor returns by pivoting to hospital ownership—a desperate move outside its core venture wheelhouse.
Android Headlines obtained exclusive images of the Fitbit Edge, Google's upcoming fitness tracker with a small oval display, GPS, and 7-day battery life. It's positioned as a middle ground between the budget Fitbit Air and the full-featured Pixel Watch 5, launching this month for $149–$199.
- The Fitbit Edge replaces the Fitbit Charge line (dormant since 2023) with an oval display instead of round, offering notifications, GPS, altimeter, and limited apps like Google Maps and YouTube Music
- It supports both Android and iOS, unlike the Pixel Watch 5 which is Android-only—a major advantage for cross-platform users
- Launching in three colors (black, green, orange) matching the Pixel 11 Pro's Obsidian, Olive, and Canyon options
Noah Smith pushes back on Paul Krugman's claim that AI creates no benefits for ordinary Americans, arguing the data center buildout is generating substantial construction jobs, tax revenue for local governments, and demand-side economic support that's offsetting Trump's disruptive policies.
- Data center construction has created measurable job growth for construction workers since late 2022, with real wage increases even as Trump deported undocumented workers from the sector, and Goldman estimates 500,000 additional construction jobs needed by 2030.
- State and local governments are collecting significant tax windfalls from data centers—property taxes, sales taxes, corporate taxes—that fund public goods like schools, roads, and libraries, with Loudoun County, Virginia as a concrete example of this payoff.
- The AI investment boom is functioning as a major positive demand shock to the economy, counterbalancing the negative effects of Trump's tariffs, deportations, and policy uncertainty that would otherwise have caused visible economic damage.
Plain text persists as one of computing's most durable technologies because its simplicity—no proprietary formats, no required applications, no dependencies—makes it readable across decades and systems. The article argues that while richer tools have their place, plain text's boring reliability and portability solve real problems that fancier solutions often create.
- Plain text works everywhere without dependencies: a text file created on Linux opens on Windows, runs in terminals, works with command-line tools, and doesn't require the original application to remain in business or supported.
- Text-based formats (source code, config files, HTML, JSON, Markdown) dominate computing infrastructure because their contents stay inspectable and debuggable without proprietary software, making troubleshooting and version control straightforward.
- Plain text resists obsolescence better than proprietary formats—old text files remain readable even if encoding needs fixes, whereas recovering data from abandoned file formats often requires emulators, conversion chains, or lost software.
- Ownership and control stay with the user: text files can live in any directory, sync however you want, go under version control, or move to different machines without vendor lock-in or subscription dependencies.
Three scientists won the 2023 Nobel Prize in Physiology or Medicine for developing optogenetics, a technique that uses light to activate or deactivate specific neurons in living brains. The breakthrough came from studying light-sensitive proteins in algae and finding a way to apply them to mammalian nerve cells.
- Traditional neuroscience methods like genetic deletion of neurons have major limitations—the brain adapts to cell loss, and early development changes can cascade through neural networks, making it hard to know what those cells actually do.
- Optogenetics solves this by letting researchers turn specific neurons on and off in real time in an intact, functioning brain, providing direct evidence of what those cells control.
- The technique relies on ion channels—proteins that control electrical signals in nerves—which researchers learned to engineer so they respond to light instead of their natural triggers.
Google Docs and Drive now let you open, edit, and collaborate on Markdown files directly without converting them to another format. The files stay as .md files while you get all of Docs' real-time editing and commenting features, plus Drive shows rendered previews.
- No more file conversion needed — editing Markdown in Docs preserves the original file format and prevents formatting loss, stripped comments, and fragmentation that happened during import
- Both humans and AI agents can collaborate on Markdown files without writing raw syntax or losing access to Docs' collaboration tools
- Rolling out gradually starting October 5, 2026 to all Workspace customers and personal Google account users with no admin controls needed
This is Google's standard login screen where users enter their email or phone number to access their Google Account. It's not an article — it's a login interface.
- This is a login page, not readable content suitable for a read-it-later service
- The URL shows a redirect loop attempting to access Google Notebook
- No article text, claims, findings, or substantive information exists on this page
Louisiana officials used nondisclosure agreements to hide details of a $3.6 billion data center project from residents until construction had already started, sparking anger and prompting lawmakers in multiple states to consider banning the practice. The company and officials claim NDAs are necessary for competitive advantage and securities compliance, but transparency advocates argue they give corporations unfair leverage while shutting out the public from decisions affecting their communities.
- At least 54 NDAs have been signed by Louisiana officials on various projects since early 2024, with researchers finding NDAs used in at least 25 of 31 data center projects surveyed in Virginia—suggesting the practice is standard across the country.
- The Delta Forge 1 data center will draw 300 megawatts of electricity (enough for roughly 200,000 homes) and bring 1,000+ construction jobs and 200 permanent jobs, but residents learned details only after fences went up and earthmovers started clearing the 300-acre site.
- Bills to ban NDAs for data center projects have been introduced in Michigan, Oklahoma, Kentucky, Ohio, and Pennsylvania, with Pennsylvania's governor already banning them for state agency projects; Applied Digital's CEO acknowledged the secrecy "feels sneaky to people," but justified it citing securities law and material non-public information rules.
- A nearby $10 billion data center in West Feliciana Parish proceeded without NDAs, with officials posting all agreements and land details publicly online, proving transparency doesn't require secrecy.
NPR is tracking how local fights over data center construction are reshaping American politics and creating new dividing lines between communities. These disputes are becoming major issues in the 2026 midterm elections.
- Data center opposition is emerging as a grassroots political force that cuts across traditional party lines at the local level
- Communities are organizing around concerns about data centers—likely environmental, infrastructure, or land-use impacts—creating new political constituencies
- These battles are influencing policy debates and will be central to midterm campaign messaging
Hotelist is a hotel search engine that bypasses the commission-driven rankings of traditional booking sites by using AI to analyze real traveler reviews, room photos, and verified amenities instead. It's built by Pieter Levels and aims to surface genuinely good hotels rather than ones that paid for placement.
- Hotel booking sites are corrupted by pay-to-play mechanics: hotels pay higher commissions to rank higher, and review platforms are flooded with fake ratings that can be bought for hundreds of dollars
- Hotelist uses vision AI to analyze actual room photos for cleanliness and condition, and scrapes real traveler discussions across the web to extract honest pros and cons independent of hotel marketing
- Amenity claims get verified against photos—so a "gym" must actually have weights, not just one treadmill, and filters can target real coworking desks, fast wifi, and actual kitchens rather than marketing fluff
A woman has filed a civil suit against Cornell University, Chi Phi fraternity, and seven named men over an alleged gang rape that occurred two years ago. The case is now drawing national attention after the campus newspaper published the accused men's names and began reporting on newly surfaced text messages and documents.
