Hasty Briefs - Daily Tech News Brief

Bilingual

Stolen Buttons
a day ago
- The author collects 'stolen' buttons from every website they visit, presenting a stash of various UI button labels.
- The collection includes a wide range of button texts in multiple languages (English, German, Swedish) such as 'Register now', 'Log In', 'Accept', 'Shop now', etc.
- The text also includes references to a project called 'Button Stealer' and various interactive elements like 'Previous slide', 'See all symptoms', and 'Learn more'.
Computational chemistry, Locally? My rant on SSH
an hour ago
- The author works on computational chemistry for predicting protein spectra, often requiring the use of university high-performance computing clusters, which involves a cumbersome login process with VPN and 2FA that takes about 30 seconds each time, leading to decreased motivation and potential abandonment of work.
- The concept of 'clunkiness' is defined as having many intermediate steps to complete a simple task, and this friction is especially detrimental for independent researchers who work after hours and need instant, hassle-free access to computational resources.
- The author suggests that doing computational chemistry locally is becoming more feasible, as tools like Claude Code make installation easier, and local machines can efficiently handle small tasks (e.g., SCF optimization for 19 atoms) that are actually faster than cluster runs, as demonstrated by a Mac completing a task in 82 seconds versus a cluster median of 190 seconds.
Google just bet its inference future on a chip built for one model
an hour ago
- Google's Frozen v2 chip hardwires parts of Gemini's architecture but keeps weights updatable to avoid obsolescence.
- The chip aims to deliver 6-10 times more tokens per watt than current Google AI chips, helping alleviate the AI compute crunch.
- This represents a shift toward model-specific silicon, similar to the evolution of Bitcoin mining from GPUs to ASICs.
- Competitors like Taalas, d-Matrix, and SambaNova are also developing specialized hardware for AI inference.
- The approach balances hyper-efficiency of hardwired architectures with flexibility for future model updates.
Now You Can Work in C with Only BASIC Knowledge
2 hours ago
- C Source, Inc. was founded in 1983 and dissolved in 1996, with Gene W. Brown as president and Tom Hogan as a senior programmer and marketing VP.
- BASIC_C was a library for C compilers (Lattice, DeSmet, Computer Innovations) that implemented over 80 BASIC-like functions to help BASIC programmers learn C.
- The library included familiar BASIC commands like INPUT, LPRINT, LOCATE, and INKEY, and came with a well-written manual comparing BASIC and C.
- A 1984 review by Ernest Tello noted that BASIC_C eased the transition to C but warned that line-by-line translations from BASIC might not produce optimal C code.
- The library had some bugs (e.g., the COLOR function did not match IBM BASIC behavior) and lacked math functions, but offered unique functions like BLOAD, BSAVE, and SWAP.
- The product required some C knowledge and was not a full automatic converter; it was a learning aid and conversion tool for BASIC programmers.
How did the H1B visa become the 'Indian' visa
2 hours ago
- The H-1B visa was created in 1990 with a 65,000 cap, but actual annual approvals are about 4.7 times higher due to exemptions for universities, nonprofits, and renewals, totaling around 399,000 in FY2024.
- The program is dominated by Indian IT workers, who receive about 71% of H-1B approvals, while the green card per-country cap (7%) creates multi-decade waits, effectively trapping workers and enabling exploitation.
- Beyond the H-1B, larger pathways like Optional Practical Training (OPT, over 505,000 workers in FY2024) and H-4 spousal work permits expand the skilled workforce to about 1.28 million, roughly two-thirds of the U.S. software workforce.
- Wage protections are weak: the required prevailing wage is often set below market, and the $60,000 exemption for H-1B-dependent employers has not been adjusted for inflation since 1998, allowing widespread underpayment.
- Fraud and abuse are documented, including a 2008 audit finding 13.4% fraud, and recent lottery gaming reduction from 408,891 duplicate registrations to under 8,000 after a per-person rule was implemented.
