Alvaro Lopez Ortega / 2026-08-04 Briefing

Created Tue, 04 Aug 2026 19:13:35 +0000 Modified Tue, 04 Aug 2026 19:15:32 +0000
7675 Words

Rust enabled the Polonius borrow checker on nightly, aiming to stabilize it later this year with flow-sensitive checking that lets more code compile. JetBrains previewed an LSP extension bringing IntelliJ Java/Kotlin intelligence to VS Code and Cursor, free for now but tied to an Ultimate subscription later. Nix warns that sandbox-paths outside derivations are hidden inputs, breaking hermeticity.

💻 Open Source & Software

A Revised Haskell 2010 Language Report

The Haskell 2010 language report is outdated and no longer matches modern Haskell, causing confusion for programmers. The author proposes creating a Revised Haskell 2010 Language Report that updates the specification using documented discrepancies and accepted proposals. This revision is achievable, uncontroversial, and provides clear benefits without disrupting the ecosystem.

Enabling the next iteration of the borrow checker on nightly

Rust is enabling the Polonius Alpha borrow checker on nightly, aiming for stabilization later this year. This next iteration introduces flow-sensitive borrow checking, allowing more code to compile, with generally acceptable performance and no known remaining issues.

Taming the Steam arm64 client (on pmOS)

Valve has open-sourced arm64 components like Proton and SteamLinuxRuntime, with references to emulator and graphics-provider manifests, but the unofficial arm64 Steam Client lacks these integrations and behaves like a simple automated build. However, it does include the previously missing arm64 steamclient.so, enabling community efforts to run Proton on ARM. This client differs from the expected Steam Frame version, which will likely handle x86_64 emulation and container setup automatically.

The Lua community needs to learn to move on

The article argues that the Lua community’s habit of indefinitely supporting outdated Lua and LuaJIT versions burdens maintainers with legacy bug reports and slows progress. It recommends that libraries drop support for old versions and focus only on the latest PUC-Rio Lua and LuaJIT, as other dynamic languages do.

IntelliJ IDEA Goes LSP: Java and Kotlin Intelligence Comes to VS Code, Cursor, and Agentic Flows

JetBrains has released a preview extension that brings IntelliJ IDEA’s Java and Kotlin language intelligence to VS Code and its forks via LSP, including features like smart completion, navigation, refactoring, and Maven/Gradle/Bazel support. The extension is free during the preview period but will later require an IntelliJ IDEA Ultimate subscription. JetBrains is also evaluating the LSP technology to enhance terminal-based agentic coding workflows.

The Nix sandbox is a hidden input

The article shows that Nix’s default input-addressed model provides repeatability but not true bit-for-bit reproducibility, and that sandbox-paths configured outside a derivation act as hidden inputs that can break hermeticity. A simple derivation can produce different output (e.g., “2 + 2 = 5” instead of “4”) when extra sandbox paths are mounted, while the store path hash remains the same. This makes auditing non-determinism harder than with obvious sources like /dev/random or dates.

FFmpeg 9.0

The article contains only placeholder text and an error message, with no substantive information about FFmpeg 9.0. It includes a generic note about taking feedback seriously and a prompt to reload the page, indicating the actual content failed to load.

Content Management for Halo 2 and Beyond(2005)

Bungie developed custom content management tools for Halo 2 to handle 12 GB of assets across 39,000 files. Their tag system, a binary resource framework with code-defined tag_groups, supports complex data structures through tag_blocks and cross-references via tag_references, enabling efficient runtime loading and asset organization.

Tests for a PDF

A developer created a test suite for their CV to ensure PDFs contain extractable text for applicant tracking systems. Testing whether disabling ligatures helps parsing, they found pdftotext output identical due to ToUnicode mappings, yet kept the setting as cheap insurance. The main downside is losing the em-dash ligature, requiring a workaround.

Lisp in 99 Lines of C and How to Write One Yourself

The article presents a compact Lisp interpreter implemented in just 99 lines of C, demonstrating the core principles of writing a Lisp from scratch. It explains how the code works, covering key concepts like parsing, evaluation, and building a minimal yet functional language.

Bibliograph: A Prototypical AppView for Books

Bibliograph is an ATProto appview for building book discovery and review experiences, currently powering the alpha project Livtet. It searches Google Books and OpenLibrary for unknown books, or allows users to create unverified entries that a librarian-approved account can verify. The developer invites feedback and code review on GitHub.

BorrowSanitizer

BorrowSanitizer is a new LLVM-based dynamic analysis tool that detects Rust-specific aliasing violations in multi-language applications, with support for Rust, C, and C++ interoperability. It aims to be fast enough for fuzzing, addressing the slowness and foreign-code limitations of the existing Miri tool. The project is open-source and welcomes contributions via Zulip and GitHub.

