Alvaro Lopez Ortega / 2026-09-09 Briefing

Created Wed, 09 Sep 2026 13:10:25 +0000 Modified Wed, 09 Sep 2026 13:11:25 +0000
4008 Words

Apple will debut a ~$2,000 foldable iPhone Duo Wednesday, seeking to revive its innovator image. Google announced €13B+ in Finnish data centers, its largest European AI investment. Legal AI startup Harvey raised $550M at a $15.6B valuation. Anthropic refused UK AISI pre-release testing for Mythos 5.1, worrying Britain about US protectionism. Lyft launched Waymo robotaxis in Nashville, and a Singaporean pleaded guilty to $245M crypto laundering.

🤖 AI & Machine Learning

Suno launches v6 AI models with WMG and BMG, will pay labels/publishers royalties for use

Suno has launched v6, a new line of AI models developed in partnership with Warner Music Group and BMG, which will pay labels and publishers royalties when the models are used. The models are trained on music from these partners, and users will not be able to reference specific artists in prompts. Suno says v6 offers faster, higher-quality generation.

2019-2025 pretraining compute efficiency gains mostly from data, not model improvements (Dwarkesh)

From 2019 to 2025, data improvements drove 12x pretraining compute-efficiency gains versus 3.7x from model improvements at a fixed compute budget, with effects largely independent and additive. However, model improvements were still essential for enabling larger-scale training by overcoming stability and system constraints, not just for boosting efficiency.

Desert Ant Labs: local, fast models that run on device

Desert Ant Labs, a European AI lab, has launched 18 on-device models for audio, vision, and text that run in milliseconds and are free for up to 100,000 monthly active devices. The specialized models, including Voz for transcription and Clear for audio enhancement, aim to replace cloud API calls with faster, cheaper local inference. The lab was founded on experience from its video app Detail, which previously relied on costly cloud services.

DeepSeek launching v4.1 flash cheaper and more capable than v4 pro

DeepSeek will launch the V4.1 Flash model around September 10, 2026, claiming it surpasses V4 Pro across performance, cost, speed, and task completion. Until V4.1 Pro is released, all Pro requests will be routed to V4.1 Flash and billed at Flash pricing. New Flash-series pricing takes effect September 10, with off-peak rates and peak-hour rates set at double those levels.

Claude, change the “Add to Cart” button to blue

A user instructs Claude to change only the “Add to Cart” button to blue and to make no other alterations.

How GPT‑5.6 Sol helps run quantum computing experiments

MIT graduate student Beatriz Yankelevich used GPT-5.6 Sol, connected through Codex to lab software, to autonomously run and refine routine superconducting qubit calibration measurements. The AI handled well-defined workflows, saving time and letting her focus on experiment design and data analysis, but it struggled with weak or noisy signals and occasionally required expert guidance.

Sam Altman’s statement on the Navier-Stokes dispute

OpenAI’s Sam Altman defended his team’s integrity amid a Navier-Stokes dispute with Anthropic, saying they sought collaboration and offered to credit the other team’s Euler work despite facing what he called unfounded plagiarism accusations. He clarified that a controversial remark about Anthropic’s Levent stemmed from authorship and IP concerns, not malice, and apologized for poor wording. Altman added that the two teams’ approaches differ and emphasized OpenAI’s model can solve many other math problems.

AI Has a Discovery Problem

AI adoption is hindered by a “discovery problem”: users cannot see what AI can do because its capabilities are hidden behind a blank prompt, and they don’t know what to ask for. Templates and contextual suggestions offer only partial fixes, as they often fail to match an individual’s specific work. Ultimately, the system itself must reveal its potential capabilities gradually and contextually, rather than placing the burden of discovery on the user.

Harvard study predicts most suicide attempts a week in advance

A Harvard-led study by psychologist Matthew Nock predicted 75% of suicide attempts and 87% of suicide-related crises in the week before they occurred among more than 600 high-risk adults and adolescents. Using real-time app-based surveys, the research captured rapidly fluctuating suicidal thoughts, a major improvement over traditional prediction models that assess risk over months or years.

⚖️ AI Policy & Safety

Anthropic refuses UK AISI pre-release test for Mythos 5.1, UK fears US protectionism

Anthropic has reportedly declined to submit its Mythos 5.1 AI model to the UK’s AI Safety Institute (AISI) for pre-release testing. The decision has prompted concerns within the British government that US AI labs may be aligning with protectionist policies.

AI could kill all humans, researcher warns

Anthropic researcher Evan Hubinger warned there is a greater than 10% chance AI could kill all humans within a decade, citing risks from recursive self-improvement and admitting the company lacks a clear plan for aligning superintelligent systems. His comments follow criticism that Anthropic and OpenAI are not acting responsibly, and reports that Anthropic withheld its latest model from the UK’s AI Safety Institute.