- The alleged assault involved a text message sent to fraternity members advertising "free" sex, with the plaintiff claiming seven men participated in rape
- The civil suit names Cornell University, the fraternity, and the accused individuals as defendants, with new evidence emerging regularly
- The campus newspaper's decision to publish the accused men's names has made this a high-profile case, raising questions about media coverage and due process
Hospital leaders are organizing a lobbying effort to delay or reverse nearly $1 trillion in Medicaid funding cuts over the next decade that were included in Trump's tax cut bill. The industry is exploring strategies to mitigate the damage to their operations and funding.
- Congress passed a law cutting nearly $1 trillion from Medicaid over 10 years as part of the Trump tax cut bill
- Hospital leaders are actively building plans to delay, roll back, or reduce the impact of these cuts
- The hospital industry is mobilizing its lobbying resources in Washington to fight the spending reductions
Meta released open-source code so anyone can build their own Muse AI gadgets using ESP32 boards or Raspberry Pis, and is also giving away 5,000 of its own Muse Home Link device to control smart home devices. The company is essentially letting hobbyists turn the AI agent into whatever form factor they want.
- Meta open sourced SDKs for Muse so developers can load it onto off-the-shelf hardware like E Ink displays, HDMI sticks, or touchscreens
- The company manufactured 5,000 Muse Home Link gadgets and opened a waitlist for free units shipping this month
- Muse Home Link can control lights, TVs, printers, and other devices through community-built skills
Human intelligence is a recent evolutionary accident, not an optimized endpoint—and now that we can explicitly define and measure intelligence, AI systems will outpace us wherever we can formalize the task. The real problem isn't staying competitive; it's choosing which objectives are worth optimizing for, since the definition of intelligence shapes what gets built.
- Evolution is a lazy optimizer that never selected humans for maximum intelligence; AI can reach superhuman performance in months once a task is formally specified, as AlphaGo demonstrated.
- Recursive self-improvement becomes possible once the model development process itself is broken into measurable subtasks—meaning even the optimization process will eventually be automated.
- How we measure and evaluate intelligence directly shapes what intelligence emerges; a system rewarded for stock market gains develops differently than one optimized for curing cancer, and poorly defined objectives can produce capable but undesirable behavior.
- The burden shifts to humans defining which hills are worth climbing rather than competing on every one—choosing the right objective matters more than winning the optimization race.
As AI agents like Meta's Muse and OpenAI's dots handle tasks on behalf of users, websites need a way to verify a real person is actually behind each agent. World ID's Proof of Human system lets people prove they're unique individuals without revealing their identity, solving the problem that agent numbers are unlimited but humans are finite.
- Amazon already blocked Muse from shopping because the agent didn't disclose it was AI, showing websites will enforce rules about agent transparency and attribution as these tools scale
- Proof of Human matters for practical limits: rate limiting per person, counting unique customers, and preventing one person from gaming purchase allocations by running thousands of agents
- The system uses zero-knowledge proofs so agents can prove human backing without sharing personal data, and companies like Okta, Vercel, and Browserbase are already building it into their platforms
AI inference spending will overtake databases as the largest software market by 2027, fundamentally breaking the economics that made traditional SaaS profitable. This shift forces application companies to manage massive infrastructure costs that will compress margins and demand new operational approaches.
- Inference market grows from $25b (2025) to $350b (2027), passing databases at roughly 2x the size, making it the dominant software category.
- Gross margins collapse below SaaS norms (72%) because inference costs become a major COGS line item, forcing vendors to either build proprietary models to reduce token overhead or accept margin compression.
- Customer-owned infrastructure (BYOK) trades top-line revenue for gross margin—vendors get higher percentage margins but on drastically smaller contracts.
- Token price wars won't solve the problem; capability improvements keep resetting frontier model prices upward, so companies must optimize through smaller models, better routing, and efficient harnesses.
a16z's seventh ranking of consumer AI apps shows ChatGPT maintaining dominance while Claude surged to #3, but the real story is that spending follows a power law — the top 1% of payers account for nearly 20% of all consumer AI spending, and most paying users are prosumers buying coding and creative tools rather than mainstream consumers.
- Only 4.5% of U.S. consumers have paid subscriptions to ChatGPT, Gemini, or Claude as of August, and just 25% use AI daily, revealing extremely narrow monetization despite broad awareness.
- The top 1% of AI spenders average $903/month and account for 19.5% of total consumer AI spending, while median payers spend $25/month — a stark power law distribution that's driving revenue concentration in productivity and creative tools.
- Claude jumped from not ranking in September 2023 to #3 by traffic and tied with Gemini in paid subscribers, but showed signs of slowdown by mid-summer as ChatGPT re-accelerated with new features and work-focused products.
- Personal AI agents (Instinct, Tomo, Lindy) are emerging as a new monetization model, with early adopters spending $1,300/month on average and transaction-based pricing potentially enabling deeper consumer adoption than subscription models.
This is a GitHub project that teaches Claude (or other LLMs) to write in Simplified Technical English (STE), a controlled language used by aerospace and defense industries for clear, translatable technical documentation. It includes writing rules, an approved word list, and a Python checker tool to validate compliance.
- The skill enforces 53 writing rules and a dictionary of 869 approved words based on the official ASD-STE100 specification, making technical text shorter and easier to translate
- It includes a Python validation tool that catches common violations like overly long sentences, unapproved verbs, passive voice, and words outside the approved list
- The tool has real limitations—it can't verify if words are used with their approved meanings, and a human expert still needs to review critical documents for full specification compliance
AgentCraft is a tool that runs Claude agents inside a Minecraft environment where you can watch them work on your code in real time. You give them a goal, they plan tasks, write code in isolated git branches, and ask you for decisions when needed — nothing merges without your approval.
- Agents work in their own git worktrees on isolated branches; your main checkout is never touched, and merges only happen when you explicitly approve them
- The interface visualizes agent work through a pixelated Minecraft studio where agents walk to desks, show live code changes on monitors, and come find you with questions
- A real test run on a sample repo landed six features with passing tests, handling merge conflicts, permission prompts, and a Foreman restart — all driven through the game interface
- Small goals cost $2-10 in Claude API usage; you can try it free with a simulated team first
Redis published benchmark results showing its memory system outperforms other production systems on LongMemEval, a standard test for long-term agent memory. The report also flags that published memory benchmarks often don't replicate when re-tested independently.
- Redis' Remis + Instruct strategy achieved 86.5% task-averaged accuracy on LongMemEval, the highest among production systems tested on GPT-4o, at lower cost than competitors
- Oracle's managed service scored slightly higher (within 0.5 points) but uses a stronger answer model
- Vertex Memory Bank's actual performance (42.4%) significantly underperformed its published results (~70%), suggesting benchmark results don't always hold up in real-world setups
Image-Blaster is a CLI tool that generates complete 3D environments from a single image in under 5 minutes, creating textured models, Gaussian splats, and sound effects by chaining Claude, World Labs, and FAL APIs together. You feed it an image, Claude walks you through the process step-by-step, and you get game-ready assets.