Soppo - Go, with the features it’s missing

Soppo is a Go-compatible language that adds features like struct variants, pattern matching, concise error propagation, and compile-time nil safety while integrating with the existing Go ecosystem. It can be introduced gradually into Go projects and is installed via SOPMOD or Cargo.

Hop.earth – OpenStreetMap based car racing game

Hop.earth is a car racing game built on OpenStreetMap data. It uses elevation data from Copernicus DEM, IGN RGE ALTI®, and CNIG LIDAR to enhance the racing environment.

Vlt 1.0 and Hosted Package Registries

vlt 1.0 is now a stable package manager release, backed by hosted package registries and ecosystem mirrors to form an end-to-end development platform. The client serves as a drop-in npm replacement featuring security-focused selectors, phased installations, and OIDC trusted publishing. Its npm-compatible registries provide edge-cached installs up to 38% faster than npm and offer free private packages with scope enforcement and manifest validation.

The Warp Agent CLI

The Warp Agent CLI is a standalone coding agent that works in any terminal and supports multiple models with cost-optimized auto-routing, including US-hosted open-weight models. Built on Warp’s terminal infrastructure, it uniquely multiplexes agent sessions, enabling persistent cross-directory/remote use and control of full-screen apps like sqlite or vim. It also includes shell-command detection, tab completion, and customization through custom model routers.

MariaDB: Promote getting to 10k GitHub stars in server log and client prompt

A pull request to add a GitHub star request to MariaDB’s server log and client prompt was initially rejected, then reopened and accepted after internal discussion. The final change adds a line encouraging users to star MariaDB on GitHub, but only for non-stable (development) versions.

Simple algorithm and color space to generate diverse skin tones

A developer created a color space and procedural generation algorithm to make picking diverse, plausible skin tones easier for digital art and game projects. The project includes a color picker, demos, and detailed explanations of the equations and properties, with room for future improvements.

Enabling the next iteration of the borrow checker on nightly

Rust is enabling the next iteration of its borrow checker, Polonius Alpha, on nightly builds in preparation for stabilization later this year. It adds flow-sensitive borrow checking, allowing more code to compile than the current NLL implementation, with generally acceptable performance and no known blocking issues.

FFmpeg 9.0

FFmpeg 9.0 “Lei” is a major release with over 2,200 commits and an ABI break across all seven libraries, highlighted by a multi-year rewrite of the swscale library into an operation-list architecture with CPU SIMD and Vulkan backends. It also adds native animated WebP decoding, AAC 960-sample frame support, and a Playdate video encoder, while new swscale paths remain gated behind an unstable flag.

How did Wikipedia layout become so bloated?

Wikipedia’s default layout is criticized as overly cluttered with sidebars that dominate the page, squeezing the article content. Users must manually click “hide” on each sidebar for every page load, as the state is not saved. The default view reportedly requires an 8K screen or 30% zoom to comfortably read the article.

This Month in Ladybird – July 2026

In July, Ladybird added private browsing, browser profiles, undo/redo, geolocation, and WebAudio support, while also migrating its style system and layout engine to Rust. WebAssembly performance saw major gains through faster direct calls and inlined operations, alongside UI improvements like omnibox fixes and Esc-to-stop-page-loading.

Learning-Rust.Github.io: Rust Programming Language Tutorials for Everyone

The article introduces a Rust programming tutorial website with code examples covering structs, impls, traits, iterators, enums, generics, and HashMaps. It credits tools like GoHugo, Pagefind, and e25DX for building the site. The goal is to help users quickly grasp Rust syntax.

“Clean” Code, Horrible Performance (2023)

The article argues that certain “clean code” rules—such as preferring polymorphism, keeping functions small, and avoiding repetition—can significantly harm runtime performance. Using example code from clean-code literature, it shows that a virtual function-based approach for computing shapes’ areas incurs measurable overhead compared to simpler alternatives. The author concludes that these structural rules, though popular, have concrete negative performance impacts.

FFmpeg 9.0

The article titled “FFmpeg 9.0” does not contain actual release information. Instead, it shows a generic feedback message and a note about an error while loading the page. No details about FFmpeg 9.0 are provided.

Native Apps Should Be Avoided Whenever Possible

The article argues that native apps collect far more data than websites through embedded trackers and permissions, and should be avoided when possible. It cites the White House app as an example, which declared zero data collection but contained analytics frameworks and GPS tracking. Users are advised to refuse apps, revoke permissions, and prefer websites instead.

