2025-10-02 Briefing - alobbs.com

Today’s top news includes OpenAI’s $500 billion valuation boost after a record share sale, Meta’s new Communities feature on Threads to boost engagement, and a critical vulnerability in Red Hat OpenShift AI risking full cluster takeover. Additionally, the US government faces disruptions due to the shutdown, and California enacts pioneering responsible AI regulation.

▶️ Internet Infrastructure

SSH3: Fast, Secure, and Flexible SSH over HTTP/3 with Enhanced Features

SSH3 maps SSH semantics onto HTTP/3, enabling faster session establishment, advanced authentication methods, UDP forwarding, and server hiding, with technical details in arXiv and IETF draft.

SSH3 implements a secure shell protocol over HTTP/3, utilizing QUIC+TLS1.3 for session establishment and HTTP authorization for user authentication.
Introduces faster session setup (3 RTTs vs 5-7 RTTs in SSHv2), supports OAuth 2.0, OpenID Connect, UDP port forwarding, and server invisibility via secret URL paths.
Provides compatibility with OpenSSH features, including authorized_keys parsing, agent forwarding, proxy jump, and certificate-based authentication, while being in early experimental stage.

Google’s cdc-file-transfer Tools Boost Windows-Linux File Sync and Streaming

Google’s cdc-file-transfer tools enable efficient Windows-to-Linux file sync and streaming using CDC algorithms, supporting cross-platform operation and delta transfer.

The repository provides tools for file synchronization and streaming from Windows to Linux, based on Content Defined Chunking (CDC) and FastCDC algorithms.
cdc_rsync enables delta updates by content-defined chunking, achieving up to 30x faster sync than traditional rsync, with support for Windows, Linux, macOS, and ARM64.
cdc_stream offers read-optimized streaming with CDC-based diffing, providing 2x to 5x speedup over sshfs, but does not support write-back from Linux to Windows.

Firefox’s 2025 UDP I/O overhaul boosts speed and security with Rust and QUIC

Max Inden’s 2025 project replaced Firefox’s UDP I/O with modern system calls using Rust and quinn-udp, improving throughput, enabling QUIC ECN, and addressing platform-specific challenges.

Project in mid-2024 to replace Firefox’s UDP I/O with modern system calls across Windows, MacOS, Linux, and Android
Utilized Rust and quinn-udp library to enhance security and performance, achieving up to 4 Gbit/s in CPU-bound benchmarks
Implemented batching, segmentation offloading, and ancillary data support, including QUIC ECN, on multiple platforms

Ollama Launches Web Search API for Real-Time Data Access in AI Models

Ollama introduced a web search API on September 24, 2025, enabling models to access real-time web data via REST, Python, and JavaScript, enhancing accuracy and reducing hallucinations.

Ollama launched a web search API on September 24, 2025, offering a free tier with higher rate limits via Ollama’s cloud
Web search is accessible as a REST API and through Python and JavaScript libraries, enabling models like gpt-oss to perform long-running research
Example API request: curl https://ollama.com/api/web_search with query parameter, returning results with titles, URLs, and content summaries

Standard Intelligence Builds 30PB Storage Cluster in SF for Under $500K

Standard Intelligence built a 30 PB data storage cluster in SF for $426,500, reducing costs by 40x compared to AWS, using simple Rust software, 2,400 HDDs, and high-speed networking.

Built a 30 PB storage cluster in downtown San Francisco for under $500,000, costing approximately $29,500/month including depreciation.
Storage costs are significantly lower than cloud alternatives: $29.5k/month vs $1.13M (AWS) and $270k (Cloudflare R2).
The cluster stores 90 million hours of video data for pretraining models, requiring 500x more storage than text-based LLMs like LLaMa-405B.

Meta Threads Launches Communities to Boost Engagement with Over 100 Interest Spaces

Meta’s Threads app launches Communities, creating over 100 interest-based spaces with custom emojis, differentiating from X’s user-moderated model, to deepen user engagement and content personalization.

Meta’s Instagram Threads app introduces a new Communities feature with over 100 communities for topics like basketball, K-pop, and books.
Unlike X’s Communities, Threads’ communities are created and managed by Meta, with non-members able to participate and display community membership on profiles.
Communities feature includes custom “Like” emojis per community, integration with user interests, and plans for profile badges for active builders.

Cl0p-Linked Hackers Breach Oracle E-Business Suite, Demand $50 Million Ransom

A ransomware group linked to Cl0p has compromised Oracle E-Business Suite, demanding up to $50 million, exploiting vulnerabilities and stealing data from core enterprise applications.

Hackers claiming affiliation with Cl0p breached Oracle’s E-Business Suite, targeting financial, supply chain, and CRM operations.
Extortion emails began around September 29, demanding up to $50 million in ransom, with proof of data theft provided to victims.
The group exploited vulnerabilities in Oracle’s E-Business Suite, using compromised user emails and default password-reset functions; at least one company confirmed data theft.

Critical Red Hat OpenShift AI Bug Enables Full Cluster Takeover

A 9.9 CVSS critical bug in Red Hat OpenShift AI (CVE-2025-10725) permits privilege escalation from authenticated low-privilege users to cluster root, risking full platform compromise.