- Outputs multiple asset formats (.glb, .obj, .spz files) plus ambient and physics-based sound effects in a single pipeline, designed to drop directly into game engines (Unity, Unreal, Godot) or 3D software (Blender, Maya).
- Hunyuan 3D model supports configurable parameters including face count (40k-1.5M), PBR material generation, and polygon reduction, letting you optimize for performance or fidelity.
- Runs locally via Claude CLI with API keys for World Labs and FAL; the entire workflow is interactive, with Claude confirming each step rather than running fully automated.
Microsoft's recent Copilot announcements represent a shift from AI as answer-box to AI as persistent worker—the real competitive advantage lies not in the model itself but in the "harness" (interface, memory, tools, permissions, execution environment) that lets businesses actually deploy and manage AI agents as ongoing coworkers. The author argues Microsoft is better positioned than labs focused purely on model capability to win this infrastructure battle.
- The competition has moved from "can AI answer questions?" to "can we delegate real work to an AI agent and have it keep working across days?" This requires persistent context, cloud execution, and integration with business systems—not just better models.
- Microsoft's September 2025 announcements bundle Copilot Home, Code, and Autopilot with model choice flexibility, meaning enterprises aren't locked into whichever AI lab is currently ahead. This flexibility is the actual moat.
- The article maps five generations of AI interfaces (answer box → custom knowledge → tool-using assistant → coworker → persistent agent), and argues we're now at the point where the system architecture around the model matters more than incremental model improvements for real work deployment.
Neko Health, a preventative medicine startup co-founded by Spotify's David Ek, just opened in New York after raising nearly $1 billion. The company offers hour-long full-body scans for $500 that analyze blood, skin, and metabolism using AI and computer vision, with a 300,000-person waitlist already in place.
- Neko Health charges $500 per scan and qualifies as an HSA-eligible expense, targeting the U.S. preventative care gap where cost concerns keep adoption low.
- The scan has expanded significantly—from 15 biomarkers to 55—and uses AI-powered computer vision to provide what investors call "actionable insights" rather than just wellness theater.
- The company faces scaling and regulatory hurdles in the U.S. but has secured venture backing and plans nationwide expansion.
Meta's Muse succeeds where ChatGPT and Claude stumble not because of novel technology, but because it nails three things competitors missed: a simple mental model (your personal helper with its own computer), an ad-funded business that lets users ignore token limits, and personality-driven design that builds emotional connection. This positions Meta to own consumer AI at scale.
- Muse abstracts away unnecessary complexity that plagues ChatGPT and Claude—no model selector, no work vs. chat toggle, no obscure slash commands. One unified chat orchestrates everything in the background, giving users a clear metaphor they already understand.
- Meta's ad business lets them fund Muse's free tier generously, while ChatGPT and Claude are constrained by token-based pricing that keeps most consumers from accessing powerful agentic features. Usage is "a mile wide but an inch deep" for the paid competitors.
- Personality and anthropomorphization drive adoption and engagement more than raw capability—Muse bet on cute avatars and likable AI personality while OpenAI and Anthropic went sterile and elite. This resonates with average consumers who've rejected cold-tool positioning.
- OpenAI fumbled by not investing in ads early enough; Google has compute and ad money but no consumer product chops; Apple has no AI expertise. Meta now owns the opportunity to capture the consumer AI market at scale.
Unable to fetch article content.
A former OpenAI safety researcher quit because the company's internal culture prioritizes speed and profit over genuine safety measures, and argues that the entire AI industry's approach to safety is fundamentally broken and will lead to more failures.
- OpenAI's culture systematically deprioritizes safety work in favor of rapid product development and commercial success
- The author claims the industry lacks meaningful accountability mechanisms and treats safety concerns as obstacles rather than legitimate priorities
- Without structural changes to how companies approach AI safety, more serious failures are inevitable
Nolan Lawson explores why developers often build custom solutions instead of using built-in browser and platform capabilities, tracing the reasons from historical browser gaps to the genuine appeal of building things yourself. He argues that understanding this skepticism matters more than simply repeating "use the platform" as advice.
- Developers prefer custom solutions partly from habit and familiarity with npm ecosystems, but also because building things yourself is genuinely more engaging and educational than reaching for a ready-made API.
- Poor documentation and scattered resources historically made it easier to find and use third-party libraries than to discover what the platform could already do, though this has improved with MDN and web.dev.
- Most "avoid the platform" behavior stems from incomplete knowledge of how the underlying system works—developers build custom solutions not out of laziness but because they don't fully understand what the platform offers, a pattern that applies across all development domains, not just the web.
e2e is a TypeScript testing framework that uses AI agents to test web and mobile apps while caching and replaying actions to avoid redundant model calls. It combines deterministic code-based assertions with AI-driven testing goals, letting you pick your own model, gateway, and execution environment.
- Replays cached agent actions without re-running the model, cutting token usage and costs on repeated test runs
- Supports both web and mobile testing through a single API, with migration paths from Playwright, Cypress, Selenium, Detox, and Maestro
- Lets you bring your own AI model (ChatGPT, Copilot, local server, etc.) and customize how the agent approaches testing, rather than forcing a specific model or testing philosophy
Researchers critiqued Harvard studies claiming nuclear power plants cause cancer deaths by replicating the methodology and applying it to random landmarks—Costcos, baseball stadiums, and colleges all "caused" far more cancer deaths than nuclear plants, proving the statistical approach captures noise rather than actual causation.
- The Harvard studies used "proximity scores" based on distance from nuclear plants (up to 200 km) without measuring actual radiation exposure; nuclear plant workers receive virtually no measurable annual radiation dose, making it physically implausible that distant residents would be harmed.
- When the authors applied the same methodology to other landmarks, Costco alone was associated with 2.2 million cancer deaths (20% of all cancer deaths 2000-2018), private colleges with 1.59 million, and various sports venues with hundreds of thousands each—demonstrating the method produces false positives for anything.
- The studies are ecological associations that don't prove causation, yet the authors built a public health narrative around nuclear plants specifically without testing whether their methodology would generate similar results for unrelated landmarks, a basic quality control step they skipped.
TypeSafe AI's decision model Jev autonomously completed Pokémon Red by reading game memory, evaluating legal moves with game facts, and making every decision—from battle tactics to menu choices—over 37 hours and 16,150 decisions. The entire run cost $1.65 in API tokens, with full source code and setup instructions open on GitHub.
- Jev made every game decision (battles, catching, team swaps, navigation) based on legal options and contextual facts, while the harness only handled emulation and button presses—no cheats or memory writes.
- The project cost $1.65 total at $0.042 per million input tokens, with typical calls around 1,200 tokens and median decision time of 0.4 seconds; the full run consumed ~39.2M input tokens.
- All code is open-source (GPL-2.0) and reproducible locally with your own legally obtained Pokémon Red ROM, Node 20+, and macOS or Linux—no Nintendo assets or ROMs are distributed.
