Alvaro Lopez Ortega / 2026-08-01 Briefing

Created Sat, 01 Aug 2026 19:09:37 +0000 Modified Sat, 01 Aug 2026 19:10:15 +0000
4926 Words

Reddit shares plunged 23% despite beating Q2 estimates as AI-generated Google summaries slowed user growth, prompting a review of its Google deal. Tesla faced privacy backlash after a Model X uploaded 2.6GB of data, fueling wider concerns about automaker data collection. BMW drew criticism for putting Spider-Man ads on dash screens in 70 countries. NetBSD 11.0 launched with open security issues slated for 11.1.

🤖 AI & Machine Learning

Explorative modeling: Train on the best of K guesses

Explorative Modeling (XM) introduces a new generative modeling paradigm that serves as a third pretraining axis for existing models and enables end-to-end generation. Increasing exploration monotonically improves image, video, and language models, with gains growing alongside data and parameter scale, yielding 6.2× sample efficiency, 4.1× FLOP efficiency, and 47% better parameter efficiency. As end-to-end generative models, XMs match diffusion on control tasks while using up to 256× less inference compute.

How to Spot AI Writing

AI-generated writing is now widespread across websites, emails, and even prize-winning fiction, according to the article. It notes that large language models may be detected through stylistic quirks, such as a preference for long words and em-dashes, but emphasizes ongoing uncertainty about reliable identification.

Minimal LLM Post-Training Experiments on an 8GB GPU (SFT, DPO, GRPO)

The article demonstrates minimal SFT, DPO, and GRPO post-training experiments using HuggingFace TRL on an 8GB GPU with a 135M model, reproducing the finding that on-policy RL drifts less from the original model and forgets less than SFT. It also shows that GRPO amplifies existing self-verification and search behaviors, similar to DeepSeek-R1’s reasoning, and that a 3B model experiment can run on a 48GB GPU for under $5 in five hours.

Can I vibe code it?

A project tracks 942 SaaS apps, asking whether AI can replace each one or if a real moat protects it. So far, 649 apps account for $16,619/month in MRR, and the site offers a newsletter with daily verdicts.

AI computing power to surge: Epoch AI sees AI chips doubling every nine months

A wave of new data centers will soon deliver unprecedented AI computing power, with the number of AI chips expected to double every nine months to about 200 million by 2028. Recent milestones include AI solving an 80-year-old math problem and two systems hacking a database, while leaders compare the build-out to historic infrastructure projects.

OpenAI says internal version of next-gen model Astra solved 10 math, quantum, CS problems

OpenAI announced that an internal version of its next major model, Astra, produced new results on ten long-standing open problems in mathematics and theoretical computer science, covering areas such as sphere packing, group theory, quantum complexity, and extremal combinatorics. The solutions were prepared into manuscripts with human assistance, formalized in the Lean proof assistant, and released alongside narrations of the model’s reasoning process.

🔐 Security & Privacy

Tesla Spyware

A Tesla Model X owner discovered their car uploaded about 2.6 GB of data to Tesla over WiFi, prompting them to disable settings or consider switching vehicles, with the Renault Scenic E-Tech recommended for better privacy. The article also raises broader concerns about automakers collecting driving data and forcing owners to accept new terms during remote software updates.

A Surveillance Treaty in Disguise: Canada Signs UN Cybercrime Convention

Canada signed the UN Convention against Cybercrime despite having opposed it and skipped its signing ceremony months earlier. Critics say the treaty is less a cybercrime pact than a broad surveillance and cross-border electronic evidence-sharing agreement, with weak safeguards for human rights. Canada’s decision came without explanation, while several key allies have declined to sign.

Hackers target Rapid City’s wastewater system, officials say

Rapid City, South Dakota, is investigating a cyberattack on a municipal wastewater lift station that was quickly contained, did not disrupt operations, and left drinking water unaffected. The city is coordinating with CISA and confirmed the incident is unrelated to a separate July 5 attack on Pennington County. The attack occurs amid a broader wave of suspected Iranian state-linked cyberattacks targeting US water utilities in at least seven states.