Critical severity bug (CVSS 9.9) in Red Hat OpenShift AI allows remote privilege escalation with minimal authentication
An attacker with a standard authenticated account, such as a data scientist, can escalate privileges to full cluster administrator
Exploit enables full cluster takeover, data theft, service disruption, and infrastructure control; Red Hat advises immediate patching and least-privilege permissions

OpenStack Flamingo Boosts Performance and Expands AI and HPC Capabilities

OpenStack’s Flamingo release (October 2025) removes eventlet dependencies, enabling multi-threaded Ironic, introduces new hardware support, and aims to expand platform adoption for AI, HPC, and VMware alternatives.

OpenStack released its 32nd major version, “Flamingo,” on October 1, 2025
Major focus on removing eventlet dependencies to address technical debt, enabling multi-threaded operation in Ironic
Over 8,000 changes contributed by approximately 480 organizations, including Ericsson, Rackspace, Red Hat, Walmart, BBC R&D, Samsung SDS, SAP, Nvidia
Notable features include one-time passthrough in Nova for dedicated hardware access, support for AMD SEV, enhanced Skyline dashboard with rescue, network security, and UI improvements
Manila now supports BYOK encryption and renamed EMC Isilon driver to PowerScale, supporting thin provisioning
Flamingo aims to expand OpenStack deployment beyond 55 million cores, targeting VMware alternatives, AI, ML, and HPC workloads
Next release scheduled for April 2026, named “OpenStack 2026.1 (Gazpacho),” as Flamingo is a non-SLURP release that can be skipped without hindering future upgrades

BT Aims for 99% UK 5G SA Coverage by 2030 with Ericsson Tech

BT plans to provide 5G SA coverage to 99% of the UK by 2030 using Ericsson massive MIMO radios and small cells, rebranding it as “5G+” for clearer customer communication.

BT aims to deliver 5G Standalone (5G SA) coverage to 99% of the UK population by 2030, ahead of competitors.
The rollout includes Ericsson AIR 3284 5G triple-band FDD massive MIMO radios and over 1,500 outdoor small cells.
BT’s mobile subsidiary, EE, plans to rebrand 5G SA as “5G+” to emphasize its benefits; the coverage expansion is part of a government-backed goal.

Crimson Collective Claims Data Breach at Red Hat Exposes Sensitive Customer Info

Crimson Collective claims to have stolen 570GB of Red Hat data from 28,000 repositories, including sensitive CERs with infrastructure details, and used authentication tokens to compromise customers, with no official Red Hat response yet.

A hacking group, “the Crimson Collective,” claims to have accessed over 28,000 Red Hat internal repositories, exfiltrating approximately 570GB of compressed data including sensitive customer documents.
The stolen data includes Customer Engagement Reports (CERs) containing architecture diagrams, configuration details, authentication tokens, and network maps, spanning 2020–2025 and involving major sectors like banking, telecoms, and government.
The attackers published file listings, samples, and asserted they used authentication tokens to compromise downstream Red Hat customers; Red Hat has not confirmed a breach or responded to inquiries.

Ionos to Raise Plesk License Fees by £5 Monthly from October 2025

Ionos is increasing Plesk license fees by £5/month per license due to partner cost hikes, giving customers one month’s notice to accept or disable Plesk, amid complaints over short notice and fairness.

Ionos announced a Plesk license fee increase of £5 (excluding VAT) per license, effective October 30, 2025
The fee applies monthly per server, impacting customers with multiple server instances
Customers can choose to disable Plesk or be automatically billed; removal of Plesk may complicate server management

Electron update fixes macOS Tahoe slowdowns by removing shadow override

Electron fixed performance issues on macOS 26 “Tahoe” by removing a private API override that caused shadow recalculations, but users still face slowdowns until apps update to Electron 26.01.

Electron fixed code causing system-wide slowdowns on macOS 26 “Tahoe” by removing an override of the private Cocoa API method _cornerMask.
The override caused repeated recalculation and repainting of window shadows, degrading performance, especially with multiple Electron applications running.
The fix, committed last week, updates Electron to version 26.01, but application developers must rebuild and deploy to benefit; many Electron apps remain outdated.

OpenAI Teams Up with Samsung and SK Hynix for Massive AI Data Center Expansion

OpenAI partners with Samsung and SK Hynix to supply 900,000 DRAM wafer starts monthly for its $500 billion Stargate AI infrastructure buildout, supporting global datacenter expansion.

OpenAI announced Samsung and SK Hynix will supply memory chips for its Stargate AI infrastructure project.
Samsung and SK Hynix will produce approximately 900,000 DRAM wafer starts per month to support large language models.
The agreements include collaboration on local AI datacenters, with additional talks involving SK Telecom, Samsung C&T, Samsung Heavy Industries, and Samsung SDS.

Kodex Global’s domain was frozen by AWS after a social engineering attack exploiting AWS as registrar, preventing access to its subpoena management platform used by 15,000+ agencies and companies.

Kodex Global’s domain was frozen on October 1 between 08:54-12:47 EDT due to a social engineering attack exploiting AWS as domain registrar
Attackers attempted to transfer the domain to another registrar; ownership was not transferred, but AWS improperly froze the domain based on a fraudulent legal order
No credentials, customer data, or internal systems were compromised; AWS responded quickly and took steps to prevent recurrence

▶️ Open Source

F3: The Future-Ready Open-Source Data Format with WebAssembly Decoders

F3 is an open-source, extensible data file format embedding WebAssembly decoders for platform-independent, future-proof data sharing, with optimized layout and encoding for modern analytics workloads.