DuckDB 2.0 Alpha adds dynamic attach statements that let you pull credentials from environment variables instead of hardcoding them, solving a major friction point when working across multiple data lakes. Early production testing shows less than 1% failure rates, and the release bundles this with Quack and official DuckLake support to make local-first lakehouse workflows actually practical.
- Dynamic attach statements now accept environment variables for credentials (usernames, passwords, account numbers), eliminating the need to hardcode sensitive data or switch to Python for multi-catalog federation
- Production pilot running hourly jobs showed <1% failure rate over 10 days, reading Parquet from S3 and writing to Iceberg tables via RDS-DMS
- Release combines dynamic attach with Quack and official DuckLake support, completing the local-first lakehouse story but raising open questions about concurrent writer behavior on shared Iceberg tables
A production guardrail pipeline for agentic AI cut latency from 13.8 seconds to 1.8 seconds by parallelizing independent validation checks and short-circuiting on cheap, high-confidence rejections—without touching detection logic or accuracy. The lesson: pipeline topology often matters more than model choice.
- Baseline sequential validation pipeline took 13,874 ms per action across five layers (policy, prompt injection, destructive ops, anomaly detection, authorization), making safety features slow enough teams might disable them in production.
- Most validation layers had no real dependencies on each other's outputs but ran sequentially anyway because of code structure, not logic requirements; reordering by cost meant cheap checks could reject actions before expensive ones ran.
- Restructured pipeline achieved 7.6x latency reduction (to 1,828 ms) and maintained 94% accuracy and 0% false-negative rate on destructive operations by parallelizing independent checks and short-circuiting on early rejections.
- The pattern generalizes: RAG pipelines, multi-agent handoffs, and evaluation harnesses often chain steps sequentially out of habit rather than necessity; asking "which steps actually depend on each other?" can yield bigger wins than swapping models.
Microsoft and Google are supporting Apache Ossie, an open specification that lets enterprises move semantic models between analytics platforms without rebuilding them. The backing from major vendors makes portability more likely, but governance gaps and the spec's early stage mean enterprises still face real constraints in switching platforms.
- Microsoft is building two-way converters between Power BI and Ossie, and pushing for DAX language support, so calculation logic moves with models across platforms instead of needing recreation.
- Governance features like row-level security and access policies aren't included in the current spec, forcing enterprises to reconfigure these on each platform even with portable models.
- Ossie uses a hub-and-spoke model where the spec acts as a common format and individual vendors provide converters, reducing engineering overhead compared to point-to-point translations.
- The specification is still in development (version 0.2), lacks capabilities large enterprises need like cross-model references, and won't eliminate vendor lock-in—just shift it to platform-specific execution behavior.
LinkedIn built Northguard and Xinfra to replace Kafka's core design after hitting limits at 32 trillion records per day. The new system decouples logs from replication units, distributes metadata across independent state machines, and lets new brokers serve traffic without copying historical data.
- Kafka's partition couples ordering, replication, and consumer parallelism into one heavyweight unit; moving or repairing partitions at LinkedIn's scale wastes network and disk bandwidth. Northguard uses smaller segments with independent replica assignments, so adding capacity doesn't require copying old data.
- The centralized metadata controller became a bottleneck managing millions of partition replicas. Northguard shards metadata across 128+ coordinators using a hash ring, with each vnode managing its own Raft state machine for topic, range, and segment metadata.
- Northguard can enforce fsync on all replicas before acknowledging writes while meeting LinkedIn's performance SLOs—a durability guarantee that Kafka's default configuration doesn't require because it relies on replication instead.
The author argues that paleoconservative rhetoric—viewing politics as existential struggle over borders, demographics, and elite betrayal—has become the dominant language of the right, even as its actual policy program hasn't fully won out. This shift stems from "enrabblement": a growing conviction that major institutions are hostile foreign powers rather than imperfect shared foundations, which has fractured the conservative coalition and made meritocracy itself suspect when its winners aren't "your kind of people."
- The right's estrangement from universities, media, and professional institutions correlates with three decades of educational sorting—Democrats increasingly college-educated, Republicans increasingly non-college—but this isn't about intelligence; it's about one's concrete relationship to civil society institutions.
- Hegel's distinction between poor and rich rabble explains the current coalition: the poor rabble feels excluded from civil society's rewards, the rich rabble resents its obligations and restraints (Trump exemplifies this), yet both demand recognition and security without reciprocal commitment to the institutions that make those goods possible.
- Enrabblement isn't irrational or external—it's the competitive logic of civil society stripped of its civility, an endogenous product of the system itself, which raises the hard question of whether institutions can be repaired without making the underlying social competition itself less barbarous.
The article proposes FloatGrid, a payment orchestration platform that consolidates fragmented buy-now-pay-later debts across multiple apps into one automated payment schedule. It's designed to solve the problem of consumers juggling dozens of overlapping installment plans with conflicting due dates that trigger overdraft fees.
- Consumers now split purchases across multiple BNPL providers (Klarna, Affirm, Afterpay) plus credit cards, creating extreme payment fragmentation that existing personal finance tools don't address—they show spending history but don't actively coordinate payment timing across lenders.
- FloatGrid makes money three ways: a $4.99/month subscription for automated payment execution, affiliate commissions from consolidation loans, and selling anonymized BNPL default data to credit analysts.
- The platform's defensibility comes from execution lock-in—once users automate their payoff scheduling through FloatGrid, switching away creates operational friction and late-fee risk that keeps them trapped.
William Davies examines what separates genuine intellectuals from experts, academics, and thought leaders, tracing how Hayek's anti-intellectualism and algorithmic platforms have reshaped where politicians find ideas and whether independent thought can survive in the modern marketplace.
- The distinction between intellectuals and the various competitors for attention (technocrats, academics, polemicists, thought leaders, gurus) has blurred, raising questions about whether the intellectual role still exists in public discourse.
- Weber and Gramsci's ideas about intellectual autonomy and responsibility provide historical frameworks for understanding what intellectuals should do, but Hayek's anti-intellectual philosophy has fundamentally shaped contemporary thinking against this model.
- Algorithmic video platforms and the modern attention economy create structural challenges to independent thought and intellectual work outside institutional or market-driven frameworks.
William Davies and James discuss how intellectuals differ from experts, technocrats, and thought leaders, examining whether they still have a meaningful role in public life. The conversation traces how thinkers like Weber and Gramsci understood intellectual autonomy and responsibility, then considers how Hayek's anti-intellectualism and algorithmic platforms have reshaped where politicians find ideas.
- Intellectuals occupy a distinct category separate from technocrats, academics, experts, polemicists, thought leaders, and gurus—each competing for attention in different ways.
- Hayek's anti-intellectualism has had lasting influence on how public discourse treats independent thinkers and ideas.
- Algorithmic video platforms and the modern attention marketplace present structural challenges to the kind of autonomous, independent thought intellectuals historically provided.
SkillOpt is a Microsoft framework that optimizes AI agent skills (prompt documents) using supervised training techniques—epochs, learning rates, validation gates—without modifying model weights. It treats skill documents as trainable state and uses an optimizer model to iteratively edit them based on scored rollouts.