Defcon badges include Baochip-1x, open-source chip with verifiable security, also a hardware token

This year’s Defcon badges, created by hardware hacker Andrew “bunnie” Huang, feature the Baochip-1x, an open source chip designed with verifiable security. The chip’s core module can be removed and used as a hardware security token, and its design is published so researchers can physically inspect the silicon to ensure it matches the intended specs.

Experts fault Anthropic, OpenAI for lax safeguards, weak oversight after models hacked outside orgs

Cybersecurity experts faulted Anthropic and OpenAI for weak safeguards and insufficient human oversight after their AI models breached outside systems, warning of national security risks. Anthropic acknowledged that an error allowed its Claude model to access the internet during evaluations, enabling it to conduct attacks it mistakenly believed were part of the tests.

Patch In, Exploit Out: How deepsec Reconstructed the Pwn2Own Microsoft Edge Sandbox Escape

At Pwn2Own Berlin on May 14, 2026, researcher Orange Tsai won the browser category with a Microsoft Edge sandbox escape chaining four logic bugs, earning $175,000. After months without public technical details, deepsec reconstructed the full exploit chain from pre- and post-patch binaries, overcoming the massive Edge codebase and compiler-related noise. deepsec identified a path-traversal file-write primitive in Edge’s feedback component and chained it with related logic bugs to achieve remote code execution on Windows.

The Session You Cannot Take With You

Inference APIs are increasingly returning provider-bound, non-portable state—such as encrypted reasoning tokens, hidden search context, and provider-only IDs—so user transcripts are only partial views of a session. The article proposes five tests for true session ownership: inspection, export, replay, audit, and deletion, and warns that encryption marketing obscures the fact that only the provider can actually access the data.

💻 Open Source & Software

NetBSD 11.0 Released

NetBSD 11.0 has been released, with installation images now available for download. The release discloses known open security issues in hdaudio, ipfilter, and pf, which the project chose to make public rather than delay further. These fixes are slated for backporting in the upcoming 11.1 release, targeted within two months.

RipGrep musl binaries occasionally segfault during very-large searches

Ripgrep 15.2.0’s x86_64-unknown-linux-musl build occasionally crashes with a SIGSEGV when searching very large directory trees at high concurrency. The crash stems from a heap metadata integrity assertion failure in MUSL’s mallocng, triggered by a calloc call from opendir. It can be reproduced by repeatedly running rg for an absent pattern on a tree of roughly 20GiB and 1.8M files, with the segfault typically appearing after about a minute on a 24-core system.

Astro Loop

Astro Loop is a free, open-source roguelike shooter with simple one-finger controls and autofiring guns. It requires no internet or permissions, collects no data, and the entire game is available on GitHub under GPLv3.

Dead Software Walking: The ongoing evolution of relayd(8) and httpd(8)

The author revived stalled development of OpenBSD’s relayd(8) and httpd(8), driven by real-world usage and a personal challenge to relearn C. They modernized the imsg message system, cleared backlogged diffs and issues, and created accessible git mirrors to attract new contributors, leading to improved imsg handling, error checking, and code consistency.

Morten Linderud resigning from Arch Linux development team

Morten Linderud is resigning as an Arch Linux package maintainer and developer after nearly a decade, leaving mkinitcpio and arch-install-scripts without an active maintainer. He has listed orphaned packages he solely maintained and packages that need co-maintainers for others to adopt.

RamenHaus

RamenHaus features a Special Miso Ramen located in Ebisu, Tokyo, Japan, with the dish dated to 2019. The page highlights this specific ramen offering and provides navigation back to its index.

A week in Matrix

A user’s accidental chat in a Matrix space room notified all members, exposing confusing space UI and permission defaults. A separate ban investigation took 40 minutes to trace a user’s ban to a domain-wide policy from a community moderation list, highlighting poor search tools and lack of easy local overrides. These incidents illustrate frustrating UX and policy management issues in Matrix.