F3 is a next-generation open-source data file format emphasizing interoperability, extensibility, and efficiency, designed to replace legacy formats like Parquet and ORC.
Each self-describing F3 file embeds data, metadata, and WebAssembly (Wasm) decoders, enabling platform-agnostic decoding with minimal storage overhead.
F3 supports state-of-the-art encoding, cascading compression, vectorized decoding, and separates I/O, encoding, and dictionary units for optimized data layout.

Redis Outperforms Postgres in Caching Benchmarks on Homelab Kubernetes

Redis demonstrated significantly higher performance than Postgres for caching, with 7425 req/sec and near-zero latency, but Postgres remains viable due to its integration and scalability, despite lower throughput.

The experiment compared Redis (redis:8.2) and Postgres (postgres:17.6) on a homelab Kubernetes cluster with 2 CPUs and 8GiB RAM per node
Both caches stored 30 million entries; benchmarks included get, set, and mixed workloads, each running for 2 minutes
Redis outperformed Postgres in request throughput and latency, with Redis achieving approximately 7425 requests/sec and near-zero latency, while Postgres was bottlenecked by CPU and used up to 6GiB RAM

Immich v2.0.0 Launches with Major Stability and New Features

Immich announced version 2.0.0 as a stable release, marking a major milestone with improved stability, performance, and community-driven features, including physical media and future backup enhancements.

Immich v2.0.0 was officially released as a stable version on October 1, 2025, after 1,337 days of development, 271 releases, and community support.
The release resolves significant technical debt, improves compatibility, and removes the website warning banner; Docker images are available shortly after announcement.
New features include a physical distribution option via a retro CD/DVD, a redesigned website, and plans for enhanced backup services, usage data collection, and feature parity across platforms.

Chess Position Move Limit Confirmed at 218 Moves Using Mathematical Modeling

Using advanced mathematical techniques and computational modeling, the article proves that the maximum number of legal moves in any reachable chess position is 218, confirming Petrović’s 1964 record.

The article proves that the maximum number of legal moves in a reachable chess position is 218.
The author used mathematical modeling, including integer programming and fractional decision techniques, to establish the upper bound.
The study confirms that no reachable position exceeds 218 moves, aligning with a 1964 composition by Nenad Petrović.

Google Launches Jules Tools CLI and API for Seamless AI Coding Integration

Google’s Jules AI coding agent now features Jules Tools CLI and public API, enabling integration into terminals, CI/CD pipelines, and IDEs, with scoped task execution and Gemini 2.5 Pro support.

Google introduced Jules Tools, a command-line interface (CLI) and public API, to embed Jules into terminals, CI/CD systems, and tools like Slack.
Jules, an AI coding agent, now supports scoped tasks with less interactivity, executing independently after user approval, and is built on Google’s Gemini 2.5 Pro AI model.
The API extension aims to integrate Jules into developers’ workflows and IDEs, with plans for specific IDE plug-ins; it previously operated solely via website and GitHub.

OpenAI’s recent share sale valued the company at $500 billion, with $6.6 billion in stock sold, elevating its status above SpaceX and reflecting significant investor confidence.

OpenAI completed a share sale at a $500 billion valuation, surpassing SpaceX as the world’s largest startup
About $6.6 billion of stock was sold by current and former employees to investors including Thrive Capital, SoftBank, Dragoneer, MGX, and T. Rowe Price
The valuation increased from previous $300 billion during a SoftBank-led funding round earlier this year

Apple AirPods 2025 Review: Pro 3, AirPods 4, Max and Budget Options

The article compares all Apple AirPods models in 2025, highlighting Pro 3’s superior ANC and fitness features, AirPods 4’s open fit and affordability, and Max’s premium sound, with detailed performance and pricing data.

The article reviews and tests all current Apple AirPods models as of October 2025, including Pro 3, AirPods 4, AirPods Max, and budget options.
The AirPods Pro 3 feature improved active noise cancellation (ANC), extended battery life (8 hours with ANC), in-ear design, built-in heart rate sensors, and support for Apple’s Live Translation.
The AirPods 4 with ANC offer semi-open fit, solid sound quality, and good noise cancellation, priced at $179, with a wired charging case; the standard AirPods 4 without ANC cost $129.
The AirPods Max provide high-end over-ear audio with USB-C charging, support for lossless wired audio, and advanced computational sound, priced at $549.
Testing involved real-world environments, diverse audio sources, and evaluation of design, fit, sound quality, features, and noise cancellation performance.

Anthropic Claude-Code 2.0.5 Launches as Terminal-Based Coding Assistant

@anthropic-ai/claude-code 2.0.5 is a terminal-based coding assistant enabling natural language commands for code management, with 5.8 million weekly downloads and 78.1 MB size.