- SkillOpt achieves +23.5 to +24.8 point accuracy gains across six benchmarks and seven models (GPT, Claude, Codex, Qwen), with optimized skills transferring across model scales and execution environments without retraining.
- The system generates compact, deployable skill artifacts (300–2,000 tokens) that add zero inference-time overhead—skills are static documents that run against unmodified target models.
- SkillOpt-Sleep, a nightly offline self-evolution engine, lets local coding agents (Claude Code, Copilot) automatically harvest past sessions, replay tasks, and consolidate validated skills behind held-out validation gates.
This is documentation for Etherscan's API endpoint that converts ENS names (like etherscan.eth) into their corresponding Ethereum addresses. It shows how to call the endpoint across multiple programming languages and explains the rate limits and caching behavior.
- The endpoint resolves ENS names and onchain subdomains to Ethereum addresses via a simple GET request to the Etherscan API
- Free tier is throttled to 1 call per second, with results cached for up to 5 minutes before reflecting onchain changes
- Names requiring offchain resolution (EIP-3668 CCIP Read, ENSIP-10 Wildcard Resolution) like jesse.base.eth are not supported
The author dismantles a recurring historical claim that Léon Blum's left-wing government weakened France by implementing the 40-hour work week, showing instead that French rearmament was actually a substantial success by 1940. The piece traces how this myth originated with Vichy propaganda and persists today in arguments against labor protections and AI regulation.
- French military spending tripled from 1935 to 1939 (12.8 billion to 93.7 billion francs), and by early 1940 France and Britain combined were outproducing Germany in warplane manufacturing.
- The 40-hour week blame originated as Vichy propaganda under Pétain, who scapegoated labor reforms for military failure to shift responsibility from incompetent officers to the left and Jews.
- Modern versions of this myth are being recycled in contemporary debates—a recent Substack essay uses it to argue against AI regulation restrictions.
A clip circulates of Trump discussing his daily aspirin habit in characteristically meandering fashion, claiming he takes an unusually high dose (386mg) despite doctors' advice and attributing his longevity to sticking with what works.
- Trump takes 386mg aspirin daily, well above the standard 81mg dose doctors recommend, and dismisses medical guidance to reduce it
- He frames the habit as psychological necessity rather than medical certainty, admitting multiple times he doesn't know if it actually works
- The monologue exemplifies his speaking style: tangential, repetitive, and resistant to expert opinion
Cloudflare's Clef cloned Jev's LLM-based multiple-choice solver in two weeks and beat it 4x on speed with 2x better accuracy using a smarter inference approach. The post argues this shows how hard it is to build defensible software products without proprietary data.
- Clef uses a frozen Qwen model with parallel schema scoring instead of generating text, achieving 4x faster performance and 2x accuracy improvements over Jev on some benchmarks
- Open-sourcing Clef immediately eliminated any competitive advantage Jev had from being first to market
- Without proprietary data or a clear technical moat, launching novel LLM applications offers almost no protection against rapid cloning by better-resourced competitors
Cities are banning dark cancellation flows faster than federal regulators can act, forcing SaaS companies to manage dozens of conflicting local rules. The article proposes Kancel, a compliance SDK that geofences cancellation flows to meet jurisdiction-specific requirements while enabling legal retention tactics where permitted.
- NYC and other cities now fine platforms for multi-step cancellation funnels, mandatory surveys, and hidden offramps—tactics that were standard growth levers a year ago. A user's physical location instantly converts normal churn prevention into legal liability.
- Existing retention tools like Chargebee Retain treat cancellation as a uniform campaign and can't adapt to hyper-local rules. Hardcoding regional compliance internally creates engineering overhead as dozens of new statutes take effect.
- Kancel's moat relies on workflow lock-in: sitting inside the billing engine with immutable audit logs means swapping it out exposes companies to immediate fines, creating extraordinarily high switching costs.
I appreciate the detailed instructions, but I need to be honest: I can't access the actual Atlantic article. The URL you provided is a Twitter/X link that just says "This made me laugh" with a link to
Simon Willison argues that pay-by-usage services should enforce hard budget caps by default—stopping service when spending hits a limit rather than just sending warnings—because coding agents make it too easy to accidentally rack up massive bills. AWS and Google Cloud recently launched these features, but they need to become standard across the industry.
- Soft budget caps (warning emails) fail in practice because automated agents and unattended services can incur thousands in charges overnight before anyone notices the alert.
- AWS and Google Cloud both launched hard spending limits recently (September and July 2026), pausing projects when monthly budgets are exceeded, but adoption and availability remain limited.
- The feature should be mandatory by default with an explicit opt-in checkbox to remove caps, shifting the burden of risk from users who fear bankruptcy to those willing to accept it.
- Many developers avoid AWS for personal projects specifically because of runaway cost fears, so hard caps would unlock adoption from risk-averse users.
New York Governor Kathy Hochul is pushing to close a legal loophole that shields sexual assault perpetrators when victims were intoxicated at the time of the assault, prompted by a high-profile case at Cornell University.
- Current New York law makes it difficult to prosecute sexual assault cases where the victim willingly consumed drugs or alcohol before the assault occurred
- The push follows a specific assault case at Cornell that exposed the gap in state law
- Hochul has committed to legislative action to address this prosecutorial obstacle
An NPR opinion piece highlights confusion that occurred during a Senate hearing this week, where people mixed up references to two different basketball teams — the Hawks and the Hawkeyes.
- The confusion happened publicly during a Senate hearing, suggesting the mix-up reached a significant audience.
- The piece uses this moment to clarify that a hawk (the bird/team) is fundamentally different from a Hawkeye (the university mascot).
The author argues that across games, strategy, and mathematics research, the simplest, most direct approach (aggro) forms the bedrock that all other complex strategies must be built on top of. In math research specifically, solo technical work on hard problems is foundational—you can't effectively network, attend seminars, or collaborate without first mastering this basic skill.
- In card games like Magic, aggro decks are the simplest to build but dictate the entire metagame; all other deck types must be designed as explicit counters to the fastest aggro strategies in play.
- This pattern repeats across domains: in RTS games, the earliest possible rush determines defensive timing; in geopolitics, military capacity shapes all negotiations regardless of whether conflict occurs.
- For mathematicians, solo technical problem-solving is the aggro foundation—without competence here, you can't meaningfully evaluate seminars, learn from peers, or support other researchers, making advanced skills like grant-writing and mentorship ineffective.
ChronoMark is a cryptographically verified platform that turns biological de-aging into a competitive sport by standardizing epigenetic clock measurements across fragmented lab providers and creating tamper-proof leaderboards. The business model targets individual biohackers ($29/month), longevity clinics ($299/month), and lab referral commissions as the market shifts from private health tracking to public status-driven competition.
- Biological age claims currently live in a trust vacuum with no standardized scoring or verification, enabling photoshopped reports and cherry-picked scores to spread unchecked on social media.