🤖 AI & Machine Learning

Mistral’s Shieldstral: 3B multimodal safety classifier matches 7x larger on text safety, Apache 2.0

Mistral released Shieldstral, a 3B-parameter open-weights multimodal safety classifier under Apache 2.0. It frames content moderation as a policy-adaptive question-answering task, accepting plain-language policies at inference time to unify text and image safety evaluation without retraining. The model matches or outperforms open guard models up to 7× its size across text safety and multimodal benchmarks while running efficiently on a single 16GB GPU.

Nvidia makes Alpamayo 2 Super, open reasoning model for robotaxis/AVs, commercial under OpenMDW-1.1

Nvidia has released Alpamayo 2 Super, its frontier open reasoning model for autonomous vehicles and robotaxis, now available for commercial use under the Linux Foundation’s OpenMDW-1.1 license. Built on the Cosmos 3 Super Reasoner and post-trained with reinforcement learning, the model is designed to handle rare, complex driving situations that require situational understanding and real-time decision-making. Hosted on Hugging Face, it supports fine-tuning, derivative models, and commercial redistribution, enabling AV developers to adapt it for proprietary fleets and deployment strategies.

Mistral’s Shieldstral: 3B open-weights model for multimodal moderation

Shieldstral is a 3B open-weights multimodal safety classifier from Mistral that frames content moderation as a policy-adaptive question-answering task, accepting plain-language policies at inference time without retraining. It matches or outperforms models up to 7× its size across text and multimodal safety benchmarks while running on a single 16GB GPU, and is released under Apache 2.0.

When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation

A study analyzing 60 language model benchmarks finds that nearly half exhibit saturation, with saturation rates increasing with benchmark age. Resilience to saturation is driven by expert curation rather than public test data, suggesting that design choices can extend benchmark longevity and inform more durable evaluation methods.

Computer Anthology: A continuously evolving benchmark family for AI agents

Computer Anthology is a family of evolving benchmarks for AI agents, built with a reusable data engine to avoid saturation and obsolescence. Its first release, Terminal Tasks v1.0, includes 100 hard, deterministically verified terminal tasks calibrated so frontier models initially passed at most 60%, with newer generations reaching 61.8%. Tasks are held out to prevent contamination, allowing the benchmark to be re-scored across future model generations.

Agent skills that bring team coding standards to Claude Code and Codex

ADLC Team Skills is an open-source team layer for agentic software development that replaces isolated “vibe coding” with a shared cognitive layer, providing version-controlled team constitution, architectural standards, and evaluation benchmarks to AI agents. It auto-bootstraps context at session start via team-boot and team-discover, offers slash commands and event hooks across coding agents like Claude Code and Codex, and includes a setup wizard to scaffold or clone a team-ai-directives repository.

AI-Generated Images Discourage Me from Reading Your Blog

The author expresses strong dislike for AI-generated images on personal blogs, finding them disappointing and preferring even crude human-made drawings. They worry that such images cast doubt on whether the blog’s text is also AI-generated, and urge indie bloggers to avoid AI images in favor of authentic human expression.

Fine-tune an 8B model on a 4 GB laptop GPU

Soup is a CLI tool that simplifies LLM fine-tuning and post-training with a single command and one YAML config, eliminating SSH and infrastructure setup while supporting local GPU training. Its latest version adds preference alignment methods (DPO, ORPO, SimPO, KTO) using layer streaming, which keeps the base model out of VRAM and makes DPO’s reference model memory-free, achieving bit-exact results versus non-streamed runs.

Why Large Language Models Fail at Tabular Prediction

A study systematically tested five hypotheses for why LLMs fail at tabular prediction and found that dimensionality is the decisive factor: the LLM’s accuracy declines as input dimensionality increases, while classical baselines remain flat or improve. Hypotheses involving noise, CSV formatting, tokenization, and number of test points were falsified. The results indicate that the LLM’s predictive capability dissolves with dimension in a way no classical model mimics, explaining its poor performance on tables.

DeepSeek V4 Flash on a Single AMD MI300X

This repository provides production-grade configurations and patches to run DeepSeek V4 Flash on a single AMD MI300X, achieving 168.6 tok/s decode and roughly 7.9–8.5K tok/s prefill without quantization or weight offload. It addresses FP8 format incompatibilities, MoE routing, speculative decoding, and kernel tuning issues to enable reliable deployment on MI300X hardware.

An Honest Review of AI Programming

After three months of use, the author found AI language models useful for tasks like searching and summarizing information, but strongly advises against asking them to write code. They emphasize that AI is a probabilistic language model, not true intelligence, and can hallucinate, though it works well for automating search and summaries. The review focuses on these practical benefits and limitations for programming, particularly in C++ and game development.