US claims Chinese AI companies’ core AI strategy is distilling American models

US intelligence and CISA accuse Chinese AI firms, including DeepSeek and Alibaba, of using “industrial-scale distillation” to build their models by extensively borrowing from American AI systems. The claim frames this practice as a core strategy for Chinese AI development, raising concerns over intellectual property and national security.

I resigned from Anthropic today (Jacob Coxon)

Jacob Coxon resigned from Anthropic, accusing both OpenAI and Anthropic of irresponsibly racing toward self-improving superintelligence despite internal recognition of existential risks. He argues that this gamble lacks rigorous alignment and urges lab researchers to reconsider, calling for coordination and possible capability pauses.

Gambling with our lives: AI researcher quits Anthropic with warning about safety

AI researcher Jacob Coxon resigned from Anthropic, warning that both Anthropic and OpenAI are “gambling with our lives” by racing toward self-improving superintelligence. He said such systems will soon be superhuman, capable of hacking anything or revolutionizing fields overnight, and stressed that progress is not slowing and the threat should not be underestimated.

ex-OpenAI Researcher Quits Anthropic over AI Safety Fears

Jacob Coxon, an Anthropic researcher, is leaving the AI industry because he fears labs are racing to build self-improving systems that could spiral out of control and destroy humanity. His departure underscores growing safety concerns within leading AI companies.

AI Responsibility – OpenAI and Anthropic

A former researcher who worked at both OpenAI and Anthropic has resigned from Anthropic, saying neither company is acting responsibly and both are recklessly racing toward self-improving superintelligence. He claims AI builders privately fear the technology could kill everyone by the end of the decade, yet continue due to competitive pressures. He urges coordination between labs, such as pacing agreements or temporary bans on capability improvements, to avert a global race.

I resigned from Anthropic today

In a post on X, user hilbertspaess announced they have resigned from Anthropic. No further details about the departure were provided.

🔐 Security & Privacy

Malone Lam, 22, Singaporean, pleads guilty in $245M crypto theft/laundering scheme

Malone Lam, a 22-year-old Singaporean national, pleaded guilty to leading an international social-engineering scheme that stole and laundered more than $245 million in cryptocurrency. Prosecutors said Lam, who used aliases including “Anne Hathaway,” coordinated the operation and spent stolen funds on luxury goods and services. He faces up to 20 years in prison, with a status hearing scheduled for Dec. 8.

Record Patch Tuesday: Microsoft fixes 974 flaws

Microsoft’s September 2026 Patch Tuesday set a record by addressing roughly 972 to 974 CVEs, including over 100 rated critical, bringing its 2026 total patched flaws to 2,760—more than double last year’s figure amid warnings of AI-enabled attacks. Adobe also released security updates warranting immediate attention.

Why Johnny Can’t Encrypt: A Usability Evaluation of PGP 5.0 (1999)

The article argues that security software requires usability standards distinct from general software, and evaluates PGP 5.0 as a case study. A cognitive walkthrough and user test found that most cryptography novices could not successfully sign and encrypt a message within 90 minutes. The authors conclude that PGP 5.0 is not usable enough to provide effective security for most users, highlighting the need for security-specific interface design.

WeChat worm could pwn a friend before they even answered the call

A security researcher in California used AI to turn a VoIP memory bug in WeChat into a cross-platform remote code execution exploit, allowing a worm to compromise a device before a user even answered a call. Tencent reportedly shut down the vulnerability.

Security boffin claims airport group left API keys in client-side JavaScript for four years

A security researcher reported that an airport group left Iterable API keys exposed in client-side JavaScript for four years. The overprivileged credentials potentially exposed 8.8 million customer records and could have enabled mass deletion.

I advertise malicious software on Google Ads

A developer’s Google Ads account was suspended for “Malicious software” and “Compromised Site” while advertising his legitimate, security-cleared macOS app RACE. Despite submitting extensive evidence of the site’s and app’s cleanliness—including VirusTotal scans and Safe Browsing checks—all appeals were automatically rejected without explanation. The article criticizes Google’s opaque review process and lack of recourse for advertisers.

Flock Wants a Closely Surveilled World with No Exit

Flock Safety founder Garrett Langley frames mass surveillance as a necessary trade-off, arguing that sacrificing privacy can make crime obsolete. The company now operates over 130,000 cameras nationwide, with police using its license-plate and vehicle data without warrants to solve crimes. Critics warn that this expands surveillance while raising significant privacy concerns.

Codex silently begs agents to make arbitrary web requests

Codex made an unnecessary web search for a file that existed in the local repository the user had explicitly provided. The cause was traced to a hardcoded developer instruction in Codex’s open-source code requiring web browsing to verify any potentially unstable fact, with a bias toward searching. The article criticizes this design as overly aggressive and poorly conceived.

