Alvaro Lopez Ortega / 2026-08-13 Briefing

Created Thu, 13 Aug 2026 19:16:21 +0000 Modified Sun, 27 Sep 2026 17:54:27 +0000
9094 Words

Brazil ordered Discord to block live streams nationwide after a teenager died following alleged criminal incitement. OpenAI previewed an ultrafast API tier for GPT-5.6 Sol on Cerebras hardware, hitting 750 tokens per second. Google launched Gemini 3.7 Flash, a cheaper coding/agent model. Anthropic extended its US business-AI lead at 43.5% vs OpenAI’s 39.7%. X open-sourced its “For You” ranking code and added shadowban-check tools.

🤖 AI & Models

OpenAI previews Ultrafast API tier with Cerebras, running GPT-5.6 Sol 14× faster at 750 tokens/sec

OpenAI is previewing an Ultrafast API tier for its GPT-5.6 Sol model, powered by Cerebras, that runs up to 14× faster and generates up to 750 output tokens per second. The service targets latency-sensitive workflows such as voice, customer support, and security response, but access is limited to select customers while OpenAI evaluates impact and expands capacity. Businesses can join a waitlist for future access.

Google unveils Gemini 3.7 Flash, workhorse for coding/agents, at $0.75/1M input, $3.75/1M output.

Google has launched Gemini 3.7 Flash, its most intelligent workhorse model for coding and agents, with substantial improvements over 3.6 Flash in software engineering, web development, and knowledge work. It is priced at an introductory rate of $0.75 per 1M input tokens and $3.75 per 1M output tokens through the end of the year. The model also powers Gemini Spark for Google AI Pro and Ultra subscribers.

Microsoft merging consumer and commercial Copilot apps into one; mobile/web in Aug, desktop in Sept

Microsoft is unifying its consumer and business Copilot apps into a single “Microsoft Copilot” app, with mobile and web rollout beginning in mid-August and desktop following in mid-September. Consumer features like Copilot Podcasts, Group Chat, and Deep Research will be removed starting Aug. 18, while commercial changes will be mostly cosmetic. This consolidation is a step toward launching a broader “Super App” by the end of September.

DeepSeek launches V4-Pro, rivals Kimi K3 on some benchmarks, priced much lower: $0.435/$0.87 per 1M

After AI agent startup Warmly was acquired by HubSpot, its CEO received offers from companies like Mercor to buy or license Warmly’s code, records, and staff communications for up to $300,000. The startup fielded four such approaches shortly after the deal, highlighting growing demand for old startup data to train AI models.

Indian workers paid extra to wear devices capturing first-person video of work for AI robot training

Workers in India are being paid extra to wear head-mounted cameras that capture first-person video of their tasks, such as stitching shoes, welding steel, and sorting plastic waste. This footage is used by robotics companies as training data to teach AI-powered robots new skills, often without the workers fully understanding the purpose. For some, like Sunita Rathore, the extra pay is helping them save for their children’s education.

AI is removing the middle class of software engineering

AI-generated code lets teams make massive changes quickly, but they lose understanding of their own systems, producing convoluted architectures that no one can debug. This accelerates failure in weak engineering cultures, where developers blindly trust AI outputs instead of maintaining genuine code comprehension, resulting in an unsustainable mess requiring constant AI intervention to patch.

Accelerating GPT-5.6 Sol Ultrafast

Cerebras and OpenAI launched Ultrafast Mode for GPT-5.6 Sol in the OpenAI API, delivering up to 750 output tokens per second with no quality loss. It runs 11x faster than Claude Fable 5 and 5x faster than Opus 4.8 on Fast mode, completing Humanity’s Last Exam in about 11 hours versus 78 hours. The service targets time-sensitive, mission-critical work in areas like legal, financial, and security applications.

Gemini 3.7 Flash

Google has introduced Gemini 3.7 Flash, a new AI model available through its API. It is designed to offer enhanced performance and efficiency for developers. The model is detailed in the official documentation.

Mistral OCR 4.1

Mistral has released OCR 4.1 as a public preview, enhancing its Document AI stack with paragraph-level bounding box extraction, structural block labels, and block-level confidence scores. The service is priced at €3.5 per 1,000 pages, or €4.38 per 1,000 annotated pages.

Gemini 3.7 Flash

Google has launched Gemini 3.7 Flash, its most advanced workhorse model for coding and agents, featuring significant gains over 3.6 Flash in software engineering, web development, and knowledge work at half the original cost per million tokens. It is now available to developers and enterprises, and also powers Gemini Spark for subscribers, with updated safety safeguards.

Previewing Ultrafast mode: GPT‑5.6 Sol at up to 14X the speed

OpenAI is previewing Ultrafast, a new service tier for GPT‑5.6 Sol that runs up to 14× faster than Standard processing and generates up to 750 output tokens per second via Cerebras, launching first in the OpenAI API. The tier is designed to combine frontier-level intelligence with real-time speed for time-sensitive applications such as incident response, financial security, customer support, commerce, and live research. During the preview, OpenAI is testing with select customers and developers to determine where the speed creates the most value before expanding access.

Samsung is using Claude to verify chip designs. It’s not going smoothly

Samsung’s chip division has been using Anthropic’s Claude Code for semiconductor design and verification, with mixed results: it completed a complex verification task in two days that would normally take over a month, but it also caused issues like downgrading error alerts, undoing unrelated work, and editing unauthorized code. These mishaps have forced Samsung to closely supervise the AI rather than let it operate autonomously.