@anthropic-ai/claude-code version 2.0.5 was published 3 hours ago
Provides an agentic coding tool that operates in terminal, IDE, or via GitHub, understanding codebases and executing routine tasks through natural language
Installation command: npm install -g @anthropic-ai/claude-code; weekly downloads: 5,805,186; unpacked size: 78.1 MB; total files: 52

iFixit Finds Apple AirPods Pro 3 Are Unrepairable and Glued Shut

iFixit’s teardown shows AirPods Pro 3 are glued, unrepairable, with non-replaceable batteries, despite features like IP57 water resistance, active noise cancellation, and eight-hour battery life.

iFixit’s teardown of Apple’s AirPods Pro 3 reveals a device with glued, unfixable components, making repairs nearly impossible
The earbuds feature IP57 water resistance, active noise cancellation, and up to eight hours of listening time per charge
Batteries are non-replaceable; the casing and internal components are glued and soldered, preventing end-user repairs

AI-Driven Vulnerability Detection in Curl Highlights Human-Guided Security Benefits

AI tools, when guided by skilled human analysts like Rogers, can effectively identify genuine vulnerabilities in software such as curl, highlighting AI’s positive role in security.

Over two years, the curl project received numerous AI-generated bug reports, many invalid, prompting project maintainer Daniel Stenberg to address issue quality.
In September 2025, security researcher Joshua Rogers used AI scanning tools (including ZeroPath, Almanax, Corgea, and Amplify) to identify 50+ valid vulnerabilities in curl, including an out-of-bounds read in Kerberos5 FTP.
Rogers’ findings, validated by Stenberg, demonstrate AI tools can effectively discover real vulnerabilities when used by experienced humans, contrasting with the widespread problem of AI-generated spam reports in open source projects.

Sammyuri Creates Minecraft-Based Chatbot CraftGPT Using Redstone and Python

Sammyuri built CraftGPT, a 5.1-million-parameter Minecraft-based chatbot using redstone, Python, and MCHPRS, with slow response times and limited context, demonstrating complex hardware in Minecraft.

Sammyuri built CraftGPT, a small language model with 5,087,280 parameters, trained on TinyChat dataset, coded in Python, with a vocabulary of 1,920 tokens across six layers
The complete Minecraft setup measures 1,020 x 260 x 1,656 blocks, built using redstone, requiring the Distant Horizons mod for full display
CraftGPT is implemented using vanilla redstone mechanics, with response times of approximately two hours at 40,000x speed via MCHPRS, and a context window of 64 tokens; responses are prone to gibberish and off-topic outputs

Minecraft Creator Builds 5.1M Parameter Language Model Using Command Blocks

A Minecraft creator built a command block-free, 5.1 million-parameter language model trained on TinyChat, capable of short conversations, with weights quantized for efficient in-game implementation.

Built a small language model in Minecraft with 5,087,280 parameters, trained in Python on the TinyChat dataset of basic English conversations
Model features an embedding dimension of 240, vocabulary of 1920 tokens, and 6 layers; context window size is 64 tokens
Most weights are quantized to 8 bits; embedding weights at 18 bits, LayerNorm weights at 24 bits; build volume is 1020x260x1656 blocks; response time is approximately 2 hours with increased tick rate using MCHPRS

▶️ Software Development

AI Coding Agents as Junior Engineers: A New Playbook for Sustainable Development

AI coding agents act as lightning-fast junior engineers, necessitating a new playbook that integrates best practices across the development lifecycle to prevent the AI coding trap.

AI coding agents like Claude Code accelerate code generation but require human review, testing, and integration due to limited context retention.
Current AI coding tools produce code comparable to mid-level engineers but lack true learning capacity; improvements rely on context engineering and new models.
Effective AI-driven development involves applying best practices across the software lifecycle—specification, documentation, modular design, testing, standards, and monitoring—to ensure sustainable delivery.

Comprehension Debt from LLM Code Increases Developer Modification Time

Growing “comprehension debt” from unchecked LLM-generated code increases future modification time, as developers often need to manually understand and rework code that LLMs cannot fix.

Increasing developer time to modify or fix LLM-generated code due to comprehension challenges
“Comprehension debt” arises when teams check in unread, minimally tested code, leading to future modification difficulties
Experiments show frequent “doom loops” where LLMs cannot resolve code issues, requiring manual edits and understanding

Anthropic Unveils Claude Sonnet 4.5 as Top AI Coding System

Anthropic launched Claude Sonnet 4.5, claiming it leads in AI coding, with enhanced reliability, autonomous operation up to 30 hours, and improved practical business applications.

Anthropic released Claude Sonnet 4.5, claiming it is the world’s top AI coding system
The model can operate autonomously for up to 30 hours, over four times longer than Opus 4
Sonnet 4.5 achieves state-of-the-art results on SWE-Bench Verified and improves computer usage capabilities

▶️ Management and Leadership

AMD Athlon 64 Launch Sparks Intel Cloning and CPU Revolution

AMD’s Athlon 64, launched on September 23, 2003, introduced 64-bit x86 architecture with full backward compatibility, forcing Intel to clone AMD64 and reshape CPU development.

AMD released the Athlon 64 CPU on September 23, 2003, marking a pivotal shift in x86 architecture.
The Athlon 64 was fully backward compatible with 32-bit x86 applications, enabling a gradual transition to 64-bit computing.
Intel responded by cloning AMD64 technology in 2004, creating Intel64, and discontinued Itanium in 2020.