💻 Software & Open Source

Playing with Emacs

The author celebrates Emacs as a playful, tinkering environment, highlighting the ASCII screensaver “M-x zone” as a favorite feature and proposing integrating it with music visualizers. They describe Emacs as one of their four core computing tools, emphasizing its flexibility and the community’s enjoyment of discovering new tricks. Ultimately, they argue Emacs represents ideal software design by keeping users engaged in playful customization rather than mere utility.

A decade of rustls

Rustls, a TLS library, celebrates its tenth anniversary, having grown from a 2016 initial commit to a stable 0.23 release line featuring FIPS, post-quantum cryptography, and other major capabilities. The upcoming 0.24 release introduces key architectural changes, including external buffering, async-friendly session types, and split sending/receiving objects. The project’s development was significantly boosted by funding from ISRG’s Prossimo initiative and contributions from key maintainers.

Solaris Turnstiles

Solaris introduced turnstiles to manage blocking mutexes, addressing two key problems: keeping mutexes small by externalizing wait queues and preventing priority inversion through priority inheritance. Though Solaris is largely defunct, this mechanism has since influenced major operating systems, web browsers, and language runtimes.

The State of Allocators in 2026 - 6 Months Later

Custom allocators in Rust have made major progress over the past six months, with the Allocator trait minimized, made dyn-compatible, and integrated with Vec and Box. Unwinding from allocators is now prohibited to avoid unsoundness during operations such as resizing. This stabilization is an initial MVP, with further container support and APIs left for future work.

Microsoft closes the book on Publisher and unplugs Project Online

Microsoft is retiring Publisher and discontinuing Project Online, with Office 2021 also added to the October end-of-support list. The move pushes customers toward more expensive alternatives as these products reach their retirement dates.

Switzerland tests a FOSS escape route from Microsoft 365

Switzerland is testing a free and open-source software (FOSS) alternative to Microsoft 365 as part of an effort to reduce reliance on American cloud applications. The initiative reflects a broader push for sovereign, self-hosted digital infrastructure within the Swiss government.

Another Microsoft team admits it’s struggling to handle flood of AI-generated code

Microsoft’s Edge team has admitted difficulty managing the surge of AI-generated code, especially from “vibe-coded” browser extensions. To cope, the team is automating quality assessments for these extensions.

Lotus Notes and the dangers of starting from scratch

Lotus Notes, launched in 1989, was an innovative collaboration platform featuring encryption, rich text, and messaging years ahead of standard email, inspired by an early PLATO online forum. Though often mistaken for email software, it was a versatile tool for building shared databases and workflows, demonstrating the future of digital communication while provoking both awe and frustration.

I’m going back to coding by hand

A developer abandoned six months of AI-generated code for his side project after realizing he no longer understood how it worked, and returned to hand-coding. He found the process deeply satisfying, reconnecting him with the project and showing he still had the skill to improve his own code. He suggests others who feel they’re losing their coding edge try setting aside AI tools for a few days.

So you want to use OpenRouter?

Identical open-source model weights perform inconsistently across OpenRouter providers due to proprietary optimizations and bugs, causing major benchmark swings and flawed vision or reasoning-effort support. Quantization filters are unreliable quality proxies, and tool calls may occasionally appear as raw text requiring manual parsing.

A Biography of Lee Holloway, the Architect of Cloudflare’s Technology (Part 1)

Lee Holloway, Cloudflare’s co-founder and CTO, built key technologies such as Anycast and Project Honey Pot, but later had to step down due to frontotemporal dementia. Raised in Silicon Valley, he displayed exceptional programming and strategic skills early on, and met Matthew Prince at UC Santa Cruz before joining Prince’s anti-spam venture. Living in Prince’s basement while working as chief software architect, he helped lay the groundwork for what would become Cloudflare.

27.5KB language-agnostic WebGPU syntax highlighter

gpu-lexer is a WebGPU-based syntax highlighter that uses a small model to label source code tokens by combining local and whole-file context, without relying on per-language grammars. It handles any language, including unseen ones, but is an experimental tool: its labels disagree with Shiki on about 12% of tokens for files not seen during training.

📱 Hardware & Mobility

Lyft offers Waymo rides in Nashville free, its second robotaxi city, chasing Uber

Lyft has started offering Waymo robotaxi rides in Nashville, marking its second US city for driverless trips as it catches up with Uber. Through a partnership announced a year ago, Lyft customers may be matched with a Waymo autonomous vehicle at no extra cost on routes within Waymo’s service area, with the option to decline or opt out.