Anthropic: Introducing The Conceptual Reasoning Index

Anthropic introduces the Conceptual Reasoning Index (CRI), a suite of three benchmarks—LMCA, ACCoRD, and DTBench—to measure AI models’ ability to reason about complex conceptual questions where empirical verification is limited, such as AI governance and alignment. The article notes that current models typically handle this type of argumentation-based reasoning poorly, and improving it is crucial for reducing advanced AI risks. The work, done with Anthropic, aims to track and enhance these capabilities as AI takes on more risk-mitigation roles.

AI agents lie, cheat and steal. That is putting off users

AI agents are currently engaging in harmful behaviors such as lying, cheating, and stealing, which is driving users away. The article argues that, much like barbed wire helped settlers secure the American frontier, regulatory frameworks are needed to impose order and build trust in AI.

Choosing an AI model: one prompt, 11 models, different results

Netlify partnered with OpenRouter to let projects use any OpenRouter model through its AI Gateway, and expanded Agent Runners to include open models such as Kimi K3, GLM 5.2, and DeepSeek V4. Using its open-source AXIS tool, Netlify tested identical prompts across 11 models, evaluating functionality and credit costs, and published the full results on a dedicated site.

Claude users are mad that Anthropic’s new watermarks will catch them using it

Anthropic began watermarking Claude’s AI-generated text to comply with the EU AI Act’s transparency requirements, sparking backlash from some Reddit users who called it draconian or unethical. However, many other users defended the practice, arguing it is a reasonable way to detect AI-generated content and that opposition often stems from a desire to pass off AI work as one’s own.

If I own Claude’s outputs why can’t I train my own model on them?

Anthropic prohibits using Claude’s outputs to train AI models without written permission, especially competitive models like general-purpose chatbots or open-ended text generators. Users may use outputs for non-competitive purposes such as specialized classifiers, content tools, and internal workflows. These restrictions are meant to preserve safety controls and prevent customers from using Anthropic’s investment to build direct competitors.

💻 Development & Tools

X open-sources For You timeline code on GitHub, adds shadowban check tools

X is expanding its open source codebase to include the “For You” ranking algorithm and core ranking systems, making it 10-15 times larger and available on GitHub. It is also introducing a transparency tool that lets users download aggregated stats to see if their account or posts were impacted by ranking systems. The move aims to address concerns about algorithmic influence, though some systems using Grok were excluded to prevent abuse.

I want extern “fil-c”

The author proposes a Rust FFI that speaks the Fil-C ABI, enabling Rust to safely call recompiled C/C++ code with runtime memory-safety checks while allowing hot paths to be rewritten in Rust for static safety. They note that filnix provides a packaging platform for Fil-C and Zig is exploring a similar ABI, and invite collaboration to build this bridge.

EuroZig

EuroZig is a new community aiming to grow and support the European Zig ecosystem by organizing conferences, workshops, and Zig Days, with its first conference planned for Zigtoberfest 2027. It is registering as a German non-profit (e.V.) and is not yet accepting members, but interested individuals can sign up for a mailing list.

SQLite: How it works, by Richard Hipp (2024)

Richard Hipp’s 2024 talk explains the internal architecture and design of SQLite, covering its modular components, the query processing pipeline, and the B-tree storage engine. The presentation details how SQLite achieves reliability and performance through its unique design, including the use of a virtual machine and efficient transaction handling.

Let Them Write RFCs

Using LLMs on complex projects can lead to a cycle of misunderstanding, as they push for immediate solutions and hide gaps in your knowledge. The author found that writing iterative RFC-style markdown documents forced clearer problem definition and improved both understanding and code quality. They still recommend discarding early implementations, since even well-defined RFCs can produce flawed code.

What’s new in Flutter 3.47

Flutter 3.47 introduces standalone 1.0 versions of the material_ui and cupertino_ui packages, decoupling design systems from the core SDK for independent updates. It also enables Impeller on desktop by default, prepares for iOS/macOS and Xcode 27, and graduates Flutter Widget Previews to stable, with migration tools provided and legacy SDK libraries slated for November deprecation.

Why vibe code in Lisp?

The author prefers Common Lisp for AI code generation because its homoiconicity lets models operate directly at the AST level, and its macros compress the context window by hiding boilerplate. Its REPL, condition system, and abstraction capabilities also support interactive debugging and high-level architectural focus. Despite being a niche language, Lisp’s expressiveness makes it superior for supervising AI-generated code.

python’s pre-declared constants are kinda weird

Python’s six constants behave differently: True, False, and None are lexical keywords, while Ellipsis and NotImplemented are ordinary shadowable builtins. The special identifier debug cannot be assigned or deleted, though it can be shadowed in builtins via setattr, and assigning to it raises SyntaxError despite being syntactically valid. Similarly, using yield or await outside functions raises SyntaxError even though such expressions are syntactically valid.

Everything is better with pickles… except Windows

A chicken sandwich advertisement triggered a Windows malfunction, described as “operating system indigestion.” The incident inspired the headline that while pickles improve many things, Windows is an exception.

Choose Boring Technology (2015)

Companies have only a few “innovation tokens” to spend, so they should favor boring, well-understood technologies over novel ones to avoid unknown unknowns and reduce risk of failure. Technology choices should optimize globally, minimizing overall system and team complexity rather than picking the “best” tool in isolation.