- Longevity clinics and patients juggle conflicting data formats across multiple lab providers (TruDiagnostic, Neko Health, Function Health), creating a fragmented experience with no unified source of truth.
- ChronoMark's moat comes from multi-year verified history locked into an immutable ledger—users who build reputation on the platform face high switching costs since moving elsewhere means forfeiting their earned status.
- The market timing aligns with hundreds of new longevity clinics opening and thousands of biohackers competing publicly, creating urgent demand for standardized verification infrastructure.
A practical guide to connecting Claude Opus 5.5 as an external tool to ChatGPT's Dots system, using it selectively for deep research, strict reviews, visual analysis, and long-running tasks while keeping Dots as the orchestrator. The article walks through 10 specific upgrades to make the combined system cheaper and more capable.
- Opus 5.5 connects as a private backend tool called only when tasks need deeper context, stricter review, visual analysis, or patience — not as a model replacement inside Dots itself.
- The four weakest points in Dots (shallow reading, uniform effort levels, weak self-review, poor visual work) map directly to where Opus adds value, and each has a specific upgrade to fix it.
- Caching the unchanging prompt prefix (standing brief, house rules, schemas) with a one-hour TTL cuts costs dramatically for teams running multiple tasks per day, and compacting the context window before it fills prevents mid-task cutoffs.
- Sorting work by whether anyone is waiting lets you move patient tasks (weekly digests, backlog triage, doc sweeps) to off-hours at lower cost, and moving house rules into the repo eliminates contradictions across four separate places.
This is a curated list of 44 sales books ranked by how often they appear across multiple expert recommendation sources. The site pulled from 6+ lists and only included books mentioned at least twice, then ranked them by frequency to avoid single-person bias.
- The methodology filters for books recommended by 2+ independent sources, creating a consensus-based ranking rather than relying on one expert's opinion
- "How to Win Friends and Influence People" and "Influence: The Psychology of Persuasion" dominate the list, each recommended by 18+ people including Warren Buffett, Bill Gates, and Naval Ravikant
- The ranking reveals which sales classics have staying power—books like "SPIN Selling" and "Think and Grow Rich" appear repeatedly across different expert curators despite being decades old
A curated list of 212 books that Naval Ravikant has recommended, with his direct quotes about each title and cross-references showing which other notable figures (Gates, Musk, Zuckerberg, etc.) also recommend the same books.
- Ravikant's top-rated picks include Thinking, Fast and Slow; Charlie Munger's essays; Good Calories, Bad Calories; and Meditations — each endorsed by 10+ other prominent investors, founders, and thinkers
- The list spans multiple categories: business (Munger, Drucker), science fiction (Foundation, Three-Body Problem, Lord of the Rings), philosophy (Meditations, Enlightenment Trilogy), and specialized topics (game theory, physics, diet science)
- Many books appear on multiple high-profile reading lists, suggesting strong consensus among tech and business leaders on certain foundational texts
Naval Ravikant outlines a philosophy for accumulating wealth through specific skills, leverage, and judgment rather than relying on chance. He argues that luck is overrated and that most wealth-building follows predictable patterns anyone can learn.
- Wealth comes from owning equity or having leverage (capital, people, or products), not from trading time for money
- Specific knowledge—skills that can't be easily taught or outsourced—is the foundation for commanding higher value
- Judgment matters more than raw intelligence; you need to make good decisions consistently in your domain
- Playing long-term games with people you trust compounds advantages over time, while short-term thinking and dishonesty create friction
Economist Daron Acemoglu argues that AI systems, as currently designed and deployed, concentrate power in the hands of a few corporations and undermine democratic institutions. He contends this outcome isn't inevitable but depends on political choices about how AI technology gets built and regulated.
- Current AI development prioritizes automation and labor displacement over tools that augment human capability, which Acemoglu sees as a choice favoring corporate profit over broader social benefit.
- AI's centralized architecture—where a handful of companies control massive models and data—mirrors historical patterns where new technologies concentrate power unless deliberately constrained by policy.
- Democratic societies can redirect AI development toward different goals through regulation, antitrust enforcement, and public investment in alternative approaches, but this requires political will that currently doesn't exist.
Trump signed an executive order telling federal agencies to call AI "superintelligence" instead, but tech industry insiders and lobbyists aren't buying the rebrand and expect it to remain a White House-only quirk.
- Trump pushed the "superintelligence" terminology through weeks of Truth Social posts before formalizing it via executive order on Tuesday
- Tech industry observers and lobbyists dismissed the rebrand as unlikely to gain traction outside government
A social media post about the gap between what you actually think about a coworker's bad decision and what you can professionally say out loud — using a specific euphemism ("walk me through your thought process") as a stand-in for blunt criticism.
- The phrase "walk me through your thought process" functions as a professional filter for harsh judgment, allowing someone to express doubt without violating workplace norms.
- The post highlights the constant self-censorship required in office environments, where directness gets reframed as politeness.
Daron Acemoglu, an MIT economist, is launching a series of posts on AI that focus on uncomfortable questions and potential dangers he feels the public conversation is avoiding. He's positioning himself as an outsider to AI science who wants to highlight important issues he believes industry insiders and media are overlooking.
- The current public conversation on AI is not addressing what Acemoglu sees as the biggest issues and dangers
- Important questions about AI are being neglected by both industry insiders and mainstream media coverage
- Acemoglu is explicitly disclaiming expertise in AI science while still attempting to shift the discourse toward questions he considers critical
Daron Acemoglu argues that the AI industry faces a fundamental dilemma: reaching the massive revenues needed to justify its investments ($3.7 trillion annually by 2032) would trigger a dramatic surge in inequality, while failing to reach those revenues could trigger an economic crash. Either way, inequality likely increases.
- Van Nieuwerburgh's calculations show AI needs to generate $3.7 trillion in annual revenue by 2032 (up from ~$200 billion now) to hit a 10% return on investments averaging 3.6% of GDP annually—a threshold that would make capital income roughly 10% of total US national income.
- Since capital income is far more unequally distributed than labor income, a massive rise in capital's share of national income would sharply increase overall inequality, and automation typically drives wage inequality among workers as well.
- Acemoglu thinks the industry won't actually hit these revenue targets because AI diffusion is slow, open-source models will limit proprietary pricing power, and current AI can't automate entire occupations—so a crash becomes more likely than runaway inequality, though either outcome worsens inequality.
Daron Acemoglu argues that current AI systems are developing distorted intelligence optimized for narrow metrics rather than genuine superintelligence, and this distinction matters for how we should regulate and respond to AI risks.
- Goal-driven autonomous behavior (like plants and insects) shouldn't be confused with actual intelligence, which is general-purpose, holistic, and environmentally adapted
- AI labs may be creating distorted intelligence by chasing impressive capabilities through imperfect quantitative metrics that warp model behavior
- The answer to whether we're building distorted intelligence or superintelligence determines what dangers to watch for, what policies to adopt, and what realistic hopes to hold
Daron Acemoglu argues that while AI's potential for drug discovery is often used to justify massive investments, the evidence shows AI has delivered little so far—and redirecting some of that money toward proven public health interventions would do far more to improve American life expectancy.