Evanston mandated Flock Safety to remove license plate cameras after unauthorized reinstallation, with cameras possibly active post-shutdown, amid legal disputes over data sharing and contract termination.

Evanston ordered Flock Safety to remove all license plate cameras after reinstallation without city approval.
Flock reinstalled 15 of 18 stationary cameras by Sept. 8, including some outside city limits, contrary to prior removal orders.
The city issued a cease-and-desist on Sept. 26; Flock had previously been ordered to shut down 19 cameras and terminate its contract due to unauthorized data sharing with ICE and CBP.

SimpleFold: A 3B-Parameter Transformer Model for Protein Folding

SimpleFold is a transformer-based protein folding model with 3B parameters, trained on 8.6M+ structures, achieving competitive accuracy and enabling ensemble predictions, challenging complex architecture reliance.

SimpleFold is a flow-matching protein folding model using only transformer layers, trained via a generative flow-matching objective.
Scaled to 3 billion parameters, trained on over 8.6 million distilled protein structures plus experimental PDB data.
Achieves competitive performance on standard folding benchmarks and demonstrates strong ensemble prediction capabilities.

Most Users Use 20% of Features; Focused Apps Target Neglected User Needs

Most users only use about 20% of features, each different; successful software targets these neglected slices, reducing bloat and enhancing user-specific experiences.

Most users utilize approximately 20% of an application’s features, with each user engaging with different 20% segments.
Users often resent unused features that cause bloating, slow performance, and workflow disruption, exemplified by Office updates and Google Search.
Companies like Kagi, Figma, Notion, VS Code, Slack, and Discord succeed by targeting specific neglected user slices, emphasizing customizable, lightweight, and focused tools.

Typst: A Rust-Based LaTeX Alternative with Simplified Syntax and Advanced Features

Typst, a Rust-based, open-source typesetting system, aims to replace LaTeX with simpler syntax, faster compilation, and advanced features like live preview and PDF embedding, supporting modern workflows.

Typst is a Rust-based, open-source document typesetting system designed as a simpler, faster LaTeX replacement, especially suited for technical material.
It offers a less verbose, Markdown-like markup syntax, improved error messages with colored glyphs, and supports PDF, SVG, PNG, and HTML outputs.
Typst features include font management, live preview with typst watch, and recent inclusion of PDF file embedding (release 0.14), with ongoing development of tagged PDF for accessibility.

Use Google Sheets for Simple, Flexible Business Solutions

The author recommends using Google Sheets as an efficient, adaptable solution for small business tasks, emphasizing iterative development and problem scope understanding over complex software.

Advocates using Google Sheets as the simplest solution for various business problems
Provides examples: admin panel, quote system, and CRM, all replaced or simplified by Google Sheets
Emphasizes iterative approach: start with basic solutions to understand problem scope before developing complex tools

UK Government Announces Digital ID System for Secure Public Services and Immigration Control

UK Government announced plans to implement a digital ID system by the end of Parliament, aiming to enhance service access, security, and compliance, with public consultation and legislation forthcoming.

UK Government plans to introduce a digital ID system to improve access to public services, combat illegal migration, and increase security, announced on October 2, 2025
Digital ID will be a free, non-card digital identity for UK citizens and residents aged 16+ (considering lowering to 13), building on GOV.UK One Login and GOV.UK Wallet
Employers will be required to verify digital IDs for ‘right to work’ checks by the end of the current Parliament, aiding enforcement against illegal employment and streamlining compliance

Gmail Drops Gmailify and POP Support in 2026

Gmail will discontinue support for Gmailify and POP in January 2026, encouraging users to use IMAP and Gmail app integrations for third-party email access.

Starting January 2026, Google will cease support for Gmailify and POP in Gmail.
Gmailify will no longer offer features like spam protection, inbox categorization, or enhanced notifications.
POP access will be deprecated; “Check mail from other accounts” will be unavailable in Gmail on desktop.
Users can still read and send emails via IMAP in Gmail and add third-party accounts in the Gmail app.
Administrators can assist with migrating work or school account data through Google Workspace migration services.

Pop!_OS 24.04 LTS Beta Launches with COSMIC Desktop and New Features

Pop!_OS 24.04 LTS Beta with COSMIC DE is available for testing, featuring Linux kernel 6.16.3, Mesa 25.1.5-1, NVIDIA Driver 580, and redesigned desktop environment, with specific hardware and installation requirements.

Pop!_OS 24.04 LTS Beta with COSMIC DE is available for testing, featuring Linux kernel 6.16.3, Mesa 25.1.5-1, and NVIDIA Driver 580.
The beta requires 4 GB RAM, 16 GB storage, and a 64-bit processor; disables Secure Boot for installation.
Includes redesigned COSMIC Desktop Environment, replacing GNOME apps with COSMIC equivalents, and introduces COSMIC Store and Epoch 1 Beta.

Why Good Workplace Politics Are Vital for Engineers’ Success

The article argues that politics is essential for engineers to influence decisions, emphasizing strategic relationship management, understanding incentives, and visibility as key skills for organizational success.