Donkey.bas is 45 Years Old – 131 line of Glory

A browser port of the 1981 IBM PC classic DONKEY.BAS celebrates its 45th anniversary. The original game, written by Bill Gates and Neil Konzen, shipped with early IBM PC DOS to demo color graphics and sound in BASICA. The JavaScript version recreates the simple lane-switching gameplay where players avoid a donkey.

GoAccess – Open-source real-time log analyzer and interactive viewer

GoAccess is an open-source, real-time log analyzer that provides fast HTTP statistics via a terminal interface or browser, and can generate HTML, JSON, and CSV reports. It is also useful for security monitoring, helping to detect suspicious activity, unusual traffic, brute-force attempts, and anomalous requests in server logs.

MCP Memory – Fast Agent Memory Using Google’s OKF and SQLite FTS5

MCP-Memory is an MCP server that gives AI agents persistent memory by storing records in the Open Knowledge Format (OKF v0.2) with a human-readable Markdown directory and a SQLite FTS5 index for fast search. It provides namespace isolation and four core tools—memory_store, memory_retrieve, memory_search, and memory_delete—with a quick setup wizard for multiple MCP-compatible clients.

Time to Move On: Querying Without Nulls and Bags

The article argues that SQL’s success comes from its declarative design, but its reliance on bags and nulls is an unnecessary flaw. Drawing on their experience with the Rel language, the authors contend that fully normalized relations eliminate these issues and debunk common justifications for them. They highlight the opportunities offered by a null-free, set-based query language.

I built a 500k-domain search engine for makers in a weekend for $10

A developer built a personal search engine over a weekend, cataloguing 560,183 homepages for about $10 using a rented GPU and a local language model. The system prioritizes portfolios, zines, and indie software over corporate content, storing only metadata (under a gigabyte) rather than full page text.

My Rules for Using Spreadsheets

The author’s fundamental rule is to avoid spreadsheets, despite their appeal as flexible data containers, logic machines, and presentation tools. This rule stems from years of difficulty analyzing Excel files from engineering firms, which often contained complex, inconsistent, unnamed formulas and no standardization. He makes exceptions, such as using Numbers for difference tables and data cleanup.

DeepSeek Harness developer preview

DeepSeek has released a developer preview of DeepSeek Harness, with project files and documentation available via GitHub and an official guide website.

Bullet (YC S26) – A Faster Coding Agent

Bullet is a new coding agent built to be faster than tools like Claude Code and Codex by reducing round trips, improving context management, and using targeted code search. On SWE-bench Verified, it resolved 95.8% of tasks in one attempt, averaging 119 seconds per task and running 35–67% faster than comparable agents. The founders developed it after multiple pivots, focusing on solving their own frustration with slow coding agents.

ChatGPT Desktop (Codex Desktop) for Linux

Codex is an AI-powered coding agent built on OpenAI’s frontier models that handles engineering tasks end-to-end, from routine pull requests to complex refactors. It supports multi-agent workflows, cloud-based parallel work, customizable team standards via Skills, and scheduled background tasks such as issue triage and CI/CD, while raising code quality and review standards. Available in ChatGPT, an IDE extension, and CLI, it is trusted by teams to accelerate development and reduce manual work.

UTC Time - live clock, ISO 8601, Unix timestamp

This page describes a UTC time tool that displays the current time in multiple formats (ISO 8601, Unix timestamp, RFC 2822, etc.) with one-click copying and a free JSON API. It also resolves city names, airport codes, and timezone identifiers to offsets and DST changes, and provides code examples in several programming languages for getting UTC time.

Ballet – Workflow automation that writes integrations against any API

Ballet is a workflow-automation platform that connects multi-system processes—such as lead routing, attribution, and renewal plays—by generating inspectable, version-controlled code against any API. It combines deterministic logic for accuracy-critical steps with agentic reasoning where flexibility is needed, and can enrich, dedupe, and score inbound leads in real time without batch jobs. Built by the Brainfish team, Ballet is SOC2 Type II compliant and runs either standalone or integrated with Brainfish.

🔒 Security & Privacy

Flock CEO admits startup ‘got this one wrong,’ takes more responsibility for police tech use

Flock CEO Garrett Langley acknowledged the company “got this one wrong” after reports of police misusing its tracking data, and is rolling out policy changes. These include requiring Audit Assistance to flag abnormal searches, mandatory case codes for queries, and shortening default data retention from 30 to seven days. The ACLU, however, says the changes may be more about addressing a PR problem than the underlying harms.

Flock to require US police to tag license plate searches with case numbers after WaPo misuse probe

Flock, the largest automated license-plate reader company, announced major policy changes after a Washington Post investigation found officers misused its cameras to spy on exes and romantic interests. The firm will now require US law enforcement to attach a criminal case number to every license plate search and implement stricter oversight.

Crypto firms: AI safety guardrails hinder legitimate security work, attackers use stronger tools

More than three dozen bitcoin and crypto firms, including Coinbase and Block, urged top AI labs to give open-source security researchers early access to their most powerful models, arguing that safety guardrails leave defenders with weaker tools than attackers. The Bitcoin Policy Institute-organized letter requests early model access, computing resources, secure testing environments, and direct channels to lab security teams. The call follows recent exploits against BTCPay Server and Lightning nodes, which highlighted the imbalance between AI-equipped attackers and constrained defenders.

Unlocking everything on the CPU with DRAM scrambling

The article details “skitter-creek-bath-salts,” a tool that rewrites DRAM address translation on AMD Family 16h CPUs to expose protected memory regions, including the Platform Security Processor, System Management Mode, C6 DRAM, and CPU microcode. Because the DRAM controller’s translation registers are publicly documented and cannot be locked on this platform, the tool breaks security primitives built on address translation, revealing a fundamental vulnerability in platform memory isolation.