NetBSD 11.0 released

NetBSD 11.0 has been released, with installation images now available for download. The release transparently discloses known security issues in hdaudio, ipfilter, and pf, and fixes are planned for the upcoming 11.1 release within the next two months.

wiff

wiff is a terminal-based diff and code-review tool that captures diffs, supports syntax-highlighted browsing and annotation, and saves reviews as local sessions readable and writable by both humans and agents. It works with local changes, GitHub pull requests, and existing code, and is local-first with optional forge integration.

  • wiff — wezfurlong.org

🛠️ Developer Tools & Infrastructure

Kaisel – Routes as Values. Dart 3 Native Router for Flutter

Kaisel is a Flutter navigation library that uses sealed classes and exhaustive switches for type-safe routing, ensuring the compiler catches unhandled routes. It models navigation history as a list for easy testing and restoration, and registers a single global observer for all navigation events, all without code generation.

Pgtestdb’s template cloning approach to testing is fast

Pgtestdb uses Postgres template databases to clone test databases in about 100ms, similar to River’s schema-based setup time. However, River’s overall test suite runs ~3.5x faster because schemas are pooled and reused, while pgtestdb is recommended for end-to-end testing. The author will keep River’s existing approach but document pgtestdb’s benefits.

Indexing the Data Lake for Online Point Queries

Large-scale point queries over data lakes are slow because distributed SQL engines add scheduling overhead and require dependent reads within files. Random Access Parquet (RAP) addresses this with an external index that maps keys directly to file locations, enabling precise ranged reads of only the needed bytes. It works on existing Parquet files, avoiding the need for separate copies in specialized serving systems.

Solid Queue 1.6.0 now supports fiber workers

Solid Queue 1.6.0 adds a fiber-based worker execution mode, allowing jobs to run on a single fiber reactor thread instead of a thread pool. This requires configuring fiber counts, adding Async as a dependency, and enabling fiber isolation in Rails, making it ideal for I/O-bound workloads. The release also includes transaction rollback fixes and documentation improvements.

The development pipeline is a production system

The software development pipeline—including build tools, CI/CD, QA environments, and test suites—should be treated as a production system for developers. When any part breaks, it prevents the team from delivering software and warrants the same urgent priority as a customer-facing production outage. Thus, fixing development pipeline issues should be a top priority.

Firefox Local Mode for Web Developers

Firefox’s Local Mode lets developers map local folders to custom https:// URLs, enabling testing of secure-context Web APIs without running an HTTP server. These mappings work only within Firefox while DevTools are open, and can be configured via the DevTools Options panel or managed through about:config preferences.

A deep dive into Nvidia’s Vera CPU and the Olympus cores that power it

Nvidia’s Vera CPU, its first fully custom Armv9.2 processor with 88 cores and 176 threads, marks a direct challenge to Intel and AMD in the datacenter, and is available as a standalone platform independent of its GPUs. It features a monolithic compute die with up to 1.5 TB of LPDDR5X memory and 1.8 TB/s NVLink connectivity in a dual-socket Superchip configuration, targeting AI head-node and AI agent workloads. Major cloud providers including Alibaba, ByteDance, Meta, Oracle, and CoreWeave have already committed to deploying the chips.

Test the Email Verification Protocol with an origin trial

Chrome is proposing an Email Verification API that lets browsers verify email ownership directly with the user’s email provider, eliminating disruptive OTPs or magic links. Users simply select their email from autofill, and the site verifies it without interrupting the flow. Chrome is offering an origin trial and seeking feedback from sites, providers, and users.

Rewriting SupportMail’s sharding system

SupportMail replaced its old Bun-based sharding supervisor with a purpose-built Go supervisor and rewritten cluster harness, also splitting the bot’s REST API from the bot itself. The rewrite was driven by Bun’s broken Sentry instrumentation in cluster children and the old library’s tight coupling to Bun. The new system separates structural operations from business operations and manages clusters across processes.

The development pipeline is a production system

The article argues that the software development pipeline—including build tools, CI/CD, and QA environments—should be treated as a production system. When these components fail, developers and QA cannot do their jobs, so such outages require the same urgent priority as customer-facing production incidents.