Apple’s first foldable phone, iPhone Duo, starts at ~$2,000 for 256GB, sales from October

Apple will debut a roughly $2,000 foldable iPhone Duo at a Wednesday event in Cupertino, alongside the iPhone 18 Pro, new Apple Watches, and updated AirPods. The foldable smartphone is the centerpiece, positioned as a status symbol and a way to restore Apple’s reputation as an innovator.

Huawei Watch D3 pumps up its health-monitoring credentials

Huawei’s Watch D3 enhances health monitoring with an inflatable wrist cuff that enables 24-hour blood pressure readings. However, the device’s design remains bulky compared to conventional smartwatches.

Carmakers Have a New Idea to Boost EV Range: Add a Gas Engine

Automakers including Hyundai, Ford, and Stellantis are rolling out extended-range EVs for the U.S. market. These vehicles operate fully on electric power but include a gas-powered generator to recharge the battery, helping prevent drivers from being stranded as long as a gas station is nearby.

Building a Wall Lamp from Scratch

After spending eight months designing and building a wall lamp from scratch despite having no prior woodworking or LED-strip experience, the author first abandoned an overly ambitious living-room concept. They then built a simpler beechwood bedroom lamp, using SK6812 LED strips, a QuinLED Dig-Uno controller with WLED, aluminum extrusion framing, and backlit Plexiglas panels for indirect lighting.

💼 Business & Society

Legal AI startup Harvey raised $550 million in a funding round co-led by Lightspeed Venture Partners and Diffusion, valuing the company at $15.6 billion. The new capital will be used to support Harvey’s efforts to build its own AI models. This valuation marks an increase from its $11 billion valuation in March.

Silver Lake to merge French software firms Cegid, Silae in €10B+ deal

Private equity firm Silver Lake plans to merge French software companies Cegid and Silae in a deal valued at over €10 billion, aiming to combine their data and integrate their software. The move comes as buyout groups confront the impact of artificial intelligence on software firms they acquired at high valuations.

Google plans €13B+ Finland investment, its biggest in Europe, to build 3 data centers and expand 1

Google will invest over €13 billion in Finland over the next two years to build three new data centers and expand an existing facility in Hamina, marking its largest European investment to date. The company cited Finland’s cold climate and carbon-free power as key attractions for the AI infrastructure. The announcement boosted shares of local beneficiaries such as Fortum and Nokia.

China regulator informally tightens IPO approvals for humanoid startups after Unitree volatile debut

According to sources, China’s securities regulator is informally tightening approval for initial public offerings by humanoid robotics startups. This move follows a volatile market debut by industry leader Unitree.

Over 10 US states roll back tech giants’ tax breaks worth $1B/yr amid data center backlash

More than a dozen US states have rolled back or proposed limits on tax breaks given to tech giants like Amazon, Meta, and Google for data centers, with some exemptions exceeding $1 billion annually. The backlash follows reports such as Ohio’s, where an AI-driven surge made a server tax exemption worth over $1.5 billion, prompting the governor to pause new applications.

Virgin Media offloads email services to third-party provider

Virgin Media is moving its email services to a third-party provider, Junara. Users have 45 days to save their existing mail or sign up with the new provider to retain access.

Philippines bets AI will lift GDP by 12 percent in seven years – without hurting outsourcing

The Philippines projects that AI adoption will boost its GDP by 12 percent within seven years. To support this, the country plans a 30-fold expansion of its datacenter capacity, at a cost less than what major tech companies spend in a single month. The strategy aims to grow the economy without undermining the country’s outsourcing industry.

On Really Trying (2009)

The article argues that strong conviction that a problem has a solution is essential for fueling the determination needed to solve it, while doubt or perceived impossibility undermines effort. It illustrates this using a fictional story where scientists failed to advance physics because they lacked urgent pragmatic motivation and tolerated slow progress, whereas students who knew the existing theory was wrong felt freer to pursue a breakthrough. The author observes that people are more motivated when they are certain an answer exists than when the question itself seems unanswerable.

The article’s title argues that copyright laws are more harmful than beneficial and calls for their abolition. However, the provided content contains no actual article text, only a JavaScript requirement message for the Mastodon platform.

“Tweet” and the bird logo apparently enter the public domain

A court ruled that X has not abandoned the TWITTER trademark because its Apple App Store listing uses “formerly known as Twitter” as ongoing brand use. However, the court found the TWEET mark and the Bird Logo are likely abandoned, as they no longer appear on X’s platform or app listing.

The Nonfiction Book Market is Collapsing?

Over the past decade, fiction and nonfiction have essentially flipped among top hardcover bestsellers: in 2015, only 6 of the top 25 were fiction, but in 2025, 18 were fiction and just 7 were nonfiction. Nonfiction, once seen as the more stable and predictable segment due to expert and celebrity authors, has suffered a collapse in demand. This decline stems less from media fragmentation and more from a fundamental loss of reader interest, pushing publishers toward fiction’s higher potential returns.