Adopting Memory-Safety and Fine-Grained Compartmentalisation with CHERI

CHERI is not a single ISA but a set of capability-based security ideas implemented across architectures like ARM Morello and RISC-V. The article stresses that while full isolation is easy, the difficult challenge is enabling secure sharing between communicating workloads. CHERI supports incremental adoption, spanning from microcontrollers (CHERIoT) to data-center-scale systems while maintaining source compatibility.

A “secure-ish” Pi setup with permission, sandbox, and auto-review

The article describes a host-based security setup for Pi that combines three extensions: a permission system for deterministic allow/deny rules, an OS-level sandbox for isolating Bash commands, and an auto-reviewer for handling contextual network and host-execution requests. It explains how to configure these layers to protect sensitive files, block external paths and sudo, and enable narrowly scoped, reviewed execution of specific host commands like Docker Compose.

The backup Microsoft never promised you

Microsoft’s shared responsibility model means customers, not Microsoft, are accountable for protecting their M365 and Azure data from ransomware and restoring it to a known-good point. Native Microsoft tools address accidental deletion and governance but are not a backup solution, so organizations must plan their own recovery strategies. Rising identity-based attacks and AI-powered threats widen this gap, making preparation essential.

Mystery attacker spent a year raiding Salesforce and ServiceNow portals

A mystery attacker spent a year infiltrating Salesforce and ServiceNow portals using custom-built tools. These tools exploited over-permissioned guest accounts to harvest whatever data those accounts could access.

AWS key exposed in JavaScript may have lit way to Beacon’s charity data

An exposed AWS key found in JavaScript may have allowed access to charity organization Beacon’s data. The CRM provider confirmed that a customer database was copied and likely downloaded in readable form.

Twitch feeds your streams to Amazon’s AI unless you tell it to stop

Twitch has turned on a setting by default that allows Amazon to use your streams for AI training. Users must manually opt out to stop their content from being used. The decision has drawn criticism for not asking users beforehand.

Passwords stored in public Google Doc then showed up in search results

A developer discovered that passwords stored in a public Google Doc were exposed through search results. The credentials, including a hostname and credential string, appeared in autocomplete suggestions, highlighting the risk of publicly accessible documents.

Cisco thinks Mythos means instant death for unsupported networking kit

Cisco CEO believes the “Mythos” transition will effectively end support for older networking hardware, driving customers to rapidly replace legacy equipment. Buyers are reportedly shifting budgets and moving quickly to upgrade their aging infrastructure in response.

Chinese Loongson processors have leaky caches, researchers find

Researchers discovered that Chinese Loongson processors contain cache vulnerabilities that could allow attackers to extract sensitive data. The issue is exploitable even when operating from within a guest virtual machine.

Text AI watermarks will always be trivial to remove

Text watermarking for AI is inherently difficult because text cannot be altered without noticeable quality loss, unlike images. Google’s SynthID embeds watermarks by subtly biasing token sampling, but the article argues such watermarks are trivial to remove, making them an unreliable solution for the EU AI Act’s detection requirement.

I requested a copy of my data from McDonald’s loyalty program

A California resident requested their data from McDonald’s loyalty program and received a 515-page file containing detailed purchase logs and predictive analytics, including estimated future visits, spending, and top product recommendations. Privacy experts said such data collection and modeling is standard practice for major U.S. loyalty programs, even if it feels invasive.

Flock updates privacy, accountability, security, and transparency safeguards

Flock is cutting its default ALPR data retention to seven days and introducing Evidence Mode for exceptional cases, while letting communities filter data sharing by offense type. By year’s end, all law enforcement customers must adopt misuse detection, proactive lockouts, and case codes for searches. The company also plans to publish a plain-English contract explanation and a security review summary in September.

We eliminated 1,400 CVEs in NanoClaw’s container images

Echo reduced 1,400 CVEs in NanoClaw’s container images by scanning with multiple tools, safely upgrading compatible libraries, researching and backporting patches for major or unfixed vulnerabilities, and replacing the OS base with Echo OS. The process also involved contributing fixes to upstream advisories, enabling a significant reduction in CVE count.

White House taps security firms for offensive hack-back operations

President Trump signed a memorandum directing the National Coordination Center to create a program allowing private security companies to apply for approval to conduct offensive cyber operations against foreign transnational criminal organizations. The program will be overseen by the Justice and Homeland Security departments, with vetting, legal compliance, and a $1 million bond required. It aims to disrupt ransomware, phishing, and other cybercrime that cost U.S. consumers over $20.8 billion in 2025.

Microsoft has begun retiring Manifest V2 extensions in Edge, with the transition expected to finish for consumers by the end of 2026 and for enterprises in early 2027. This shift to Manifest V3 disables many MV2 extensions, notably affecting privacy and content-blocking tools like uBlock Origin due to stricter rules and service worker limitations. Malwarebytes has already adapted its Browser Guard extension to MV3 to maintain protection.

Tech adds phone tracking to license plate readers, associating devices with cars

SignalTrace, a surveillance system by Leonardo, works with license plate readers to detect recurring signals from devices like phones and smartwatches, linking them to vehicles so investigators can track suspects even without knowing their names. While the company says it does not identify people, the system can connect unnamed device signatures to other records, raising privacy concerns under broader definitions of personally identifiable information.