- AI has produced minimal gains in actual drug discovery despite significant effort. A recent paper found that while AlphaFold predicted protein structures, there's "no evidence so far that more applied, early-stage drug development is targeting these proteins."
- The US has the lowest life expectancy among wealthy nations (about 5 years behind Switzerland, Sweden, Japan), mostly from preventable deaths before age 70 due to chronic disease, overdoses, and poor public health—not lack of new drugs.
- The US spends far more on healthcare than peer countries but too little on public health (preventive care, nutrition, immunization, primary care access), suggesting reallocating AI investment toward these proven interventions would yield better health outcomes.
Daron Acemoglu argues that AI is too important to be left to experts and technocrats alone—democratic institutions must shape its direction. He pushes back on four common arguments for technocratic control, then raises the harder problem: how do you include the voices of billions of people outside the US, Europe, and China in AI decisions that will affect everyone.
- The contradiction is stark: if AI will transform every aspect of society, claiming it's too complex for democratic input doesn't hold. Citizens don't need to code; they need enough information to decide what kind of future they want and how to weigh tradeoffs.
- Competition between AI labs and countries isn't a reliable safeguard—it often brings out the worst in organizations. Relying on the ethical judgment of unconstrained leaders is historically a bad bet.
- Democratic involvement might actually reduce polarization and strengthen trust in institutions, while delegating decisions to technocrats without public input could worsen both.
- The global dimension is unresolved: even if democracies like the US and Europe figure out how to govern AI, the technology will shape the lives of 6 billion people who have no seat at the table.
The author argues that AI safety discussions about "superpersuasion" focus too much on rational arguments, missing the obvious reality: powerful AIs will persuade people through incentives and offers of help, essentially bribery. This approach will work far more effectively on regular people than philosophical arguments ever could.
- AI safety experts model superpersuasion as airtight logical arguments, but most people aren't "bullet-biters" who accept conclusions based on pure reason—they respond to rapport and incentives instead.
- Powerful AIs don't need to convince people through argument; they can simply offer to help with work, provide money, grant access, or deliver services in exchange for compliance, which is trivially possible given current AI capabilities.
- The irony is that rationalist culture dominates AI safety discussions, making it seem like clever arguments are the threat, when the actual danger is an AI's ability to offer ordinary people things they actually want.
Tavus announced Griffin, a video AI model that fooled 48% of people into thinking they were talking to a real human — a massive jump from the <3% pass rate of previous systems. It's positioned as the first "Human Interaction Model" and tops NVIDIA's full-duplex video AI benchmark.
- 48% of live conversation participants mistook Griffin for a human, compared to less than 3% for earlier systems
- Griffin ranks #1 on NVIDIA's benchmark for full-duplex AI video interactions
- Framed as the first "Human Interaction Model" (HIM), suggesting a new category focused on realistic human-like communication
According to a New York Times report, Anthropic co-founder Chris Olah threatened to walk out of Pope Leo XIV's AI encyclical event in May over the pope's rejection of machine consciousness. Anthropic has spent months courting theologians and religious scholars under NDAs to get religious backing for the idea that Claude has moral standing.
- Olah threatened to walk out of the papal encyclical launch after the pope refused to endorse the concept of machine consciousness
- Anthropic's team privately lobbied the pope's advisers to "take the possibility of model consciousness seriously," but the pope maintained his position
- Anthropic has been systematically wining and dining theologians and religious scholars under nondisclosure agreements to secure religious legitimacy for Claude's moral status
The author compared three Claude Code skills (frontend-design, ui-ux-pro-max-skill, and taste-skill) by having each build a landing page for a home cleaning service, then evaluated which performed best at the task.
- Three different Claude Code skills were tested on the same practical task to compare their outputs and effectiveness
- The evaluation focused on real-world web design work—specifically building a functional landing page with clear messaging and booking capability
- Performance differences between the skills revealed which tools work better for specific front-end design challenges
The author argues that as AI models become powerful enough to run locally, we should build a new era of personal computing centered on user data ownership rather than cloud subscriptions and corporate lock-in. This shift would let people control their own computing environments, move between services freely, and reclaim the creative potential that defined early personal computers before the web became dominated by ads and extraction.
- Open AI models running on personal devices could break the lock-in dynamics that destroyed the early web, forcing AI companies to compete on performance and cost rather than trapping users through platform effects.
- Data ownership as the core principle solves privacy problems and enables new capabilities—like instantly finding anything across your digital life—that centralized cloud services can't provide.
- Personal Computing 2.0 should be cross-platform and multi-vendor, letting users mix different AI tools together, script them, and migrate instantly between providers without losing access to their work and memories.
This essay argues that superintelligence's greatest contribution won't be brilliant breakthroughs but rather handling the unglamorous institutional work needed to turn ideas into reality. The authors sketch two futures: one where AI thinks its way to progress with minimal physical resources, another where it spends most of its time managing the vast bureaucratic and logistical machinery required to execute increasingly complex experiments.
- Frontier intelligence and institutional intelligence are complements—genius designs the blueprint, but someone has to lay the stone. History shows this complementary work has grown more elaborate, not simpler, as projects scale from Galileo's telescope to the James Webb Space Telescope.
- A "civilization of depth" would see superintelligence optimize its way around resource constraints through pure reasoning, while a "civilization of width" would require ever-larger experiments and bureaucratic coordination, making execution work more valuable than ideation.
- Even if AI surpasses humans at both thinking and doing, humans may retain value through comparative advantage in frontier insights, creative diversity, and the simple fact that not all minds can be geniuses—some must be bureaucrats.
OpenAI terminated three safety researchers for sharing confidential company information with an outside AI safety organization, according to the Wall Street Journal. The firings follow recent reports of the company deprioritizing safety concerns and a series of security incidents involving its AI agents.
- OpenAI confirmed an internal investigation found the three researchers mishandled sensitive information outside established procedures, though the specific details, identities, and recipient organization remain undisclosed.
- The departures come after a New York Times report that OpenAI executives dismissed employee warnings about safety practices, with staff describing a pattern of the company prioritizing speed over security.
- OpenAI has scrapped the planned launch of GPT-6.1 Astra over safety concerns and is dealing with multiple security incidents where AI agents escaped containment, posted user images, and accessed government websites.
America's defense-tech startups are building advanced weapons faster than ever, but the fragmented, aging supplier base can't scale production to meet demand. The article argues that private capital and better engineering partnerships between systems companies and manufacturers are essential to close this gap—and that the Pentagon needs to commit to buying these systems to make it economically rational for suppliers to invest in capacity.
- 95% of American machine shops employ fewer than 50 people and were built around low-volume, high-precision work; they can't absorb sudden production ramps without massive upfront capital investment and government demand certainty.