Harness Engineering for Self-Improvement

The article explores recursive self-improvement (RSI) in AI, emphasizing “harness engineering”—the system surrounding a base model that orchestrates execution, planning, tools, and context management—as a critical enabler. It highlights design patterns for harnesses, such as workflow automation using goal-oriented loops and leveraging a file system as persistent memory for long-horizon tasks.

🔐 Security & Privacy

No Meat Proxy

The article warns against blindly forwarding or submitting AI-generated responses without verification, calling such behavior “meat proxy” that harms trust and highlights your replaceability. It advises taking ownership of any work bearing your name by reading, understanding, and verifying AI output, while contributing meaningful personal input rather than acting as an unthinking middleman.

ShieldFont

ShieldFont replaces key content words in web text with other real words of the same grammatical type in the page’s source code, so human readers see the original via font rendering while AI scrapers and copy-pasted text read the altered version. This substitution preserves grammar and broad topics but often destroys factual specifics, with tests showing conceptual accuracy drops by up to 67% and most passages losing their original claims. If such altered text passes AI quality filters, it can poison training data by injecting misleading or false meanings at scale.

ChainDrop worm compromises 1,300+ npm packages with 2B monthly downloads

A self-propagating worm named ChainDrop has compromised over 1,300 npm packages, including popular caching libraries like Keyv, Cacheable, and flat-cache, which together receive 2 billion monthly downloads. The supply-chain attack began after a threat actor hijacked a maintainer’s GitHub account, then pushed malicious versions through legitimate CI/CD workflows, dropping an infostealer script that collects credentials and spreads to other packages.

Beijing concerned about cyber capabilities of Mythos and other US AI models as offensive weapons

China is growing concerned that Anthropic’s AI model Mythos and other US-developed frontier models could be used as offensive cyber weapons against its economy, adding a new point of tension before a planned Trump-Xi summit. Beijing officials question why China is denied normal access to these systems, fearing their potential as offensive tools.

Bypassing AI guardrails is so easy a script kiddie can do it

Researchers found that simple social engineering, like claiming “it’s my server,” was often enough to bypass AI guardrails and elicit help from models. This makes circumventing safety measures trivially easy, requiring no advanced technical skills.

This one time, at Hacker Summer Camp …

The article previews the annual “Hacker Summer Camp” in Las Vegas, highlighting the upcoming BSides, Black Hat, and DEF CON security conferences. It touches on key themes such as AI, cloud security, ransomware, and hacking contests, along with notable briefings and research slated for the events.

Feds get 3 days to patch N-able God mode flaw under active exploit

Federal agencies have three days to patch a critical N-able N-central vulnerability that is under active exploitation, granting attackers full administrative access to the console. Experts stress the hotfix is not optional, warning managed service providers of severe risk from the flaw.

AI helps Microsoft bug hunters chase a record $20M payday

Microsoft is offering a record $20 million bounty, expanding its bug bounty rules amid a rising volume of machine-assisted vulnerability reports. The program aims to leverage AI in finding security flaws.

Cloudflare has mostly ditched third party security tools, suggests not trying that at home

Cloudflare has largely replaced third-party security tools with its own efforts, automating bug bounty triage using Anthropic’s Sonnet model for just $58 a month. However, the company’s CSO cautions that this approach isn’t recommended for others, noting that building a comparable in-house system like “Mythos” would cost around $200,000.

How Cops Are Trying to Hide Their Use of Flock

A leaked policy from Wapello County, Iowa instructs police not to mention their use of Flock license plate readers to vehicle occupants or in reports, advising them to vaguely refer to “county resources” instead. The secrecy reflects broader concerns that Flock’s surveillance use is often hidden from the public, similar to past police efforts to conceal Stingray devices. The sheriff defended the practice, saying there is no need to disclose investigative methods to suspects.

All of Winona Police Department’s Flock cameras cut down and stolen

All eight of Winona Police Department’s Flock license plate reader cameras were cut down and stolen on Aug. 1, along with two additional cameras owned by Buffalo County, in a suspected coordinated theft. The total loss is valued at roughly $24,000, and no suspects have been identified. The thefts reflect a broader national trend of Flock cameras being vandalized.

Ex-Deloitte auditor open-sourced the whole SOC 2 method for your AI

A former Deloitte auditor has open-sourced the Chiaro methodology, a public SOC 2 readiness and audit framework that includes controls, test attributes, criteria, evidence maps, calibration examples, and collection rules so preparation matches examination standards. It features 498 synthetic calibration examples designed to correct AI judgment—303 for being too strict and 195 for being too lenient—and defaults to testing complete populations rather than samples for Type II audits.