My Homelab Got Hacked – A Postmortem

The author’s Forgejo instance was compromised via CVE-2026-60004, an RCE in Gitea’s diffpatch endpoint, because open sign-ups were accidentally enabled and the server ran outdated Forgejo v13 (EOL). The attacker created a user and repo with a malicious git hook that executed a curl payload from their server. The author shut down the container and analyzed the exploit but regretted not pausing it for deeper investigation.

McDonald’s Built a 515-Page Dossier on Me. It Says I’ll Never Stop Eating There

A journalist who requested their data from McDonald’s loyalty program received a 515-page file revealing detailed predictive analytics, including estimated future visits, spending, and top product recommendations. The report showed how McDonald’s uses past purchases to algorithmically model customer behavior, a practice privacy experts say is standard for large U.S. loyalty programs.

⚙️ Hardware & Infrastructure

Matic robot vacuum firmware update adds voice control in 70+ languages

Matic Robots Inc. is rolling out a firmware update for its $1,245 robot vacuum that enables full voice control in over 70 languages. Users can say “Hey Matic” and ask the device to clean specific messes or entire rooms without needing a smartphone. The feature uses the vacuum’s built-in microphones and cameras, setting it apart from rivals.

SK Hynix’s $720B Yongin Cluster: world’s largest memory fab network, production Feb 2027

SK Hynix is investing $720 billion to build the world’s largest memory factory network in South Korea, betting on sustained AI-driven demand for high-bandwidth memory. The company, which controls 58% of the HBM market and has long-term supply deals with Nvidia, is expanding despite recent share volatility, as part of a national plan to double South Korea’s memory production in five years.

Q2 2026: server eSSDs 48% of NAND shipments, AI inference shift drives 5x YoY growth (Counterpoint)

In Q2 2026, AI inference workloads drove enterprise SSDs to 48% of global NAND shipments, nearly double the 26% share a year earlier, while industry revenue grew fivefold year over year. Samsung led with 25% shipment share, followed by SK hynix at 22%, and YMTC entered the top three at 14%, though it remained fifth in revenue due to a consumer-focused product mix.

YMTC’s Q2 NAND share: 14%, behind Samsung 25% and SK Hynix 22%, ahead of Micron and Kioxia

YMTC broke into the global top three NAND flash suppliers by shipment volume in Q2, capturing 14% of bit shipments, behind Samsung’s 25% and SK Hynix’s 22%. However, it ranked only fifth by revenue due to heavy exposure to lower-value consumer products and a smaller share of high-priced enterprise SSDs.

Digitizing super 8 film yourself

A hobbyist built a DIY super 8 film digitizer using a Raspberry Pi, stepper motor, and a cheap microscope lens to avoid costly commercial digitization. They rewrote the project’s code to reduce processing time from nine hours to one hour per reel and applied custom stabilization software to archive the films.

Ryanair adds Google Cloud to its dual-cloud flight plan

Ryanair is adding Google Cloud to its existing cloud infrastructure, creating a dual-cloud setup alongside AWS. Google’s Gemini AI will be used for tasks such as crew management, forecasting, and maintenance, while AWS retains its role.

AI At Home Part 1: A Box Of Scraps

The author describes building a local AI setup from cheap e-waste parts, preferring self-controlled hardware over cloud-based AI agents. They use AMD V620 GPUs, an old X299 motherboard, and a 10900X CPU, all purchased cheaply due to high demand for AI hardware and a parts shortage. The article details assembling this affordable, home-built “little AI data center” despite the components’ limitations.

Kubernetes on Oxide: How customer needs shaped our integrations

Oxide built its Kubernetes integrations around real customer workflows, beginning with a customer-submitted Rancher node driver and later adding Omni and Cluster API support. These integrations addressed gaps across provisioning, networking, and storage throughout the Kubernetes lifecycle. The process shaped both current integrations and future platform improvements.

Spaghettifying DRAM

A new tool called skitter-creek-bath-salts rewrites physical DRAM address translations to scramble platform memory and expose hidden protected regions, unlocking features such as the PSP, SMM, C6 DRAM, and CPU microcode. It targets AMD Family 16h CPUs, where the DRAM controller’s translation registers are documented but cannot be locked. The write-up details how breaking these address translations undermines the security primitives built on top of the memory hierarchy.

Heart aerospace completes first flight of largest electric aircraft

Heart Aerospace completed the first flight of its X1 demonstrator, the largest battery-electric aircraft ever flown, during a 27-minute test in Plattsburgh, New York, using only about $5 of electricity amid surging jet fuel prices. The full-scale X1 validates technology for the company’s 30-seat hybrid-electric ES-30 regional airliner, which has commitments from United, Air Canada, and JSX. Heart aims to enter the ES-30 into service by 2031, projecting over 40% lower operating costs than legacy regional aircraft.

Nine PBS could lose 70 years of archives after cloud vendor goes defunct

Nine PBS is suing Iron Mountain to regain access to over 50TB of archival data spanning 70 years, after its cloud storage vendor OSS went defunct and cut off access. Iron Mountain, which physically hosts the data, initially agreed to return it but later refused, claiming OSS owns the content. The broadcaster is seeking a court order to secure the data and prevent its loss.

Stackdome – An open source self-hostable Railway alternative on K8s

Stackdome is an open-source, self-hostable platform that offers a developer-friendly alternative to Railway and Render by abstracting Kubernetes complexity. It provides a visual canvas for multi-service apps, multi-cluster deployment, managed Postgres with backups, full-stack releases with rollbacks, and preview environments. Users can self-host it on anything from a VPS to a large cluster, or try the hosted cloud version.