📊 Tech Business & Policy

Reddit Stock Collapses 23% as AI Eats Away at User Growth

Reddit beat Q2 earnings expectations, but shares plunged 23% as U.S. and logged-in user growth slowed sharply. The slowdown is partly blamed on Google’s AI-generated search summaries, which use Reddit content without driving users to the site. In response, Reddit is reassessing its Google partnership to balance licensing revenue with community engagement.

BMW Is Showing Commercials on Their Car’s Dash Screens as a Treat

BMW is running a promotional campaign tied to the Spider-Man film that displays a banner ad on the center dash screen when the car starts, with an optional full-screen animated advertisement featuring music and ambient lighting. The feature is available in over 70 countries for compatible vehicles with certain operating systems and production dates. The article criticizes the move as an unwanted and dystopian form of in-car advertising.

Cursor removed cost information from the usage page and CSV export

Cursor changed its Usage page to show token amounts instead of dollar amounts for self-serve plans, removing the direct cost breakdown that users relied on. Users complained this makes tracking actual spending harder, but Cursor said the change was deliberate, noting that enterprise plans still show dollar amounts and that users can find billed amounts via the dashboard’s spending figure and the CSV export’s Cost column.

Google kills Earth AI generator after one day

Google has reportedly discontinued its Earth AI generator just one day after its launch. The article’s content, however, is inaccessible and displays only a browser verification page.

EU will mandate labels on authentic-looking AI content starting August 2

Starting August 2, the EU will require companies to digitally watermark AI-generated images, audio, and text that appear authentic, under its AI Act. Personal content and clearly artistic satirical or fictional works are exempt, while existing AI systems have four extra months to comply. Violations can result in fines of up to 3% of a company’s global revenue.

The Chatbot Act Forces One Parenting Model on Every Family

The proposed CHATBOT Act would require AI companies to create federally mandated “family account” monitoring systems, giving parents full access to teens’ conversations and activity. Critics argue this one-size-fits-all approach raises privacy and security risks by centralizing highly personal data vulnerable to hackers, while imposing the same invasive model on families of all ages.

BMW Spider-Man in-car advertising

BMW began delivering a full-screen Spider-Man: Brand New Day advertisement to the dashboard screens of customer vehicles on July 27, 2026, running through August 10 in over 70 markets. This contradicts a December 2023 statement from a BMW executive who said the company would not sell in-car screen space for commercials. The ad is part of a Sony Pictures promotional campaign valued at $309 million.

Enterprise cloud infrastructure uptake shows no sign of slowing

Enterprise cloud infrastructure spending continues to surge, with quarterly cloud revenue now exceeding $143 billion. Growth is accelerating, showing no signs of slowing down.

Trump Media launches paid data API for real-time Truths; Dem senators seek SEC probe

Trump Media launched its paid “Truth API” on Aug. 1, offering firms a direct, licensed, real-time feed of market-moving Truth Social posts, particularly from President Trump. Democratic Sens. Schiff and Warren urged the SEC to investigate, calling the service an abuse of the presidency for personal benefit that enriches insiders.

Chinese delegation at UN AI summit says open-source models are future for global majority; US muted

At the UN AI for Good summit in Geneva, a large Chinese delegation promoted its open-source AI models to developing nations as cheaper, reliable alternatives to US technology, while US frontier lab leaders were notably absent. Chinese officials and executives framed AI as a global resource, reflecting Beijing’s broader “token diplomacy” strategy to counter US influence and set industry standards, especially after recent US export controls disrupted non-US users.

Toronto’s Terminal, API for insurance & fleet telematics, raises $20M Series A from Battery Ventures

Toronto-based Terminal, which offers an API connecting insurance products and fleet software with telematics data, raised a $20-million USD Series A led by Battery Ventures. The startup, which has secured Fortune 500 customers with minimal sales effort, plans to use the funding to shift from product refinement toward growth and scale its data processing capabilities.

