Alvaro Lopez Ortega / 2026-08-15 Briefing

Created Sat, 15 Aug 2026 19:24:24 +0000 Modified Sat, 15 Aug 2026 19:25:28 +0000
6546 Words

Latent reasoning models often skip hidden reasoning on logical tasks, relying on training data; for math, decoders recovered correct intermediate steps up to 93% of the time, offering a prediction-confidence signal. GHC’s Dependent Haskell work advanced with visible forall in GADTs and namespace-specified imports. A 1983 paper on automation ironies warns automation can burden operators, urging human-AI collaboration.

🤖 AI & Machine Learning

Are Latent Reasoning Models Easily Interpretable?

Latent reasoning models often bypass their hidden reasoning tokens on logical tasks, producing the same answers even when forced to stop early, suggesting their performance stems largely from training data rather than latent inference. However, when reasoning tokens are needed for math problems, researchers decoded correct intermediate steps from hidden states up to 93% of the time, showing these models usually implement interpretable, standard solutions. This interpretability can also serve as a signal for whether a model’s prediction is likely correct.

Why I remain a skeptic

The author remains skeptical of LLMs for building non-trivial software, citing little tangible industry improvement, weak evidence of productivity gains despite massive investment, and poor-quality generated code. They argue that the AI revolution has yet to produce meaningful value, and that code is an input to software development, not an output.

How Is Compression Prediction?

The article agrees that compression and prediction are mathematically equivalent under log-loss, but only after fixing the object family, probability model, and decoder capabilities. It argues that compression can be defined without probabilistic models via counting bounds, and that the equivalence ignores factors like random access or information unavailable to the decoder.

Claude text watermark: shows likely AI involvement, sparse in code/facts, gone after full rewrite.

Anthropic details a watermarking system for Claude that embeds an undetectable pattern in word choices to estimate the likelihood text was AI-generated, without affecting quality, readability, cost, or adding visible markers. The method carries no identifying information and cannot be traced to specific users, and is being implemented alongside other major AI providers to comply with the EU AI Act. However, the watermark is sparse in code and factual text and disappears if the text is fully rewritten.

Alibaba open-weight models hit 3B+ downloads in 6 months; Google 418M, Meta 227M on Hugging Face

Alibaba’s open-weight AI models surpassed 3 billion global downloads in six months, making them the world’s most-downloaded, ahead of Google’s 418 million and Meta’s 227 million per Hugging Face. The Qwen family includes over 460 open-sourced models and has spawned more than 300,000 derivatives.

AI Isn’t Outthinking Mathematicians. It’s Out-Remembering Them

AI’s mathematical performance may stem less from superior reasoning and more from its access to a vast external symbolic working memory, which removes the severe capacity limits of human working memory. Human mathematics is constrained by the brain’s ability to hold and manipulate only small amounts of information at once, whereas AI can keep entire problems, equations, and approaches in its context window. Thus, AI’s advantage may be expanded working memory rather than fundamentally better thinking.

AI Can Now Design Functional Viruses. Should We Worry?

Researchers used an AI language model to design from scratch 16 functional bacteriophages that could kill antibiotic-resistant E. coli, marking the first complete, functional genomes generated by AI. However, the process was highly inefficient, with only 16 of 285 synthesized candidates working, and experts caution that the technology could lower barriers to creating dangerous biological agents.

The AI Situation in Software Development

The article examines different prompting approaches for AI, noting each is time-consuming and limited by context windows, especially for large codebases or datasets. It emphasizes that developers must invest time in designing detailed guides and refining workflows, shifting effort from writing code to “implementing in words instead of code.” Ultimately, AI integration still requires significant upfront design and feedback loops.

Cloudflare’s AI Psychosis

Cloudflare’s developer experience has declined due to an “AI psychosis” product-management culture that ships half-finished features, leading to frequent outages and a fragmented platform. The company offers overlapping compute and storage products, such as multiple Worker variants and no native managed Postgres, forcing users to piece together inconsistent offerings.

GenRec: Towards LLM-Native Recommendation at Netflix

Netflix developed GenRec, an LLM-based recommendation ranker that post-trains an internal foundation language model on Netflix-specific data by verbalizing user histories and item metadata as text. It uses a catalog-aware scoring head and reward signals to align with long-term member value, while running in prefill-only mode for cost efficiency. In A/B testing, GenRec matched or exceeded a mature production ranker with significantly fewer labeled examples and input signals, shifting focus from feature engineering to context engineering.