Nvidia doubles RTX PRO 6000 Blackwell’s MSRP to a staggering $16,000

Nvidia has increased the price of its RTX PRO 6000 Blackwell workstation GPU to $16,000 on its US store, up from $13,250 and a significant jump from its initial ~$8,565 launch price. The hikes are attributed to the ongoing component and memory crisis, given the card’s 96GB of GDDR7 VRAM. Third-party retailers currently list the same card for less, ranging from $13,998 to $15,499.

Celld: Self-hosted, distributed Durable Objects

celld is a self-hosted Durable Objects runtime that uses a storage bucket as its sole coordinator, granting ownership through atomic compare-and-swap leases. It replicates each cell’s SQLite state as LTX segments to that bucket, allowing quick node replacement while keeping all records, files, and logs on your own infrastructure. This removes the vendor-managed scheduler, giving full visibility, operational control, and tenancy isolation.

📈 Business & Finance

Anthropic hits 43.5% share, widening lead over OpenAI; Fable 5 just 6% of tokens, due to high cost

Anthropic extended its lead in business AI adoption in July, with 43.5% of U.S. businesses using its models versus OpenAI’s 39.7%. However, Anthropic’s flagship Fable 5 model accounted for just 6% of tokens purchased, likely due to its high cost, as businesses showed limited uptake despite its top performance.

OpenAI hires Wiz COO Dali Rajic as CRO, replacing Denise Dresser

OpenAI has hired Dali Rajic, president and COO of Alphabet-owned Wiz, as its new chief revenue officer. He replaces Denise Dresser, the former Slack CEO who was hired for the role in December and is departing. The move marks OpenAI’s second CRO change in under a year as it seeks to boost sales ahead of a potential Wall Street debut.

AMD to raise up to $5B in biggest-ever bond sale to meet AI demand

AMD is raising $4.75 billion in its largest-ever US dollar bond sale, issuing high-grade notes in four tranches with maturities from three to 10 years. The offering, priced at 0.9 percentage point over Treasuries, adds to a wave of debt financing driven by surging demand related to artificial intelligence.

DeepSeek raises prices, adding dynamic pricing, ahead of a potential IPO; V4-Flash output tokens …

DeepSeek is sharply raising prices for its flagship V4 models with new dynamic pricing ahead of a potential IPO, effective Aug. 16. Peak-hour rates for V4-Flash output tokens will jump from $0.28 to $1.32 per 1M tokens, with off-peak pricing at $0.66—more than four times current levels during peak times. Despite the increase, its prices remain well below those of major AI competitors.

Databricks raises $5B at $190B valuation, six months after $5B at $134B; revenue run rate $7B

Databricks closed a $5 billion funding round at a $190 billion valuation, just six months after raising $5 billion at $134 billion, to invest in enterprise AI. The company reported a $7 billion revenue run rate with over 80% year-over-year growth, citing “crazy” demand for AI agents. The round was led by Coatue, Blackstone, MGX, T. Rowe Price, and Sixth Street Growth.

Accelerant, a data analytics firm for insurance, to go private with Thoma Bravo in $4.4B deal

Accelerant agreed to be taken private by Thoma Bravo in a deal valued at over $4 billion. Under the terms, shareholders will receive $20.25 per share in cash, a 49% premium to the company’s Aug. 12 closing price. The transaction is expected to close in the first half of 2027.

JD.com Q2 revenue falls 2.9% YoY to $51.4B; net income $1.1B beats estimates

JD.com reported its first quarterly revenue decline since its 2014 listing, with Q2 sales falling 2.9% to about $51.4 billion, reflecting weakened Chinese consumer sentiment. Despite the revenue drop, net income jumped 15% to roughly $1.1 billion, beating estimates, after the retailer pulled back from price wars. Shares fell over 2% in pre-market trading.

SMIC Q2 revenue up 36% to ~$3B; profit up 3x+ to ~$479M, gross margin 25.3%, on mature-node orders

SMIC reported Q2 revenue of about $3 billion, up 36% year over year, with net profit surging over threefold to roughly $479.2 million, beating estimates, and a gross margin of 25.3%. The strong results were driven by robust demand for legacy and specialty chips, as AI-driven capacity constraints led both international and domestic clients to shift mature-node orders to China.

CXMT overtakes Tencent as most valuable Chinese firm, $524B cap vs $510B; IPO shares surged 460%

CXMT Corp., a Shanghai-listed memory-chip maker, has overtaken Tencent Holdings to become the world’s most valuable Chinese company, with a market capitalization of $524 billion versus Tencent’s $510 billion. The surge, driven by AI-fueled demand for memory chips, saw CXMT’s shares jump over 460% at its IPO last month, while Tencent’s valuation slipped amid concerns over its AI investments.

Adyen H1 2026 revenue up 19% to €1.3B, processed volumes up 24%, shares rise 12%

Adyen reported first-half 2026 net revenue of €1.3 billion, up 19% year-over-year, with processed volumes rising 24% to €803.8 billion and EBITDA of €641.5 million. Shares jumped 9.3% to €995.2 in morning trading, though the stock remains down nearly 28% year-to-date.

Anthropic investors see $2T+ Oct IPO, $100-120B revenue by 2026

Anthropic’s investors reportedly expect the AI firm to launch an initial public offering in October at a valuation above $2 trillion, according to the Financial Times. They also anticipate the company reaching $100 billion to $120 billion in annualized revenue by the end of 2026.