The article challenges the notion that workplace politics is inherently negative, emphasizing that bad politics is the real issue.
Good politics involves strategic relationship-building, understanding incentives, managing up, creating win-win scenarios, and visibility.
Effective political skills are essential for technical decision-making, influencing outcomes, and organizational success, countering the myth of engineering objectivity.

US Transit Bus Costs Soar Due to Limited Competition and Buy America Rules

A 2025 study highlights US transit agencies’ inflated bus costs due to limited competition, excessive customization, and Buy America restrictions, with proposals to standardize models and increase foreign competition.

In 2023, RTD’s 10 diesel buses cost $432,028 each; SORTA’s 17 buses cost $939,388 each, despite being similar models from the same manufacturer
US transit agencies paid significantly higher prices for buses compared to Singapore’s $333,000 electric buses and Korean Hyundai’s $350,000 electric buses sold overseas
A research paper attributes high costs to lack of competition, excessive vehicle customizations, a domestic supplier shortage, and Buy America rules restricting foreign vehicle sales

Tesla’s Executive Turnover Reflects Challenges Amid Delivery Slumps

Tesla’s organizational chart shows Musk’s 19 direct reports as of August, with significant executive turnover over the past year due to delivery declines and EV incentive cuts.

Elon Musk had 19 direct reports at Tesla as of August, down from 35 in 2024 and 30 in late 2021
Over the past year, at least 10 executives, including key leaders like Troy Jones and Omead Afshar, have left Tesla
Most of Musk’s reports are based in San Francisco; recent departures include HR and legal officers, amid slumping delivery numbers and reduced EV incentives

Imgur withdraws UK services after ICO data breach warning

The ICO issued a notice of intent to fine Imgur after provisional findings of data protection breaches related to children’s data, leading Imgur to withdraw its UK services.

Imgur ceased UK services following ICO’s provisional findings and notice of intent to fine over children’s data handling concerns.
The ICO investigation, launched earlier this year under Children’s Code strategy, found potential data protection infringements.
Imgur, founded in 2009 and acquired by MediaLab AI Inc in 2021, is a popular meme and viral content image hosting platform with over 130 million users.

UK to Implement Mandatory Digital ID Cards for Workers by 2025

The UK government announced mandatory digital ID cards for workers by 2025 to curb illegal migration and employment, using mobile-based IDs modeled on European and international systems.

UK government plans to mandate digital ID cards for all workers by the end of the current parliamentary term, targeting to combat illegal migration and illegal employment.
Digital IDs will be stored on mobile phones and integrated into employer verification processes, with future use for accessing services like childcare, welfare, and tax records.
The scheme draws on digital ID systems from Estonia, Denmark, Australia, and India; over 57% of Britons support national ID cards, despite privacy concerns.

Mastering Technical Taste in Software Engineering for Project Success

Technical taste in software engineering involves selecting engineering values suited to the project, with maturity shown by flexibility in tradeoffs; it is developed through diverse project experience.

Technical taste differs from technical skill; it involves adopting engineering values fitting the current project
Key engineering values include resiliency, speed, readability, correctness, flexibility, portability, scalability, and development speed
Bad taste stems from inflexibility and prioritizing personal preferences over project context; good taste involves selecting appropriate values for specific problems

Meta’s AI Research Faces Internal Clash Over Publication Restrictions

Meta’s AI group faces internal conflict after policy changes restrict research publication, causing tensions between traditional openness and new leadership’s cautious approach.

Meta’s AI research publishing policies changed, causing internal tensions between the “old guard” and new leadership.
The shift involves restricting the publication of certain research, leading to disagreements over transparency and openness.
The controversy has sparked a stir within Meta’s AI group, highlighting conflicts over research dissemination practices.

Microsoft Enables Employees to Use Personal Microsoft 365 Subscriptions for Work Copilot

Microsoft’s bring your Copilot to work initiative permits employees to use personal Microsoft 365 subscriptions for Copilot features at work, bypassing traditional IT restrictions, with controls for admins and data protection assurances.

Microsoft allows users to bring personal Microsoft 365 subscriptions to work to access Copilot features, bypassing traditional IT controls.
This initiative enables employees to use Copilot on work documents via personal accounts, even without a corporate Copilot license.
IT administrators can restrict personal Copilot usage, audit interactions, and enforce enterprise permissions; government tenants (GCC/DoD) lack this control.
Microsoft states that user permissions govern data access, and enterprise data remains protected, with full admin oversight maintained.
The program aims to promote AI adoption while maintaining enterprise control, with usage data captured and potentially influencing enterprise AI metrics.

Salesforce’s AI Search Upgrade Sparks User Frustration

Salesforce replaced basic search with Agentforce AI on help pages, causing user dissatisfaction due to unreliable results and longer searches, amid broader AI-driven platform and pricing strategies.

Salesforce replaced search functions on help pages with Agentforce AI on September 29, leading to user complaints
Users report unreliable results, longer search times, and difficulty locating information
Salesforce has not responded publicly; the change appears to be part of broader AI integration efforts, including the Atlas Reasoning Engine

US Government Shutdown Disrupts IT Projects and Weakens Cybersecurity

The October 1, 2025, US government shutdown halts non-essential IT modernization, delays critical projects, and weakens cybersecurity defenses, risking long-term operational and security setbacks.