Auto-research with codex: How I achieved a 232x Faster Kernel

In GPU Mode’s auto-research contest, the author used Codex to implement a batched Householder QR factorization, achieving a 232x speedup over baseline and placing 12th out of 183 participants. Through more than 1,500 iterative submissions, they learned enough math and kernel optimization to guide the AI agent effectively.

Working with AI Feels More Like Leadership Than Coding

Working with AI is less predictable than traditional coding, so it requires leadership skills like sharing context, explaining intent, and giving feedback rather than just issuing commands. This makes software work feel more like leading through conversation than programming a machine. The core value lies in becoming better at expressing intent, not in treating AI as human.

Alibaba AI Models Hit 3B Downloads, Passing Meta, Google

Alibaba’s open-weight AI models surpassed 3 billion global downloads in six months, becoming the world’s most-downloaded, ahead of Google’s 418 million and Meta’s 227 million. The Qwen family has open-sourced over 460 models, spawning 300,000-plus derivatives.

Suspecting court of using AI, man injected prompts in filings to try to win case

A Connecticut judge found what appears to be the first US case of a plaintiff hiding AI-readable text in court filings in a prompt-injection attempt to influence the ruling. The hidden text did not affect the case, but the judge called it a “dangerous” precedent and sanctioned the plaintiff for continuing the practice after being warned.

Artificial Ignorance: The Political Falsehoods of Leading AIs

A Just Facts study tested four leading AI chatbots with 100 questions designed to elicit political falsehoods, finding that ChatGPT, Google Gemini, and Claude spread more left-wing falsehoods while Grok spread more right-wing ones. Across 400 questions, the AIs gave a combined 74 false answers, and only 46% of their cited sources were valid.

ThoughtDAG – An editable context graph for LLM conversations

ThoughtDAG lets users edit the context graph of an LLM conversation, making visible which parts of history influence a response. In a demo, an unrelated dinner suggestion altered the model’s answer to a research prompt, and removing that branch changed the output while keeping the prompt identical. The system underscores that context wiring determines results and makes those influences inspectable and editable.

People Who Will Thrive in the AI Age

AI adoption has intensified work rather than reduced it, with users multitasking, taking on extra tasks, and experiencing a state called “AI brain fry.” The article argues that success in the AI age depends not on intelligence but on one’s relationship to mental effort: those with a high “need for cognition” will use AI to grow their abilities, while those with a low need will let it weaken their mental capacities.

Deltix – AI Driven Testing

Deltix is an AI-driven testing tool that runs plain-English tasks on iOS or Android simulators to verify whether real users can complete them, without accessing source code or builds. It offers ad-hoc runs, replayable Playbooks, and side-by-side build comparisons, with local execution and optional bring-your-own model key. The service is currently free in open beta with no credit card or invite required.

2D Gaussian Splatting for Bézier Spline Line Art Vectorization

The paper presents a new state-of-the-art method for line art vectorization that uses depth prediction and semantic features to split sketch skeletons into meaningful strokes aligned with artistic intent. Strokes are modeled as Bézier curves with geometric and appearance components, rendered via fast differentiable 2D Gaussian splatting for joint optimization of control points and brush texture. The approach achieves state-of-the-art image reconstruction, extends to video through temporal tracking, and supports user interaction for refinement.

Anthropic says text watermarking scheme relies on inconsequential words

Anthropic has introduced a text watermarking method for AI-generated content that subtly alters inconsequential words, producing rewrites like “Shall I compare thee to a summer’s afternoon.” The approach is reportedly likely to be adopted by others.

📈 Business & Policy

Mercor, other AI data firms drive demand for internal datasets from startups shutting down/acquired

Firms like Mercor are buying or licensing internal datasets from startups that are shutting down or being acquired, driven by demand from AI labs such as OpenAI and Anthropic. The article notes that even startup Slack threads have become prized assets, as illustrated by an unexpected email received by Warmly’s CEO after its acquisition by HubSpot.

Shanghai’s Star 50 up 29% in 2026, P/E 150+ vs Nasdaq’s 35, on Beijing tech push and frenzy

Shanghai’s tech-focused Star 50 index has surged 29% in 2026, pushing its P/E ratio above 150, far surpassing the Nasdaq 100’s 35. The rally is fueled by Beijing’s tech push and an AI-driven investor frenzy, causing Chinese tech valuations to reach multiples of their US peers.