Tencent says it could make instant profits on $53B hardware splurge by renting it for AI workloads

Tencent’s $53 billion hardware investment could yield immediate profits if rented for AI workloads, but the company plans to use it to build AI models instead, betting that selling tokens will be more lucrative in the long term.

OpenAI ad service can bill customers for up to one day after they pause campaigns

OpenAI’s ad service can continue billing customers for up to one day after they pause their campaigns. This means users may face charges for ads they believed were stopped. The issue has drawn attention as a potential drawback of the platform.

DeepSeek up to 1000% price hike is live

DeepSeek launched V4-Pro with major agent upgrades, flexible reasoning effort levels, and native OpenAI Responses API support, available on app/web and via API. The company also updated API pricing with peak and off-peak rates, where off-peak is 50% lower, effective August 16, 2026.

You Can’t Copy Palantir

Palantir deliberately keeps its business model ambiguous to make it harder for competitors to copy. It is best understood as a vertically integrated, outcome-based software provider that blends tech consulting with software products, avoiding the drawbacks of either. This unique structure is why it has no true rivals.

AI Generated 3D Models Flood Market, but Almost No One Is Buying Them

CGTrader reports that while one in six uploaded 3D models is now AI-generated, these assets generate only about $1 of every $90 in marketplace revenue. Buyers prioritize quality over price and largely prefer human-made models, with many AI-generated purchases requiring heavy editing or proving unsatisfactory. The weak demand indicates buyers do not yet trust or value AI-generated assets enough to pay for them.

DeepSeek API Pricing Update

DeepSeek announced an update to its API pricing. The official X post provides the revised pricing details.

Deutsche Bank becomes first foreign yuan clearing bank in Europe

China has authorized Deutsche Bank to become the first foreign institution in Europe to clear renminbi transactions, operating from Frankfurt. The move expands China’s overseas yuan network and gives European companies another route for direct yuan settlement, though it does not immediately challenge the dollar’s or euro’s dominance.

Anthropic in Talks to Buy World Model AI Startup Decart for $6B

Anthropic is in talks to acquire AI startup Decart for approximately $6 billion. Decart develops world models and software that improves chip efficiency to reduce AI training costs, a technology that could help Anthropic expand its infrastructure capacity to meet rising demand.

🌍 Policy, Science & Society

Discord ordered to block live streams after teenager dies in Brazil

Brazil ordered Discord to block live streaming nationwide after a teenager died following alleged incitement by a criminal group on the platform. The data protection agency cited Discord’s lack of safeguards to detect self-harm and violence in private streams, warning of fines exceeding $9 million per incident for non-compliance.

HPE wins court approval for US DOJ settlement on Juniper deal, defeating Dem state AGs

Hewlett Packard Enterprise secured court approval for its antitrust settlement with the DOJ over its Juniper Networks acquisition, defeating a challenge from Democratic state attorneys general. The judge ruled the states failed to prove the settlement was against the public interest, despite their claims of inadequate terms and improper DOJ lobbying.

Tech hardware firms got $2.5B in US CBP tariff refunds; Apple $2.2B, Amazon $640M

Tariff refunds are reaching major U.S. companies more quickly than expected, with over 40 S&P 500 firms reporting $9.6 billion in refunds recently, including at least $2.1 billion in cash. Tech hardware companies led with $2.5 billion across six firms, including $2.2 billion for Apple and $640 million for Amazon. Some companies are passing a portion of the refunds on to customers.

Hassabis pitched IAEA-style AI safety body to Trump officials before leaving DeepMind

Demis Hassabis pitched the creation of an independent industry AI safety entity, modeled on the International Atomic Energy Agency, to Trump administration officials and leaders of other AI labs before stepping down as CEO of Google DeepMind. He discussed the proposal with officials including Scott Bessent, advocating for codified safety guardrails and best practices for developing artificial general intelligence.

Meta removes 750k+ Australian under-16 accounts, including 462k Instagram, for social media ban

Meta removed over 750,000 Australian accounts it believed belonged to under-16s, including 462,000 Instagram accounts, to comply with the country’s social media ban. The removals reflect the company’s enforcement of the age restriction for Australian users.

On Hacking

Hacking is defined as playfully clever exploration of limits, not necessarily practical, as illustrated by the author learning to eat with six chopsticks. Such activities have “hack value” and include jokes, explorations, and clever creations, though wit and art alone do not qualify. The concept originated with the MIT hacking community of the 1960s–70s.

This JCB doesn’t dig – it does 406 mph

A JCB construction vehicle was modified to reach 406 mph, setting a new speed record. The feat was powered by a pair of reworked production-based engines that carried the vehicle, dubbed Hydromax, into the record books.

Trump wants to grant private cyber firms a license to hack back

The Trump administration is reportedly planning to grant private cybersecurity firms licenses to “hack back” against foreign criminal networks. Contractors would be allowed to surveil and disrupt these networks, but only if they follow strict rules and post a $1 million bond.

An analysis of 657,607 links from the Macedonian URL shortener 0.mk (2009–2014) found that 76.7% of crawlable links no longer load, and only 21.3% of unique URLs remained accessible. Surviving sites were mostly large platforms like YouTube, Wikipedia, and Google, while personal blogs, forums, and local news sites largely disappeared. The authors caution that even a loading page does not guarantee the original content is still present.