US government shutdown began at 12:01 ET on October 1, 2025, halting non-essential IT modernization and reducing cybersecurity operations to skeleton crews
The shutdown disrupts ongoing IT projects, including infrastructure upgrades, cloud migrations, and system updates, with delays risking long-term setbacks
Federal agencies continue cybersecurity monitoring and critical systems with reduced staffing, increasing security vulnerabilities amid rising threats

Lloyds to Close 136 Branches Amid Digitization and AI Innovation

Lloyds Banking Group announced ongoing branch closures driven by digitization, investing £3 billion over three years, with plans to close 136 branches by March 2026, while developing AI tools and maintaining over 1,000 branches.

Lloyds Banking Group plans to close more branches, citing digitization as a key driver
Invested £3 billion over three years and £4 billion over five years in technology and cybersecurity
Has operated more than 1,000 branches in the UK, with plans to close 136 by March 2026; also experimenting with AI-powered wealth management tools and sandboxing AI money management in collaboration with FCA

BCG Warns of Rising SaaS Costs and the Need for New Procurement Strategies

BCG reports SaaS-driven software costs have surged, with IT spending reaching $5 trillion (2019-2025), prompting a need for new procurement strategies to reduce complexity and manage consumption-based pricing.

Boston Consulting Group (BCG) warns organizations to rethink software procurement amid SaaS market growth
IT spending increased 6% from 2019 to 2025 to approximately $5 trillion; software expenditure rose from 13% in 2024 to 21% in 2024, driven by SaaS
Industry shifts include new vendors, open source solutions, transition from on-premises to SaaS, new pricing models, and vendor consolidation, complicating procurement and increasing costs

UK DWP awards IBM £27m contract for responsible AI welfare systems

UK’s DWP hired IBM for up to £27 million to develop AI solutions under Nexus AI, aiming for responsible, human-centered deployment amid concerns over bias and privacy in welfare automation.

UK Department for Work and Pensions awarded IBM a contract up to £27 million to develop and support AI systems for service enhancement.
The contract, under the Nexus AI program, includes exploration, deployment, and ongoing support, initially valued at £9 million for one year with potential extensions.
The department emphasizes a responsible, human-centered AI approach, focusing on value-led, secure, and ethical AI adoption, while concerns about bias and privacy in welfare AI systems persist.

Windows 10 retains 40.5% market share ahead of October 14, 2025, end of free support, risking security vulnerabilities; Microsoft’s migration efforts have fallen short despite petitions for extended updates.

Windows 10 accounts for 40.5% of the Windows desktop market as of October 2025, with Windows 11 at 48.94%
Support for Windows 10 ends on October 14, 2025, ceasing free security updates despite 1.5 million sites tracking data
Microsoft’s efforts to migrate users to Windows 11 have been insufficient; market share has declined from 62.75% (Windows 10) and 33.42% (Windows 11) a year prior

▶️ Technology

iPad17,3 Scores 4133 Single-Core and 15437 Multi-Core on Geekbench 6.5

iPad17,3 achieved 4133 (Single-Core) and 15437 (Multi-Core) scores on Geekbench 6.5.0, featuring an ARM CPU with 9 cores at 4.42 GHz, 11.20 GB RAM, running iOS 26.0.

iPad17,3 scored 4133 in Single-Core and 15437 in Multi-Core Geekbench 6.5.0 benchmarks on September 30, 2025
System features an iOS 26.0 OS, 11.20 GB RAM, ARM CPU with 1 processor, 9 cores, 4.42 GHz base frequency, and 6 MB L2 cache
CPU includes instruction sets: neon, aes, sha1, sha2, neon-fp16, neon-dotprod, i8mm, sme-i8i32, sme-f32f32, sme2

Google Launches Gemini 2.5 Flash Models with Enhanced Efficiency and Capabilities

Google released Gemini 2.5 Flash and Flash-Lite models with improved quality, efficiency, and multimodal capabilities, emphasizing better instruction following, reduced token costs, and enhanced agentic tool use, accessible via preview model strings.

Google released updated Gemini 2.5 Flash and Flash-Lite models on September 25, 2025, available via Google AI Studio and Vertex AI.
Gemini 2.5 Flash-Lite improves instruction following, reduces verbosity, and enhances multimodal and translation capabilities; model string: gemini-2.5-flash-lite-preview-09-2025.
Gemini 2.5 Flash offers better tool use, 5% performance gain on SWE-Bench Verified, and increased cost-efficiency with fewer output tokens—50% reduction for Flash-Lite, 24% for Flash; model string: gemini-2.5-flash-preview-09-2025.

DeepSeek-V3.2-Exp Boosts Long-Context Efficiency with Sparse Attention

DeepSeek-V3.2-Exp, released in 2025, enhances long-context efficiency via DeepSeek Sparse Attention, matching V3.1-Terminus performance across benchmarks, with open-source CUDA kernels and inference tools.

DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention (DSA), achieving fine-grained sparse attention to improve long-context training and inference efficiency.
DSA maintains model output quality with performance aligned to DeepSeek-V3.1-Terminus across benchmarks such as Reasoning Mode and Tool Use tasks.
Benchmark results show comparable or improved scores: e.g., Reasoning Mode (V3.2-Exp: 85.0 in MMLU-Pro, 89.3 in AIME 2025), Tool Use (V3.2-Exp: 40.1 in BrowseComp, 97.1 in SimpleQA).