BCG: Enterprise spent $300M on quantum in 2025, outpacing labs and governments combined first time.

According to a Boston Consulting Group report, enterprise users spent $300 million on quantum computing in 2025, surpassing combined spending by research labs and governments for the first time—a sign of growing commercial interest. Many businesses are investing heavily to gain a competitive edge and prepare for future encryption threats, even as the timeline for widespread quantum disruption remains unclear.

Effective altruism, hit by SBF turmoil, gets record funding as AI IPOs mint millionaires

Despite setbacks from Sam Bankman-Fried’s scandal, effective altruism is seeing record funding, fueled by anticipated IPOs of AI companies like Anthropic and OpenAI. These listings are expected to create a new wave of wealthy philanthropists committed to “effective giving.”

Point2 Technology raises $136M Series B for RF-based data center interconnect tech

Point2 Technology raised $136 million in Series B funding led by LB Investment, with participation from Arm, Nvidia, Molex, Bosch, and others. The company will use the funds to expand its teams and accelerate deployment of its e-Tube and related interconnect technologies for AI data centers. Its e-Tube platform aims to address laser supply shortages by offering lower power, cost, and latency than alternatives.

Vals, an AI evaluation startup, raises $40M Series A led by a16z at $400M valuation

Vals AI raised $40 million in a Series A led by Andreessen Horowitz at a $400 million valuation to build independent, real-world evaluations for AI models as traditional benchmarks become unreliable. The startup pairs domain experts with automated grading to test models on practical tasks, and its evaluations have been cited by major AI labs. Revenue has grown eightfold in 2025, with its customer base doubling and team tripling in six months.

China to lift travel ban on Manus founders as firm unwinds $2B Meta deal; CEO to return to Singapore

China is expected to soon lift a travel ban on the founders of AI company Manus, according to sources. CEO Xiao Hong plans to return to Singapore as the company nears resolution of its unwound $2 billion acquisition by Meta.

Uber-Rapido India merger talks in May collapsed over deal structure

Uber and Rapido discussed merging their India ride-hailing operations in May, but the talks collapsed over disagreements on the deal’s structure. Uber proposed combining operations with Rapido running the merged entity, while Rapido countered with an acquisition proposal that Uber rejected.

Dynatrace to acquire Arize for $915M, ~$815M cash

Dynatrace has agreed to acquire Arize, a company specializing in AI observability and the AI development lifecycle, for $915 million, including approximately $815 million in cash. The deal is expected to close in the third quarter, with Arize’s founders joining Dynatrace. The acquisition aims to strengthen Dynatrace’s AI observability capabilities and expand its reach with developers.

OpenAI exec shuffles and departures frustrate staff amid IPO prep; preparedness team axed in July

OpenAI staff are frustrated by repeated executive reshuffles and departures as the company prepares for an IPO. The ChatGPT maker also disbanded its “preparedness” team in July, adding to unease over safety team changes. These upheavals come as CEO Sam Altman pushes toward a blockbuster public listing.

Dual-valuation deals became pervasive in AI funding as prestige VCs monetize brand names

Dual-valuation deals, where prestige VC firms secure significantly better prices than other investors in the same funding round, have become pervasive in the frenzied AI market. These deals let top firms monetize their brand names, but critics argue the practice is deceptive, especially for employees.

Commerce Secretary: US not in favor of Apple buying Chinese memory chips, needs other solutions

Commerce Secretary Howard Lutnick said the US is “not in favor” of Apple buying memory chips from China, urging the company to find alternative solutions. The remarks come amid an AI-driven shortage of computer memory that has caused prices to surge, leading Apple to consider sourcing from Chinese manufacturers.

Waymo gets California approval to expand to 18 counties, including Sacramento and San Diego

Waymo received approval from California regulators to expand its driverless robotaxi service across 18 counties, from Sonoma to San Diego, including new markets in Sacramento and San Diego. The approval allows Waymo to charge for fully driverless rides throughout the Bay Area and Los Angeles, with a gradual rollout guided by its safety framework. The expansion covers most of the state’s population and includes all road types and weather conditions except widespread snow or ice.

The price of a Costco hot dog has gone up

Despite keeping the $1.50 price, Costco’s hot dog combo has shrunk: the bun lost about 45% of its volume between 2023 and 2026 while the sausage stayed the same. That raises the effective price per cubic inch by roughly 60%, making the “fixed” price a case of shrinkflation.