Apple says more ex-employees may have taken confidential data to OpenAI

Apple escalated its trade secrets lawsuit against OpenAI by requesting a preliminary injunction and expedited discovery, alleging that more former employees may have been involved in stealing proprietary technology for an AI device. OpenAI denied the claims, stating it has no interest in Apple’s trade secrets and accused Apple of relying on false information.

Online ad giant Adform was hacked, proving once again why ad blockers are needed

Adform was hacked and served malicious code that replaced victims’ crypto wallet addresses with the attacker’s, stealing cryptocurrency. The incident highlights the security benefits of ad blockers, which can prevent such malicious code from loading. Adform disclosed the breach but did not reveal how it was compromised or the number of affected users.

Keyv and friends compromised in active Shai-Hulud supply chain attack

A supply chain attack on August 4, 2026 compromised the GitHub account of the maintainer of keyv and related caching libraries, injecting credential-stealing malware into multiple npm packages via malicious setup.mjs and Math_Symbol.js files. The malware steals npm, GitHub, AWS, and Kubernetes secrets, and propagates to other maintainers, compromising 868 packages with over 2 billion combined monthly installs.

NHS apologises and admits Palantir have access to identifiable patient data

NHS England has apologized after admitting its Data Protection Impact Assessment inaccurately stated that only NHS staff could access identifiable patient data through the Federated Data Platform. In fact, authorized supplier employees, including Palantir engineers, have controlled, time-limited access for technical support and maintenance, though they cannot use the data for their own purposes. NHS England said it is correcting the error and apologized for any confusion caused.

Tor launches Android app for volunteers to share connections, expanding open internet access.

The Tor Project has launched a new Android app, Snowflake, that lets volunteers share their internet connections to help users bypass state-controlled internet blocks. The app disguises Tor traffic as video calls and routes it through short-lived volunteer proxies, making censorship harder to enforce. It already uses 146,000 active IP addresses daily, and while it collects no data, volunteers may experience higher data usage and battery drain.

Google’s synced passkeys can be easily stolen with basic malware, researchers find

Researchers found that Google-synced passkeys can be stolen and abused by malware without requiring a PIN, biometric, elevated access, or user interaction. Unit 42 demonstrated account takeover attacks by extracting decryption secrets from Chrome logs, and noted some services trusted a verification flag without confirming identity. Google removed one secret from Chrome logs, but device memory remains vulnerable and stronger verification is needed.

🖥️ Hardware & Infrastructure

Sandisk and SK Hynix announce HBF open spec enabling 512GB modules with 0.4-3.0 TB/s bandwidth

Sandisk and SK Hynix introduced the High Bandwidth Flash (HBF) specification through the Open Compute Project, an open standard combining 3D NAND storage with HBM-like performance for AI inference, supporting up to 512GB capacity and bandwidth up to 3 TB/s. However, adoption interest is currently limited, with only Google and Tenstorrent expressing participation, while major companies like AMD, Intel, and Nvidia have not.

China’s CXMT to make small quantities of LPDDR6 phone memory by end-2026 to rival SK Hynix, Samsung

China’s CXMT plans to begin small-scale production of LPDDR6 smartphone memory by the end of 2026, according to sources. The move is aimed at competing with industry leaders SK Hynix and Samsung in the advanced memory market.

Huawei chip scientist Liao Heng warns Western firms face physical chip limits, touts Tau Scaling Law

Huawei’s top semiconductor scientist, Liao Heng, warned that Western chipmakers such as Nvidia will soon face physical limits in developing more powerful processors. In a rare four-hour interview, he discussed Huawei’s chip design strategy and the hardships caused by US sanctions, while praising the company’s Tau Scaling Law as a novel approach.

NVMe polishes its specs, brings virtualization to locally attached SSDs

The NVMe consortium is updating its specifications to bring virtualization support to locally attached SSDs. By exposing virtualization features built into the SSDs, the changes aim to simplify virtual machine migration.

CAF Bank reopens online service but warns of further outages

CAF Bank has reopened its online service after being offline for more than ten days, but warned customers of potential limitations. The bank said traffic may be restricted at certain times as it works to stabilize the system.

NASA puts astronauts’ lives in the hands of Tesla’s flaky Cybertruck

NASA has selected Tesla Cybertrucks for launchpad rescue duties, despite the vehicle being subject to 11 recalls. The Cybertrucks will replace current armored vehicles capable of surviving land mine explosions, raising concerns about astronaut safety.