- Major Indian IT firms (Infosys, TCS, Cognizant, HCL, Wipro) are the heaviest users, with documented settlements for visa fraud, below-market pay (47-64% less), and discrimination against non-Indian staff.
- The visa system fuels India's export-driven economy, with IT services exports of $224 billion and remittances of $129 billion, while a WTO dispute filed by India in 2016 remains frozen to protect its dominant share.
- The program creates winners (Indian IT firms, U.S. companies) and losers (American workers displaced, H-1B holders stuck in long green card queues), with some layoffs requiring employees to train their replacements.
Cheap Self-Hosted Kubernetes on Hetzner Cloud
an hour ago
- terraform-hcloud-kube-hetzner is an open-source Terraform project that simplifies deploying a production-ready K3s cluster on Hetzner Cloud.
- Hetzner is a German infrastructure provider offering high-performance hardware at reasonable prices, with a solid API and features like private networking, load balancers, and VMs.
- Kube-Hetzner codifies best practices for running K3s reliably on Hetzner Cloud, handling provisioning, networking, load balancers, node pools, and automated rolling upgrades.
- It uses openSUSE MicroOS, an immutable, container-centric OS with atomic updates and a read-only root filesystem for security and reliability.
- K3s is a lightweight, conformant Kubernetes distribution that packages components into a single binary, simplifying cluster operations.
- The blog plans a highly available cluster with 3 control plane and 3 worker nodes across three Hetzner zones, enabling local-path provisioning and optional Longhorn for replicated storage.
- Monthly cost estimate: ~€58 with VAT for the described setup, which is cheaper than typical managed Kubernetes services (e.g., $70-$75 on GCP/AWS).
- Cost-saving options include using a single load balancer, MetalLB, or removing public IPs via a bastion host or VPN.
- Deployment steps: clone starter project, install Terraform, hcloud CLI, kubectl; run terraform init, plan, apply; retrieve kubeconfig and test connectivity.
- Backups use K3s' etcd snapshots to S3, and node backups via Hetzner's built-in backup mechanism.
- Upgrades are handled by Rancher's System Upgrade Controller (SUC) and kured for automated node reboots with configurable maintenance windows.
FBI Forced to Reveal New Details on How It Redacted Epstein Files
3 hours ago
- FBI and DOJ trained agents to redact Epstein files, confirmed by officials after a FOIA lawsuit.
- Training videos instructed agents to redact mentions of Donald Trump using Excel spreadsheets.
- Agents worked 24-48 hour shifts reviewing hundreds of thousands of files, while Trump's name appeared over 38,000 times in released documents.
Delaware judge dismisses UnitedHealth Group lawsuit against the Guardian
2 hours ago
- A Delaware judge dismissed a defamation lawsuit by UnitedHealth Group against the Guardian over an investigation into secret payments to nursing homes.
- Judge Calvin Scott ruled that the Guardian's reporting was substantially true and that the company failed to state a viable claim.
- The Guardian's general counsel welcomed the decision as a victory for factual reporting and press freedom.
- UnitedHealth has not yet responded publicly and may appeal the ruling.
The Light Flip Is the Stylish Dumb Flip Phone of Your Dreams
4 hours ago
- Light Phone III was criticized for being too similar to a smartphone, despite limited apps.
- Light Phone introduces Light Flip, a dumb flip phone with no front screen, targeting Gen Z who want to avoid smartphone appearance.
- Existing flip phones on the market are poor quality, break easily, and lack proper network certification.
- Light Flip costs $299, with a 24-month subscription option at $39/month including service plan.
- The device uses a D-pad, T9 typing, and limited tools (Phone, Alarm, Calculator, etc.), with an SDK for third-party apps.
- Hardware includes plastic build, 2.8-inch OLED display, 12MP camera, replaceable battery, USB-C, headphone jack, but no NFC.
- Cultural backlash against smartphones and AI is strong, with events like the Summer of Ludd festival booing smartphone users.
- Commodore's Callback 8020 offers a similar retro flip phone that runs Android but blocks social media and browsers.
- Light Phone cofounders welcome more options in the market to normalize non-smartphones.