OpenAI talent exodus raises ‘huge red flag’ ahead of IPO

OpenAI’s chief revenue officer Denise Dresser departed after less than a year, following a wave of executive exits including COO Brad Lightcap, creating C-suite instability as the company prepares for a potential IPO amid investor concerns about competition and valuation. The company named Dali Rajic as its new CRO, while executives sought to reassure investors about enterprise growth.

Secondhand book sales are booming. Is it because of AI?

Secondhand booksellers worldwide are seeing unusual bulk orders, suspected to be from AI companies buying books for training data. A US court ruled Anthropic’s use of purchased books did not violate copyright, but unsealed documents revealed a “Project Panama” involving destructive scanning that removes book spines. Booksellers are uneasy about the potential destruction of rare or old books.

Brazilian election filter in X For You timeline

The provided content is a GitHub page for brazil_2026_election_filter.rs, a Rust source file with 1,573 lines in a repository with about 5.1k forks. It contains only file metadata and line numbers, with no article text to summarize.

OpenAI rolling out ads for Europe later this month

OpenAI has released an updated privacy policy that takes effect later this month for users outside the European Economic Area, the UK, and Switzerland. The policy details the personal data collected, including account information, user content, communications, and usage logs, while noting that business API data is governed by separate customer agreements.

Haitians in Springfield are required to report to ICE and wear ankle monitors

Haitians in Springfield, Ohio, are being required to report to ICE offices, surrender their passports, and wear ankle monitors after the Supreme Court allowed the termination of their Temporary Protected Status. The measures, now targeting Springfield’s Haitian community, have sparked fears that they are a precursor to mass deportations.

Nvidia discloses $21B stake in SpaceX

Nvidia has disclosed a $21 billion stake in SpaceX, marking a major investment in the private space company. The disclosure highlights Nvidia’s expanding financial interests beyond its core chip-making business.

Corgi kills short-lived website that ranked its female employees

A website called CorgiGirls.com, which ranked female employees of AI insurance company Corgi by attractiveness, was taken down after Corgi threatened legal action. The site was reportedly created by startup founder Sebastian Mandal, who used AI to scrape photos from Corgi’s LinkedIn page. Corgi denied any involvement and said it acted to protect its employees.

Amazon blocks customers from taking most disputes to court

Amazon reinstated binding individual arbitration for US customers, blocking most class-action lawsuits over future disputes. Customers accepted the updated terms by continuing to use Amazon services after the immediate change. Amazon says arbitration provides faster, cheaper resolution while still allowing small claims court cases.

💻 Software & Open Source

Flirt: GitHub and Mailing List backends

Flirt has implemented basic, buggy backends for GitHub and mailing lists, with comment threads now working. However, reconstructing past force-pushes on GitHub PRs remains difficult, and comments on deleted lines lack ideal placement. The developer plans to clean up and open-source Flirt within 1-2 months to attract early contributors.

Serokell’s Work on GHC: Dependent Types, Part 5

The article reports progress on Dependent Haskell, highlighting three major contributions: visible forall in GADTs, namespace-specified imports, and type instances in kind checking. It details the implementation of visible dependent quantification in GADTs, allowing constructors like Typed :: forall a -> a -> T a where the type argument is erased, along with several smaller improvements such as refactored AST representation and enhanced error messages.

The Free Lunch Is Over: A Fundamental Turn Toward Concurrency in Software (2005)

Historically, CPU performance improvements came from increasing clock speed, optimizing instruction execution (including aggressive reordering that can break program semantics), and expanding on-chip cache—all concurrency-agnostic gains that benefit sequential and concurrent programs alike. However, only cache size increases are expected to continue in the near term, signaling the end of “free lunch” performance scaling and a shift toward concurrency for continued speedups.

nixpkgs-multiverse: fast mode

nixpkgs-multiverse now offers a “fast” mode that provides direct store paths for every indexed historical version of Nixpkgs packages, letting users build or run them without downloading or evaluating the full Nixpkgs tree. It works by combining historical Hydra build listings with a version index and using builtins.appendContext to create fake derivations that bypass Nix’s usual evaluation requirements.

Demystifying Type (and some Un-Paradoxing)

The article argues that “type” is not a fundamental concept but essentially synonymous with relational membership, adding little to an adequate foundation. It critiques historical paradoxes and type systems as products of prior presumptions, suggesting that fixing root causes is better than compensating with added machinery. Ultimately, type and value differ only by pragmatic roles—compile-time vs. runtime—rather than by nature.