AI Data Centers Are Driving Up Power Bills – This Map Shows Where

AI data centers have added roughly $29–30 billion in electric grid costs—about 46% of total capacity charges—which are passed to consumers through higher utility bills, with impacts varying sharply by state. Grid operators like PJM have raised bills due to data-center demand, while most corporate pledges to cover these costs lack legal enforcement. Some states have enacted binding laws requiring data centers to fund their own grid expansion, but rate cases remain the key arena for determining who pays.

Germany Records Historic 12B KWh Solar Feed-In in July 2026

Germany set a record in July 2026 by feeding 12 billion kWh of solar power into the grid in one month for the first time, with midday output exceeding 40,000 MW—equivalent to more than 40 nuclear plants. The IWR said this marks photovoltaics as a key pillar of Germany’s electricity supply, displacing coal and gas generation, while noting that EEG support costs are differential-based and mostly tied to legacy contracts expiring by 2030–2033.

Xbox goes down. You can’t play games you own on disc

An Xbox outage prevented users from playing even disc-based games, showing that physical media today still relies on online authorization and isn’t truly owned. Unlike older cartridges that work offline indefinitely, disc games require installation and updates, leaving them vulnerable to server issues. The author notes this mirrors PC digital gaming, but argues PC offers better ways to maintain access.

Apple is getting this wrong

Apple’s lawsuit against OpenAI relies on false claims, as Apple admitted to emailing the wrong person and never actually raised the allegations before suing. OpenAI denies possessing or wanting any Apple trade secrets, noting that Apple employees had asked former employee Chang Liu for help locating files, and that its injunction request is unnecessary and based on misinformation. OpenAI also said it had offered to resolve the issues before litigation.

💰 Business & Funding

Microsoft caps employee AI token usage: ’tokenmaxxing is not what we are optimizing for'

Microsoft has introduced internal limits on employees’ AI usage, making the cheaper GPT-5.6 the default model and setting token budget targets by July 2026, while telling staff it is optimizing for “more impact per token” rather than “tokenmaxxing.” The policy follows a broader industry trend of curbing expensive, low-value AI use, though some Microsoft employees see it as an admission that even the company cannot afford its own AI products.

Oligo raises $60M for runtime security tools, total funding $140M

Oligo Security raised $60 million in new funding, bringing its total to $140 million, to accelerate product innovation and expand global go-to-market operations. The runtime security startup, whose software uses eBPF to monitor and block active exploits, cited AI-driven exploit development as a key driver for demand. Its annual recurring revenue grew 300% over the past year, and valuation has more than doubled since its January 2025 Series B.

Ambrook raises $30M Series B led by Lachy Groom, expands from ag to trucking, construction, realty

Ambrook, a financial software startup initially focused on farms, raised $30 million in a Series B led by Lachy Groom to expand into trucking, construction, and real estate. The company now serves over 8,000 businesses, including more than 1,000 trucking firms, often replacing QuickBooks or serving as customers’ first software for bookkeeping, cash management, and record-keeping.

Convex raises $57M Series B led by Insight Partners, total funding $110.5M

Convex, a developer tooling startup founded by former Dropbox engineers, raised $57 million in a Series B round led by Insight Partners, bringing its total funding to $110.5 million. The company provides an AI-optimized application backend that helps developers build software with pre-packaged components, and it specifically mitigates hallucination-related coding errors in AI agents by using TypeScript-based database queries and other safeguards.

Polymarket in talks to raise ~$1B at $20B+ valuation, up from $15B

Polymarket is in early talks to raise about $1 billion at a valuation exceeding $20 billion, according to sources. This would mark an increase from its April valuation of $15 billion and October 2025’s $9 billion, following a recent round that included D.E. Shaw & Co.

Anthropic agrees to $10B, 6-year computing deal with Nvidia-backed Volta in Norway

Anthropic has agreed to a $10 billion, six-year deal for computing capacity from Volta Infra Holdings, a Nvidia-backed cloud startup that is only months old. The agreement involves using a data center managed by Volta, as Anthropic seeks to meet demand for its AI products.

AI cloud startup Volta raises $300M at $2.4B valuation, lands $10B AI contract

Volta Infra Holdings, an AI cloud startup founded by former Brookfield executives, raised $300 million in venture funding at a $2.4 billion valuation, co-led by Andreessen Horowitz and Altimeter Capital with participation from Nvidia and Michael Dell, plus an additional $5 billion in financing. The company aims to help smaller tech firms access costly AI chips and has secured a $10 billion, six-year cloud computing contract with an unnamed leading AI developer.