Apache Spark 4.2: Making Your Data AI‑Developer Friendly
2 hours ago
- Apache Spark 4.2 introduces Metric Views, a semantic layer for consistent business metrics across dashboards and AI systems.
- Native vector similarity functions and the NEAREST BY operator enable embedding storage and similarity queries within Spark SQL, reducing reliance on external vector databases for RAG and semantic search.
- Geospatial support is now first-class with GEOMETRY/GEOGRAPHY types and ST_* functions, simplifying location-aware AI use cases.
- PySpark enhancements include Arrow-optimized UDFs by default, modern pandas interoperability, and Real-Time Mode for millisecond-scale streaming in Python.
- Auto Change Data Capture (Auto CDC) and CHANGES clauses streamline incremental data pipelines, improving freshness and reliability for training sets.
Gumroad Says That It's Now Spending as Much on Human Employees as AI Tokens
3 hours ago
- Gumroad's human payroll dropped from $419K in June 2021 to $43K in June 2026, while AI token spend rose from $0 to $43K over the same period.
- AI now dominates engineering output, with around 150 commits compared to human contributors in the single digits.
- Customer support response time improved from 24 hours to an average of 2 minutes, driven by AI-assisted workflows.
- Gumroad maintains a small core team and emphasizes AI-assisted workflows rather than full headcount elimination.
- The trend suggests AI token spend may surpass human payroll within a few months.
Iran's IRGC claims attack on Amazon's main data hub in Bahrain
3 hours ago
- Iran's IRGC claimed its cruise missiles destroyed Amazon's central data infrastructure in Bahrain and targeted US air defense systems.
- Bahraini, US, and Amazon officials have not responded to the claim.
- The reported strikes are part of ongoing direct military confrontations between Iran and the US, which have lasted for ten consecutive days.
- Tehran warned it would target facilities in regional countries used by US forces if US attacks on Iranian infrastructure continue.
- The IRGC stated the attack was in response to US military aggression against civilian facilities in Darkhovin.
- Iran called on the IAEA and UN Security Council to condemn the US attack on the Darkhovin nuclear power plant.
- The IAEA is reviewing reports of the attack but has not issued an independent assessment.
Firefox Containers Preview
2 hours ago
- Firefox 153 introduces a native Containers Preview feature for managing separate online identities (work, shopping, personal, banking) in one browser window.
- Containers isolate cookies and ad tracking to prevent cross-container activity, stopping targeted ads from interfering across contexts.
- The feature builds on the popular Multi-Account Containers extension, integrating it directly into Firefox for easier access and support.
- Key benefits include default visibility, simplified workflow, preserved core functionality, and ongoing investment in the feature.
- In the preview, users can open tabs in specific containers, customize containers with names/colors/icons, and manage settings via Firefox Settings.
- Existing extension users can keep using the add-on alongside the native feature, as not all add-on features are yet replicated.
- Mozilla invites user feedback on the Containers Preview through Mozilla Connect to shape future development.
Meta's AI Models Are Powering the First Wave of Genesis Mission Projects
3 hours ago
- Lawrence Berkeley National Laboratory’s Advanced Light Source (ALS) produces tens of petabytes of data annually, overwhelming traditional manual analysis due to upgraded detectors capturing up to 100,000 images per second.
- The Genesis Mission, a White House initiative led by DOE, launched SYNAPS-I to use AI for real-time data analysis across X-ray and neutron science, focusing on image segmentation.
- SYNAPS-I uses Meta’s open-source SAM 3 and DINOv3 models, fine-tuned on scientific data and deployed on 300 A100 GPUs, to segment 3D volumes in about 15 minutes.
- The pipeline was demonstrated on grapevine drought response, reducing segmentation time from a month per time step to 15 minutes, enabling real-time study of biological processes.