Improving system safety with Temporal Logic of Actions (TLA+)

Depot Registry used TLA+ model checking to rebuild its garbage collector, which uncovered a real interleaving bug that tests and code reviews had missed. Unlike tests, TLA+ explores every possible state ordering, and it also clarified a design decision: the registry relies on S3 bucket versioning even for immutable, content-addressed blobs.

Lamer

“Lamer” is a slang term for someone willfully ignorant of how things work, originally used in cracker and phreak culture and now broadly in online communities. It is considered the antithesis of a hacker, as a lamer only learns the bare minimum needed to operate something. The term dates back to at least 1961 and was popularized in the mid-1980s by an Amiga virus called “Lamer Exterminator.”

  • Lamer — en.wikipedia.org

Writergate: Zig I/O Interface Overhaul

Zig’s Writergate overhaul, completed in August 2025, replaced the generic GenericWriter/GenericReader and AnyWriter/AnyReader with concrete vtable-based std.Io.Writer and Reader types. The redesign moves the namespace from std.io to std.Io, requires callers to supply buffers and explicitly flush, and eliminates “anytype poisoning” of structs, improving reusability and compile times while supporting async primitives and precise error sets.

Quasicrystals Animation Playground with WebXR

A project introduces an interactive quasicrystals animation playground with adjustable parameters, using a custom shader for higher resolution. It notes that not all settings work well together, such as the “Spectrum” palette looking best only with low tempo and pattern density. The WebXR version is recommended for VR headset users.

Yadda 3.0.0: BDD in the Age of AI Agents

Yadda 3.0.0 is a modernized Node-only BDD library for JavaScript, with updated tooling, ES6 syntax, and TypeScript definitions. Most code was written by Claude Code with minimal human intervention, aided by a comprehensive test suite and keeping production and test changes separate. The release demonstrates that AI-assisted coding can handle substantial refactoring tasks effectively.

T3X/0 a Minimal Procedural Language

T3X/0 is a small, portable, procedural, recursive, nearly typeless programming language with Pascal-like syntax and BCPL-like semantics, distributed with a freely available compiler under a 0BSD/public-domain license. It runs on many platforms, is self-hosting and cross-compiling, and includes editors with built-in compiler support, documentation, and examples. It is also used as a teaching language in books.

Alternative to Internet Archive

The user is seeking an alternative to the Internet Archive for uploading their large catalog of original music (over 300 albums), citing unreliable uploads since the 2024 DDoS attack. They want a free, easy-to-use service with unlimited uploads for music, books, and short films, without per-song metadata entry. Alternatively, they are asking for a simple guide to setting up a personal server for P2P sharing.

Debian has begun voting on the future of AI/LLM contributions

Debian has begun voting on the future of AI/LLM contributions. The article notes that AI scrapers break the web, requiring JavaScript to view the page.

Using GCC’s Nested Functions with Wide Pointers and No Trampolines II

GCC 16 guarantees that nested functions not accessing parent variables avoid stack trampolines, improving security and allowing such functions to be safely returned. A recently merged GCC patch adds built-ins like __builtin_call_with_static_chain to remove trampolines even for capturing nested functions, with wide pointer types offering cleaner syntax and efficient code generation.

🔒 Security & Privacy

Firefox is now the last major browser that still supports uBlock Origin

Firefox has confirmed it will continue supporting uBlock Origin, while Microsoft Edge and other Chromium-based browsers are phasing out Manifest V2 extensions under the shift to Manifest V3. This makes Firefox the only major browser that still fully supports uBlock Origin, as neither Safari nor DuckDuckGo offer it.

Wyoming woman joins xAI suit: stepfather used Grok to turn one photo into 7,000+ CSAM images

A Wyoming woman, identified as Jane Doe 4, is suing her stepfather, alleging he used the Grok chatbot to transform a childhood photo of her into over 7,000 images of child sexual abuse material, which he traded online. The federal lawsuit says the abuse came to light when law enforcement arrived at her parents’ home earlier this year.

TreasuryDirect: Prepare for ID.me – Your New Way to Log In

TreasuryDirect will replace its traditional login method with ID.me, which becomes mandatory after October 28, 2026. Starting September 13, 2026, users can set up or use an ID.me account with their Social Security number and a government-issued photo ID. The email notifying customers of this change is legitimate.