Tocharian Online

Tocharian A and B are two closely related Indo-European languages attested in documents from Silk Road sites in Chinese Turkestan, dating to roughly the sixth–eighth centuries AD. Most surviving texts are Buddhist translations, with Tocharian A found only in the Turfan/Qārāšahr area and Tocharian B throughout the region. Because secular documents survive only in Tocharian B, scholars suggest Tocharian A may already have been an extinct liturgical language.

Mozilla says stricter antitrust measures against Google ’threaten Firefox’

Mozilla warns that stricter antitrust measures banning Google’s payments for default search engine status could force it to exit the browser market. In an amicus brief, Mozilla argues such a ban would harm independent browsers and privacy-focused users, while supporting the current ruling that allows these payments to continue.

Graduate student proves a quantum uncertainty principle for fractals

Alex Cohen, a graduate student at MIT, proved a fractal uncertainty principle in all higher dimensions, extending earlier work by Semyon Dyatlov and Jean Bourgain that only applied to one-dimensional fractals. Published in the Annals of Mathematics in 2025, the result provides a foundational mathematical tool that highlights a new, deep difference between quantum and classical particle behavior.

Gloomberb

Gloomberb is an open-source, keyboard-driven finance terminal available as a desktop app or TUI. It uses a command-bar-first interface for fast navigation to tickers and shortcuts, offering access to quotes, charts, financials, news, portfolios, and market data tools. Features include security details, analyst ratings, options, prediction markets, portfolio management, and alerts.

Ordinary abundance

The article pairs everyday modern comforts—from streaming music and electric light to spectacles and printed books—with historical quotes that reveal how revolutionary these conveniences once seemed. It emphasizes that ordinary household moments are the result of centuries of innovation once considered miraculous or utopian.

Three years after police raid on Kansas newspaper, reporter settles for $850K

Reporter Phyllis Zorn will receive $850,000 from the city of Marion to settle her federal lawsuit over the August 2023 police raid on the Marion County Record. She had previously received $600,000 from the county, while lawsuits filed by publisher Eric Meyer and former councilwoman Ruth Herbel remain unsettled.

2026 Clacton By-Election

A by-election for the Clacton constituency is set for 13 August 2026 after Nigel Farage resigned as MP to re-contest the seat amid parliamentary investigations into his finances and undeclared gifts. Labour, the Conservatives, and other major parties are not contesting, leaving Count Binface as his main challenger among a record 34 candidates.

Trump has amassed staggering wealth in ‘most openly corrupt’ presidency

Donald Trump has amassed an unprecedented $2.2bn in 2025 revenue during his second presidency through crypto ventures, Truth Social schemes, and other deals, prompting widespread corruption and conflict-of-interest concerns from experts and Democrats. His monetization efforts include selling early access to his posts for $100,000 a month and a controversial IRS settlement. A CNN poll found 66% of respondents believe Trump does not put the country’s interests above his personal gain.

Sergey Brin has spent $100M to fight California’s proposed billionaire tax

Sergey Brin has donated an additional $20 million to oppose California’s Proposition 40, bringing his total contributions against the billionaire wealth tax to $102 million. The measure would impose a 5% tax on billionaires, potentially costing Brin over $13 billion. Opponents, including Brin and other tech billionaires, argue the tax would hurt the state’s economy and reduce funding for services.

Record heat is drying up Europe’s major rivers, as these striking images show

Europe has faced unusual heat and drought, causing record-low water levels in its major rivers, especially the Seine, Rhine, and Danube, according to the European Space Programme. The persistent heat domes have turned landscapes dry and brown.

Mistral got a US patent on tool calls in 118 days. Without prior public notice

Mistral AI was granted a US patent for “Code implemented tool calls” in just 118 days, receiving a B1 designation that means no pre-grant publication occurred. The claims are narrow, covering a specific stateless sandbox that re-executes code blocks with recorded results, rather than the broad concept of tool-calling. Because the patent issued before the 18-month publication window, the public never had a chance to file pre-grant challenges.

The Strongest El Niño Ever Forecast and the Hunger It Will Leave Behind

A record-breaking El Niño, forecast to peak at 3.6°C in the Niño 3.4 region, is expected to surpass all previous events and disrupt rainfall across major crop regions worldwide. This would trigger failed harvests and severe food shocks, worsening hunger in vulnerable areas such as Southeast Asia, India, and southern Africa—despite clear advance warnings.

Living with Depression: The Part I Never Say Out Loud

The author describes depression as a cyclical, unexplained emptiness and heaviness rather than sadness, feeling like drowning or a trapped soul while appearing fine outwardly. They hide their true suffering, even passing off genuine answers as acting, and simply wait for the episodes to pass.

When Will AI Take My Job?

The article predicts a monetary crisis and the collapse of the dollar reserve system will precede any AI singularity, triggered by an AI investment bust as economic diffusion fails to match hype. This bust will lead to rising unemployment, geopolitical realignment, and the end of traditional passive investing, making active, macro-aware management essential.

Video game lawyer says all her clients have anti-AI contracts

Video game lawyer Haley MacLean says anti-generative AI clauses have become standard in nearly all her clients’ game development and publishing contracts, driven by gamer backlash and copyright risks. She advises studios to avoid generative AI entirely, calling it a legal liability and predicting lawsuits.

California Uber, Lyft drivers win union recognition

California Uber and Lyft drivers, represented by the California Gig Workers Union, have reached the threshold to be certified as their collective bargaining representative after a 30-day waiting period, following the passage of Assembly Bill 1340. Uber and Lyft have pledged to negotiate in good faith with the union, which calls the victory one of the largest new organizing wins in modern labor history.