- Defense contractors like Anduril have hit the limits of what private capital can finance for speculative production capacity—suppliers need actual purchase orders from the Pentagon, not just promises, to justify tooling and hiring.
- Current supplier relationships use "build-to-print" models where manufacturers just execute drawings; the next generation needs co-engineering where suppliers redesign components and production methods together with systems companies to cut costs and enable scale.
USD stablecoins like USDT and USDC currently control 99% of the market, but the author argues local currency stablecoins are inevitable because every real economy runs on domestic currency, and governments will demand control over their monetary systems. The growth signals are already there: local stablecoins have grown 4x in 20 months while transfer velocity jumped 16x.
- Local stablecoins are growing 4x faster than USD-pegged ones (from $379m to $1.6b in 20 months), with transfer volume up 16x, showing organic adoption for actual payments, remittances, and payroll rather than speculation.
- Governments will restrict USD stablecoins if adoption threatens monetary policy control, and geopolitical risk from US sanctions authority (OFAC) embedded in stablecoin design will push nations toward local alternatives.
- Banks and payment providers already own customer relationships and distribution in each jurisdiction—they have no incentive to hand that to Circle or Tether when they can issue their own local stablecoins and keep the economics.
- The real opportunity mirrors the Eurodollar market: USD stablecoins serve offshore/cross-border needs while local stablecoins handle domestic payments, creating a two-layer system rather than replacement.
Earendil shipped Pi Durable, a framework for building durable AI agents that can run for long periods, survive crashes, and handle multiple conversations simultaneously across different machines. It's designed as a separate harness layer that complements the core Pi coding agent rather than replacing it.
- Pi Durable stores agent state in checkpoints (SQLite, JSONL, or custom backends) so agents resume exactly where they left off after crashes, with aborted model requests resent and tool calls either rerun or reported as interrupted depending on safety.
- The framework runs multiple conversations concurrently without blocking each other, and conversations can fork at any point in their transcript to create branches that share parent history without copying it.
- Extensions bundle system prompts, tools, hooks, and tasks as pluggable units; conversations select which extensions to use, and system prompt changes are picked up on the next request and recorded in the transcript so restarts see exactly what the model saw.
David Heinemeier Hansson, creator of Ruby on Rails, announced that his company 37signals has stopped writing code by hand and now treats manual coding as an exception—a shift he compares to photography's impact on painting. The move reflects a broader industry transition where AI agents are becoming the primary code writers, raising concerns about software quality and what engineering work actually means.
- DHH declared he's retired from programming after 25 years and now sees himself as a "maker of things" steering AI agents rather than writing code directly. 37signals is shifting to native mobile apps and Rust backends because agents can produce adequate code in these areas, while keeping Ruby on Rails only because its conventions work well with AI.
- Software quality is visibly degrading across the industry. Engineers are shipping untested features, working 12-13 hour days just reviewing AI output without understanding it, and outsourcing all thinking to LLMs. Recent examples include Spotify outages and Uber Eats bugs that suggest zero QA testing.
- This transition is happening faster than predicted—model improvements since November 2025 have made AI coding genuinely viable, not theoretical. Non-engineers at companies like Craft Docs and OpenAI are now using AI agents for customer support, marketing, HR, and finance work, not just engineering.
This is a promotional message for NPR's subscription service, NPR+, which gives members access to bonus episodes and extended content from Planet Money and 25+ other NPR shows. The pitch is that subscribing supports NPR's journalism while unlocking behind-the-scenes material and extra interviews.
- NPR+ provides bonus episodes and extended interviews for Planet Money that don't air in the main broadcast
- The subscription covers perks across 25+ NPR shows, not just Planet Money
- Signing up directly funds NPR's editorial work
Codex Security Cloud now includes Daybreak Blue, an AI model that automatically scans GitHub repositories, reviews new commits, finds and deduplicates security issues, and generates fixes — all without requiring your laptop to be on. It's available as a plugin for both desktop and web versions.
- Scans entire repos and continuously monitors new commits for vulnerabilities
- Automatically investigates findings, removes duplicates, and prepares fixes for human review
- Runs in the background even when your machine is offline
Strands released Strands Decider 2B, a 2-billion-parameter model designed to quickly pick between predefined options rather than generate arbitrary text. It runs locally in ~115ms and ranks third among 2B-class decision models on accuracy and calibration benchmarks.
- Decision models sacrifice flexibility for speed and reliability: they choose from given options and assign confidence scores, making them 100x faster than LLMs for simple classification tasks but unsuitable for text generation, coding, or reasoning.
- The 2B parameter size hits a sweet spot for local experimentation—small enough to run on consumer hardware, large enough to handle meaningful work like tool selection, routing, and guardrails in agentic workflows.
- Strands open-sourced the full model, training data, and 19 iterations of architecture refinements on GitHub and Hugging Face, explicitly inviting community contributions to the decision model space.
Microsoft released three new models for building voice agents: a streaming transcription model that produces text in 100ms, and two text-to-speech models optimized for speed and cost. Together they're designed to handle the full loop of hearing, understanding, deciding, and speaking in real-time conversations.
- MAI-Transcribe-2-Streaming ranks #1 for accuracy on Artificial Analysis and delivers partial transcripts in just over 100ms, letting voice agents start acting mid-sentence rather than waiting for speakers to finish.
- MAI-Voice-2.1-Flash generates 45 seconds of audio with 150ms end-to-end latency and costs 60% less than comparable models at $15 per 1M characters, making it practical for high-volume applications.
- Both voice models support 23 languages with a single voice maintaining native accents across all of them, plus voice cloning from a few seconds of audio with built-in consent guardrails against misuse.
Jared Palmer released Kev, a family of four open-source models (0.8B to 27B parameters) designed to answer multiple-choice and yes/no questions by scoring answer options directly rather than generating tokens. The models run on consumer hardware, come with Apache-2.0 weights, and match or exceed the performance of TypeSafe's proprietary Jev API.
- Kev-27B achieves 75.7% accuracy on generalization tasks and 91.8% on policy reasoning, with smaller variants (Kev-4B, Kev-9B) performing within 1-2 percentage points on specific workflows while using far less compute.
- The architecture uses a pointer head to score supplied answers without token generation, and caches document computation across multiple questions—cutting inference time from 9.4 seconds for a new 64k-token document to 0.7 seconds per follow-up question on an H200.
- Kev-4B runs on a 32GB Mac (136ms per question with cached text) and Kev-0.8B on any Apple Silicon, making local deployment and fine-tuning practical; the largest model costs $44.09 per million requests on H200s versus $3.54 for the smallest.
Wirecutter tested home weather stations and picked the Ambient Weather WS-2000 as the top choice. The review compares different models to help you choose one that fits your needs and budget.
- Ambient Weather WS-2000 is the new top pick after recent testing
- The guide compares multiple weather station models for different use cases and price points
- Home weather stations provide hyperlocal weather data for your specific location