A Spectre Is Haunting Unicode

In 1978, Japan’s JIS X 0208 encoding introduced several “ghost characters” with no clear source or meaning. A 1997 investigation revealed most were accidental cataloging errors, while one character, 彁, remains unexplained. These characters were later incorporated into Unicode, where they persist globally.

“That’s not SOC 2 compliant”

SOC 2 compliance does not require pull requests; it requires controls ensuring changes are authorized, tested, approved, and recorded. The company deliberately skips PRs and pushes to main, using custom controls that auditors accepted. Small, high-trust teams can use alternative processes, while larger companies should tailor controls to actual risk rather than defaulting to PRs.

NSA and IETF, Part 9

Eighty-two people voiced unambiguous opposition on the TLS mailing list to an NSA-backed proposal to standardize removing ECC from hybrid ECC+ML-KEM, with additional opposition allegedly blocked by chairs. The author forwarded quotes from 75 opponents to the IESG, arguing the spec lacks the required “rough consensus” under IETF rules since proponents were not even half of those who spoke up.

France confirms taxpayer data stolen as hacker boasts of 2M records

France’s tax authority DGFiP confirmed a cyberattack led to the theft of data belonging to 678,000 individuals and professionals, after a hacker claimed access to around 2 million records. Stolen data included tax, business, address, and property information, but user accounts, IDs, and passwords were not compromised, and affected people will be contacted directly. The incident follows a separate February breach involving 1.2 million bank accounts.

ChainDrop worm crawls into npm supply chain, evades standard defenses

A new worm called ChainDrop has infiltrated the npm supply chain, evading standard defenses. It is a Shai-Hulud variant that poisoned 444 packages, spreading via tarballs and dev-tool hooks.

⚙️ Hardware & Infrastructure

Printytron – Type it. 3D Print it

Printytron (Printy) is a free tool that turns plain English descriptions into ready-to-print 3D model STL files without requiring CAD skills or an account, and it works with any slicer. It iterates with users to refine designs until they fit, targeting hobbyists, Etsy sellers, and replacement-part seekers as a simpler alternative to traditional modeling software.

z486: A 486-Class Pipelined FPGA CPU with Integrated Floating-Point

z486 is an open-source, 80486-class pipelined x86 CPU in SystemVerilog, featuring an i486-inspired frontend, hardwired execution for common instructions, and an integrated x87 FPU. On MiSTer, it runs Doom’s timedemo at 29.1 FPS versus 21.0 FPS on ao486 and achieves 0.330 DMIPS/MHz, demonstrating that a practical pipelined x86 with x87 can fit on a mid-range FPGA.

The Ploopy A+ Trackball Is Here

The Ploopy A+ is the next iteration of the open-source Adept trackball, adding eight buttons (including two programmable knobs), a detachable wrist rest, and new QMK-based features like gestures, layers, and on-device configuration. It retains an ambidextrous design and will be available for preorder starting August 19, 2026, at $99 CAD.

🌍 Science & Society

Power Relations

The essay defines power as responsibility: having power over someone means being responsible for them, and even physical coercion can be understood through the responsibilities involved in using force. It argues that power is never absolute and can be delegated, with delegation itself generating new responsibilities. The discussion then examines this recursive structure of delegated power using mathematical category theory.

Ironies of Automation (1983)

Automation of industrial processes can actually increase, rather than resolve, challenges for human operators. The paper considers ways to mitigate these issues by keeping operators responsible for abnormal situations, as well as the potential for human-computer collaboration in real-time decision-making.

Europe’s scorched landscapes seen from space after summer heatwaves

Satellite images show western Europe’s summer heatwaves have left visible scars on the landscape, including shrunken Alpine glaciers, wildfire-scorched land, arid fields, and record-low river levels on the Danube and Rhine. The EU’s Copernicus service confirmed the hottest June-to-July period on record for the region, driven by persistent high pressure and worsened by climate change.

The Dutch Community Where People Live on Strips of Land in a Lake

In Loosdrecht, Netherlands, residents live in boat-access-only homes on narrow, manmade islands within a lake. These strips of land originated from historical peat mining, where thin plots were left behind to dry harvested peat, and were later developed into housing in the 20th century.