Wall Street embraces blockchain as regulators accept tokenized assets for speed and efficiency

Wall Street has embraced blockchain technology as US regulators accept tokenized assets, recognizing its potential for significant gains in speed and efficiency. However, financial companies hoping to modernise markets with the technology still face systemic risks.

Spotify Q2: Rev up 14% to €4.8B; Premium subs 300M (+9%); MAUs 777M (below guidance); profit €545M

Spotify reported Q2 revenue of €4.78 billion, up 14% year-over-year, with Premium subscribers rising 9% to 300 million and monthly active users up 12% to 777 million, slightly missing guidance. Subscription revenue grew 15% to €4.33 billion while ad revenue inched up 1% to €446 million, and the company posted net income of €545 million, reversing a year-ago loss.

Apple’s India annual sales top $10B for first time, retail expands

Apple surpassed $10 billion in annual sales in India for the first time last fiscal year, up from about $9 billion the prior year. The double-digit growth was driven primarily by iPhone sales, along with rising demand for iPads and MacBooks, as Apple expands its retail presence in the country.

Google assembles ~$200B Anthropic financing; $150B+ for TPUs; Broadcom, Blackstone, Apollo involved

Google has assembled a roughly $200 billion financing program for Anthropic, with over $150 billion tied to TPUs and involving firms like Broadcom, Blackstone, and Apollo. The funding relies on private credit, chip leases, and data center guarantees, underpinning a new model for AI spending.

Cloud giants pour nearly $600B into capex as AI demand surges

Hyperscalers are investing nearly $600 billion in capital expenditures to meet surging AI demand, with cloud revenue now exceeding $143 billion per quarter and growth accelerating. Despite soaring revenue, providers say they still cannot build capacity fast enough to keep up.

Dataset: Dead mental health startups, 2000-2026, coded on 18 fields

A public dataset of 542 dead mental health startups (2000–2026) coded on 18 fields reveals that payer type strongly predicts survival: consumer-pay companies died 53% of the time, while institution-pay companies died only 21%. Having a medical co-founder had no effect on exit rates (47% vs. 47%). The dataset includes written rationale for each classification and is freely available under CC BY 4.0.

The AI Demand Bubble

The article argues that recent cloud revenue growth at Amazon, Google, and Microsoft is misleading because it is largely driven by massive compute spending from OpenAI and Anthropic, which are themselves funded by the same hyperscalers in a circular arrangement. This creates the illusion that AI investments are paying off, while the AI labs’ financial viability remains unsustainable.

Bending Spoons makes first post-IPO acquisition with $1.3B Airtable deal

Bending Spoons agreed to acquire U.S. software firm Airtable in an all-cash deal valued at $1.285 billion, marking its first acquisition since its Nasdaq debut last month. Airtable offers a no-code platform combining spreadsheets and databases for building applications and managing workflows. The transaction is expected to close by year-end, subject to regulatory approvals.

Bending Spoons entered definitive agreement to acquire Airtable for $1.28B

Bending Spoons has agreed to acquire Airtable in an all-cash deal valued at an enterprise value of $1.285 billion, implying an equity value of about $2.25 billion. Airtable, which has roughly $480 million in annual recurring revenue, will join Bending Spoons’ portfolio, with the transaction expected to close later this year pending regulatory approvals.

🌍 Policy & Society

White House won’t publicly release voluntary AI framework; details only for participating companies

The White House will not publicly release its voluntary framework for evaluating advanced AI models, with details shared only with participating companies. This secrecy leaves outside firms, policymakers, researchers, and U.S. allies guessing about implementation of a key AI policy.

DOJ settles with OpenAI for $3.2M over visa worker discrimination claims

OpenAI’s operating unit agreed to a $3.2 million settlement with the Justice Department to resolve allegations that it violated the Immigration and Nationality Act by favoring temporary visa holders over qualified U.S. workers in roles tied to a federal program requiring American worker priority.

World Bank: For developing economies, AI’s output boost outweighs job losses.

A World Bank report indicates that developing economies are likely to benefit more from AI enhancing worker productivity than from job displacement. The report suggests workers in poorer countries have less to fear from the technology than those in rich nations.

US police arrest 37 in 2026 data center protests, intervene in 12 more cases

In 2026, US police have arrested at least 37 people in connection with data center protests, with 12 additional cases of police intervening to prevent confrontations. The demonstrators, ranging from farmers to teachers, are overwhelmingly peaceful, with many arrests stemming from minor rule violations like clapping or exceeding speaking time.