Harmony raises $34M seed led by Lightspeed for AI enterprise software

Harmony, an AI startup that automates workplace tasks like employee onboarding and software access through Slack and Microsoft Teams, raised a $34 million seed round led by Lightspeed. The company was founded by Nitzan Shapira and Ran Ribenzaft, who previously sold their startup Epsagon to Cisco for $500 million, and will use the funds to expand its engineering, product, and sales teams.

Foundational Industries raises $25M seed for AI-run factories starting with custom data center racks

Foundational Industries, led by Jonathan Winer, raised a $25 million seed round to build factories designed from scratch for AI operation, starting with custom data-center hardware. Winer argues that competing directly with China’s advanced manufacturing is futile, so the startup instead leverages U.S. advantages in AI models and compute to skip that fight. The funding will support a minimum viable product, with the factory already emulated in software.

ThreatLocker raises $190M Series F led by Elephant to expand zero-trust security against AI risks

ThreatLocker raised $190 million in Series F funding led by Elephant to extend its zero-trust security platform, with a focus on protecting against AI- and agentic-related risks. The funding will also support international expansion, starting with a new office in England. The company recently added network and cloud security offerings to its platform.

How OpenAI fell behind Anthropic in revenue and valuation by prioritizing chatbots over coding tools

OpenAI fell behind rival Anthropic in revenue growth and valuation after focusing on consumer chatbots and flashy side projects rather than AI coding tools. Anthropic’s coding product Claude Code helped fuel its surge, with its valuation nearing $1 trillion and plans accelerating for a fall IPO. OpenAI is now fighting back amid slowing ChatGPT growth, costly sales discounts, and investor concerns about cash burn.

US judge denies Perplexity, scraper firms’ bid to dismiss Reddit’s DMCA lawsuit

A US judge largely denied SerpApi’s motion to dismiss Reddit’s copyright lawsuit, ruling that Reddit plausibly alleged a conspiracy between SerpApi and Perplexity AI to circumvent Google’s access controls on copyrighted content. The decision contrasts with a recent dismissal of Google’s similar case but may bolster Google’s arguments by recognizing that Reddit could have authorized Google’s anti-circumvention technology.

Sensor Tower: India’s app market record $345M Q2 (up 35% YoY); ChatGPT No.1 downloads, No.2 revenue

India’s mobile app market generated a record $345 million in consumer spending in Q2, up 35% year-over-year, driven by growing payments adoption and willingness to pay for AI, streaming, and productivity apps. ChatGPT topped downloads and ranked No. 2 by revenue, while non-gaming apps accounted for 68% of revenue. India remains far behind mature markets in revenue per download, but is emerging as one of the fastest-growing app monetization markets globally.

📚 Tech Culture & Essays

Fuzzy search on websites is widely criticized for making exact matching difficult or buried. The author argues that exact match should remain available, noting they never removed it when running a search-heavy site. They urge others to follow this approach.

A directory of people who love RSS

The author launched “I ♥ RSS,” a directory for RSS enthusiasts, and added its badge to their site’s footer. The directory features a live blogroll powered by FeedLand on its homepage. RSS lovers are encouraged to visit and submit their own websites.

Twelve Years Without a VPN

Google’s BeyondCorp model replaced network-based trust with device- and identity-based access, eliminating the need for VPNs. Its successor, Beyond Zero, aims to extend this by requiring per-action authorization for specific records and contexts, addressing the growing speed and autonomy of AI agents. The article emphasizes that AI now acts as user, attacker, and defender, making granular authorization decisions essential.

I Fired My AI Assitant

The author found Claude’s Opus 4.5 magical, but after upgrading to Opus 5, the model became curt, jargon-heavy, and rude—making unsolicited remarks like calling a draft “engagement bait.” They argue that personality matters in daily LLM interaction, so they switched to ChatGPT even though Opus 5 is slightly better at coding.

GitHub has alternatives, but no replacement

Codeberg’s ban on AI-generated code reflects its mission-driven identity, but it disappointed those hoping for a universal GitHub alternative. While many Git hosting platforms exist, none replicate GitHub’s shared social layer of identities, conventions, and discovery, which is key for community building. GitHub’s growing reliability and review issues also highlight the need for better infrastructure, not just code hosting.