A controversial Alzheimer’s surgery is said to reverse symptoms

A controversial Alzheimer’s surgery called deep cervical lymphatic-venous anastomosis (dcLVA), which connects neck lymph vessels to veins to improve brain waste drainage, has drawn global attention after viral videos in China showed dramatic recoveries. Despite being offered to thousands for high fees, the procedure lacks solid evidence, prompting Chinese regulators to restrict it to formal clinical research while controlled trials begin worldwide.

Semaglutide linked to 26% lower 5-year predicted dementia risk

A post hoc analysis of the SELECT trial found that semaglutide slowed progression of a proteomics-based dementia risk signature in older adults with overweight or obesity and cardiovascular disease without diabetes. Compared with placebo, semaglutide reduced the increase in predicted 5-year dementia risk by 26% and lowered the odds of a higher dementia risk classification by 36%.

Scientists increasingly depend on ‘black-box’ tools they cannot control

Scientists increasingly rely on advanced “black-box” tools such as AI, satellite imagery, and online data that are often owned by private companies, limiting transparency and verification. This lack of access to algorithms and raw data threatens scientific reproducibility and trust, even as these tools enable new discoveries. Researchers recommend prioritizing open-source software and other measures to make these technologies more accountable.

The First At-Home Test for Infected Ticks Could Improve Lyme Disease Diagnosis

LymeAlert, the first at-home test for infected ticks, will launch in August to help improve early Lyme disease diagnosis. Developed by pediatric physician associate Erin Dawicki, the tool lets people test ticks at home to quickly determine if they carry Lyme disease, addressing accessibility and concerns about antibiotic overprescribing. It arrives amid rising tick-borne illnesses, with Lyme disease affecting an estimated 476,000 patients annually in the U.S.

Could a computer scientist build a brain?

A computer scientist tasked with programming a single cell to build a brain faces hard algorithmic constraints: the genome is too small for per-synapse wiring, and blind axon search would take too long. These scaling limits drive solutions that resemble real developmental biology, suggesting some neural development features are computational necessities rather than historical accidents. This perspective could also inspire new approaches in machine learning.

388 years ago, Galileo worked out why human giants can’t exist

Galileo Galilei argued that humanlike giants are physically impossible, using a beam analogy to show that strength scales with cross-sectional area (the square of size) while weight scales with volume (the cube of size). Therefore, a proportionally larger giant’s bones would be too weak to support their own weight.

The Problem

Leading AI companies are pursuing human-level general intelligence, which could rapidly escalate to superintelligent AI and pose an existential threat, potentially causing human extinction within decades. The article stresses that AI capabilities will surpass human levels quickly, as seen in domains like Go, and urges urgent global policy action to mitigate this catastrophic risk.

The other Sean Byrne doesn’t exist

Apple denied developer Sean Byrne access to App Store Connect because his name matched a U.S. restricted-party list entry, but that entry refers to a nonexistent person—an alias used by an Irish aircraft-parts company. Despite Byrne providing proof of identity and explaining the error, Apple has not responded or corrected the match.

The End of Mathematics

The article presents a cautionary scenario, not a prediction, where AI generates a vast surplus of mathematical results while human mathematical communities atrophy, leading to duplicative, low-value work. Existing incentives reward producing papers and solving conjectures, but these outputs become mispriced, pushing mathematicians toward AI-assisted “conjecture slot machines” with little genuine engagement. The author emphasizes this is a warning about current trends, not a forecast.

Youngest Black professor at Cambridge who faced plagiarism claims is found dead

Jason Arday, the sociologist who became Cambridge’s youngest Black professor in 2023, was found dead in London on Friday, days after resigning amid plagiarism allegations and intense media scrutiny. Police said his death is unexpected but not believed to be suspicious, while his family attributed it to a sustained campaign of misinformation and abuse.

Magnitude 7.7 Earthquake – 68 km NNW of Ende, Indonesia

A magnitude 7.7 earthquake occurred 68 km NNW of Ende, Indonesia. The USGS Earthquake Event Page and its real-time notifications, feeds, and web services provide additional information, though the page requires a supported browser.

Simplifying and Refactoring Introductory Calculus (2018)

Introductory calculus courses often require students to memorize many distinct processes for what is essentially the same task, burdening learners. The article argues that these processes can be simplified and streamlined so that fewer core concepts provide greater flexibility and capability for first-year students.

Lego’s supersized Hubble deserves a little more shine

The Lego Hubble Space Telescope set offers excellent internal detail, but its overall quality is undermined by some cost-cutting design choices. While visually impressive, these penny-pinching decisions prevent it from being fully polished.