- Meta’s open-source models allow secure deployment on government infrastructure, adapting general AI to scientific domains, and SYNAPS-I aims to create intelligent discovery platforms at national labs.
Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
4 hours ago
- Gemini introduces three new Flash models: 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber in CodeMender, targeting higher token efficiency, lower latency, and reliable performance for AI agents.
- 3.6 Flash improves coding, knowledge work, and multimodal performance, reducing output token usage by 17% compared to 3.5 Flash, with lower cost per output token.
- 3.5 Flash-Lite is the fastest model in the 3.5 series, delivering 350 output tokens per second, priced competitively for high-throughput production tasks, and outperforming prior Flash-Lite generations.
- 3.5 Flash Cyber in CodeMender is a specialized cybersecurity model for detecting and fixing vulnerabilities, exclusively available to governments and trusted partners via a limited-access pilot.
- Gemini 3.5 Pro is in testing with partners, and the team has started pre-training for Gemini 4.
- 3.6 Flash shows performance gains in benchmarks like DeepSWE (49% vs. 37%), OSWorld-Verified (83.0% vs. 78.4%), and GDPval-AA v2 (1421 vs. 1349), with enhanced safety safeguards against misuse.
- 3.5 Flash-Lite outperforms 3 Flash in agentic and coding evals, such as SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%), and supports computer use as a built-in tool.
- Both 3.6 Flash and 3.5 Flash-Lite are available starting today via Gemini API, Google AI Studio, Android Studio, Gemini Enterprise, and the Gemini app.
Why Are There No Empires in Age of Empires? (2019)
6 hours ago
- The article defines an empire as a state where a core ruling population extracts resources from a culturally distinct periphery, often by force.
- It criticizes games like Age of Empires and Civilization for portraying empires as monoethnic nation-states that destroy or replace conquered populations, missing the real mechanics of imperial diversity and resource extraction.
- Using the Roman Republic as an example, the author argues that successful empires historically derive strength from managing diverse cultures, languages, and ethnic groups, which games often oversimplify into homogeneous units.
- The piece highlights that the dominant experience of empire is that of subjugated subjects, not conquerors, a perspective games fail to represent.
- It acknowledges exceptions in games like Rise of Nations, Stellaris, and Age of Wonders that better model imperial diversity, but notes barriers like gameplay fun versus historical accuracy.
The World's 2,400 Castles
4 hours ago
- Castlemap is a free interactive map of 2,435 castles, fortresses, and palaces across 129 countries.
- The map is built entirely from open data: Wikidata, Wikimedia Commons, Wikipedia, OpenFreeMap, and Natural Earth.
- It curates only verified landmarks with real photos, English Wikipedia articles, and exact coordinates, balancing countries for explorability.
- Users can add missing castles by improving their Wikidata entries (coordinates, photo, English article) for the next refresh.
- The full dataset is free to download as GeoJSON or CSV under CC0.
- France has the most castles (183), followed by Italy, Germany, England, and Japan.
- Each castle has a 'fame score' based on Wikipedia language editions; Versailles ranks #1.
- The atlas is curated, not exhaustive, focusing on significance over quantity.

Hasty Briefsbeta

Stolen Buttons

Computational chemistry, Locally? My rant on SSH

Google just bet its inference future on a chip built for one model

Now You Can Work in C with Only BASIC Knowledge

How did the H1B visa become the 'Indian' visa

Cheap Self-Hosted Kubernetes on Hetzner Cloud

FBI Forced to Reveal New Details on How It Redacted Epstein Files

Delaware judge dismisses UnitedHealth Group lawsuit against the Guardian

The Light Flip Is the Stylish Dumb Flip Phone of Your Dreams

Apache Spark 4.2: Making Your Data AI‑Developer Friendly

Gumroad Says That It's Now Spending as Much on Human Employees as AI Tokens

Iran's IRGC claims attack on Amazon's main data hub in Bahrain

Firefox Containers Preview

Meta's AI Models Are Powering the First Wave of Genesis Mission Projects

Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Why Are There No Empires in Age of Empires? (2019)

The World's 2,400 Castles