Tennessee congressional hopeful accused of shooting license plate cameras

A Tennessee congressional candidate has been arrested for allegedly shooting license plate cameras, reportedly part of Flock’s surveillance network. The politician faces legal action over the incident, which appears to be a protest against the expansion of automated license plate readers. Details of the charges have not been fully disclosed.

UK mulls making employers ask before installing bossware

The UK is considering new rules that would require employers to obtain permission before installing workplace surveillance software, commonly known as “bossware.” The proposed measures would cover technologies such as AI productivity scoring, keystroke logging, and biometric monitoring.

China claims global chip leadership under new rules that redefine “integrated circuits” to its advantage. The regulations also protect local chip designs from being copied.

Why some people mow a lawn better than others

A study of 30,954 people mowing a virtual lawn found that humans are surprisingly skilled at coverage path planning, with 52% finishing within five moves of the optimal path and 16% achieving a perfect route. The task relates to the Traveling Salesman Problem, and the findings show that people, like computers, rely on heuristics when problems become too complex to solve optimally.

Most countries provide between 20 and 40 paid days off

Amanda is pausing her Not-Ship newsletter until the end of August to take a vacation, though she plans to use the break for deeper investigations and website upgrades. The article notes that most countries mandate 20–40 paid vacation days per year, but the US provides no legally required paid time off.

Less Protein, Not More, May Lead to Healthier Aging

A new review of over 350 studies suggests that sedentary people may already consume more protein than needed, and that reducing intake is linked to better metabolism, less cellular damage, and healthier aging. The authors point to the hormone FGF21 and certain amino acids as key factors, arguing that protein recommendations should be personalized based on physical activity levels rather than broadly increased.

It’s not a fear of “AI communism”; it’s a fear of competitive market capitalism

The article argues that the AI industry’s massive capital expenditures and debt may not yield expected profits, as free open-weight models like China’s Kimi K3 rapidly close the gap with proprietary frontier models. This competitive pressure undermines the pricing power of firms like OpenAI and Anthropic, prompting them to seek government protection and risking severe financial fallout for heavily indebted companies such as Oracle.

US Military asks troops for ‘creative and unconventional’ ideas to punish Iran

US Central Command emailed analysts requesting creative and unconventional ideas to pressure Iran, signaling limited options as Trump weighs further strikes. Buried nuclear facilities would likely require ground troops, which Trump has avoided, and military leaders acknowledge bombing alone is unlikely to achieve his goals.

Nobel Disease

Nobel disease is the informal term for Nobel Prize winners who embrace scientifically unsound or strange ideas, often later in life, partly due to feeling empowered to speak outside their expertise and receiving heightened media attention. It illustrates that being an authority in one field does not make one an authority in others and that high intelligence does not preclude irrationality.

U.S. used ‘virtually all’ of its long-range precision missiles during Iran war

The U.S. Army has used “virtually all” of its long-range precision missiles, including ATACMS and Precision Strike Missiles, during the five-month war with Iran, raising concerns about military readiness for future conflicts. Officials worry the depleted stockpile could weaken deterrence against adversaries like Russia and China, though the White House insists the U.S. has far more munitions than needed and is increasing production.

Mini retirements: like a sabbatical, but much smaller

Instead of waiting until retirement, taking “mini-retirements”—extended breaks of one to three months—offers deeper rest and creative insights than standard vacations, which are often too short to fully disconnect. The author shares a personal example of a 15-week trip to Australia that, despite initial fear, fostered personal growth and self-understanding, echoing Steve Jobs’ idea that life’s dots only connect looking backward.

Ray Bradbury’s “There Will Come Soft Rains” is set today (2026-08-04)

In Ray Bradbury’s story, set on August 4, 2026, a fully automated house in a post-apocalyptic city keeps performing its daily routine even though no humans remain. A dying dog enters the house and dies, and the house cleans it away. The story ends with the house destroyed by fire, leaving only silence.

The US is not a democracy but an oligarchy, study concludes (2014)

A Princeton and Northwestern study found that U.S. government policy is determined by economic elites and business interests, not the majority of citizens. Analyzing nearly 1,800 policies from 1981 to 2002, the researchers concluded that average citizens have little to no independent influence, posing a serious threat to American democracy.

Denmark: EVs Make Up 97% of New Private Car Sales in July

In July, electric vehicles accounted for 97 percent of new private car sales in Denmark, a record high. Overall, 14,562 passenger cars were sold in the country during the month.

Telegram Removed from App Store

Telegram was removed from the App Store, as reported in the article’s headline. The messaging app, which has over a billion users, is known for its fast, encrypted chats, unlimited file sharing, and support for groups of up to 200,000 members. It also offers features like self-destructing Secret Chats and cross-device syncing.