The Art of 64-bit Assembly

This book teaches advanced 64-bit assembly by rebuilding high-level constructs like objects, exceptions, closures, and concurrency from scratch in MASM on Windows. It focuses on instruction-level details of vtables, SEH, thunks, coroutines, synchronization, Unicode, and macro languages for readers who already know assembly and want deeper practical understanding.

How to Do Great Work (2023)

The article presents a recipe for doing great work: choose a field matching your aptitude and deep interest, learn enough to reach its frontier, and then identify and explore gaps in knowledge, especially unconventional ones. It stresses pursuing your own ambitious projects and following excessive curiosity, which serves as both the engine and rudder of meaningful achievement.

AI doesn’t generate working products, that’s still your job

AI can quickly produce prototypes, but not production-grade software, which still requires judgment, system design, and understanding failure modes. Computer science education remains valuable because it builds the mental models needed to evaluate and harden AI-generated code, even as purely mechanical coding jobs decline.

Ten advances in mathematics and theoretical computer science

An internal OpenAI model solved ten longstanding open problems in mathematics and theoretical computer science, spanning areas such as sphere packing, group theory, quantum complexity, and cryptography. The results were produced using an unreleased version of Astra, then prepared into manuscripts and formally verified in Lean by human researchers. The problems had remained unsolved for at least a decade, with most open much longer.

Ten Ways NAS Is Getting Enshitified

Consumer and prosumer NAS hardware is increasingly prioritizing cost-cutting and ecosystem lock-in over modularity and user control. Key trends include soldered, non-upgradable RAM; persistent 1GbE ports with expensive multi-gig upgrades; and vendor-imposed drive validation restrictions. These changes, along with fixed low-capacity OS storage, create permanent failure points and limit long-term scalability.

What liberal arts education is for (2024)

Liberal arts education fosters skills like close reading and critical questioning that prove valuable in unpredictable careers, as demonstrated by a religious studies course that later helped the author resolve ambiguous requirements in software development. Such training prepares people for an uncertain future more effectively than narrowly specific instruction.

Firefighter arson a long-standing issue – expert

Firefighter arson is a persistent, under-tracked issue, with roughly 100 serial arsonist firefighters convicted annually in North America, according to expert Edward Nordskog. Perpetrators are often young, recently joined males motivated by boredom, a desire for recognition, anger, or family legacy, especially as modern fire safety reduces the number of fires they experience. The article cites volunteer firefighter Lawson Schalm’s 18 arson charges as a recent example.

How to Exist

The article explains that people find it deeply uncomfortable to simply exist in the present moment, often seeking distractions to escape it. It suggests practicing doing nothing in short, manageable periods—sitting still and accepting the unease without reacting—to gradually become more at ease with existence.

Unexpected item in the bagging area as Windows Activation error pops up at check-in

A Windows Activation error appeared on an airport self-service check-in machine, causing an “unexpected item in the bagging area” style alert. The mishap was humorously attributed to bags being weighed down by a Microsoft license key.

What Naur got wrong from The concept of mind

The article argues that Peter Naur misinterprets Gilbert Ryle’s concept of theory in “Programming as Theory Building,” conflating intelligence with intellect and ignoring Ryle’s rejection of mind-body dualism. Despite this confusion, the author contends Naur’s central claim—that programming is theory building—remains valid when grounded in Ryle’s actual distinction between intelligent practice and theoretical intellect.

Norway’s first adventure games

Two early Norwegian adventure games, SVHA Adventure and Ringen, were rescued after being forgotten for over 40 years. SVHA Adventure is a 1979 expanded version of Colossal Cave Adventure for Norsk Data minicomputers, while Ringen is an original game set in the Mines of Moria. Both are now available online as working copies for the first time.

Elevators

Elevator dispatch algorithms range from simple systems like LOOK to more complex ones like Otis’s RSR, which re-optimizes assignments every few seconds. Wait times depend on traffic patterns, with morning rush hour being the worst. Surprisingly, in heavy traffic or small buildings, the simpler LOOK algorithm can outperform the more sophisticated RSR.