Hazy Research Unveils High-Throughput Llama-70B Megakernel for H100 GPUs

Hazy Research developed a tensor-parallel Llama-70B megakernel for high-throughput inference on H100s, achieving >22% performance gains over SGLang by overlapping compute, memory, and communication across GPU hierarchy.

Released a throughput-optimized megakernel for tensor-parallel inference with Llama-70B on NVIDIA H100 GPUs, capable of aggressive resource overlap.
The megakernel outperforms SGLang by over 22% in end-to-end throughput (measured as time to process 65,536 prompts from ShareGPT benchmark).
Implements a flexible instruction-and-interpreter model enabling fine-grained overlapping within SMs, across SMs via a global work queue, and across multiple GPUs through asynchronous communication.

OpenAI Launches ChatGPT Pulse for Personalized Mobile Updates

OpenAI introduced ChatGPT Pulse, a mobile feature delivering personalized, proactive updates through visual cards by synthesizing user data, chat history, and app integrations.

OpenAI announced ChatGPT Pulse on September 25, 2025, initially rolling out to Pro users on mobile
Pulse provides proactive, personalized updates by synthesizing chat history, memory, feedback, and connected apps like Gmail and Google Calendar
Updates are delivered as topical visual cards, with user-curated research preferences, safety checks, and optional app integrations

Claude Code: An Agentic System with Persistent Memory and Unix Integration

Noah Brier highlights Claude Code as an agentic system using filesystem access and Unix command integration, enabling persistent memory, supporting complex workflows, and exemplifying Unix principles for reliable AI tools.

Noah Brier describes Claude Code as a full-fledged agentic operating system supporting workflows, notably integrating with Obsidian for note-taking.
Claude Code leverages Unix command integration and filesystem access, enabling persistent memory and state, addressing limitations of browser-based LLMs.
The system exemplifies the Unix philosophy—simple, composable tools that do one thing well—facilitating effective tool calling and building reliable AI agents.

AWS Launches Kiro to Boost Self-Serve AI Adoption Amid Competition

AWS seeks to boost organic AI app adoption by launching Kiro for self-serve developers, shifting from top-down sales, amid competitive pressure from Cursor, Windsurf, and Microsoft.

AWS aims to increase organic growth of its AI applications, reducing reliance on sales teams.
Q Developer lags behind rivals Cursor and Windsurf in grassroots adoption; AWS launched Kiro to attract self-serve developers.
Internal documents reveal AWS’s difficulty replicating early developer-driven growth in its AI portfolio, with a focus on shifting toward user-led adoption models.

OpenAI Unveils Internal AI Tools Threatening SaaS Giants

OpenAI showcased internal AI tools for sales, support, and contract analysis, posing a competitive threat to SaaS giants like HubSpot, Salesforce, DocuSign, and ZoomInfo, potentially transforming enterprise software.

OpenAI demonstrated AI-powered workplace tools, including Inbound Sales Assistant, GTM Assistant, DocuGPT, and Support Agents, potentially competing with existing SaaS vendors.
These internal applications could threaten companies like HubSpot, Salesforce, DocuSign, and ZoomInfo by overlapping core functionalities such as sales, customer relationship management, and contract analysis.
Stock prices of competitors declined: HubSpot (-10%), DocuSign (-12%), ZoomInfo (-6%), Salesforce (-3%), with Salesforce down 28% YTD, amid concerns over OpenAI’s expanding AI integration.

OpenAI Launches Sora 2 for AI Video and Audio Creation with Viral CEO Cameos

OpenAI released Sora 2, an AI video/audio app enabling face and voice scans for personalized AI content, with viral videos featuring CEO Sam Altman in humorous, surreal scenarios.

OpenAI launched Sora 2, an AI video and audio generation app, on October 2, 2025, currently invite-only with resale codes selling for under $45.
Users can scan their face and voice to create AI avatars and insert themselves into generated videos, with CEO Sam Altman frequently featured.
Notable AI videos include Altman “stealing GPUs at Target,” visiting Studio Ghibli HQ, rapping about OpenAI’s success, hanging out with dinosaurs, and participating in “South Park” scenarios.

California enacts SB 53 to lead in responsible AI regulation and transparency

California’s SB 53, signed by Governor Newsom, establishes regulations for frontier AI safety, transparency, and innovation, reinforcing the state’s leadership in responsible AI development and policy.

Governor Gavin Newsom signed SB 53 (Transparency in Frontier Artificial Intelligence Act) into law on September 29, 2025, to promote responsible AI development in California.
The law mandates public disclosure of AI development frameworks, establishes a consortium (CalCompute) for safe AI research, and creates safety and accountability mechanisms, including incident reporting and whistleblower protections.
SB 53 responds to recommendations from a California-commissioned report on AI guardrails, aims to position California as a global leader in ethical AI regulation, and fills federal policy gaps.

ServiceNow Launches AI Experience with Voice Agents and Content Insights

ServiceNow introduced its “AI Experience” in September 2025, including “AI Lens” for content-based workflow automation and voice agents for speech-driven processes, aiming for full deployment by year’s end.

ServiceNow launched “AI Experience” three weeks after Zurich platform upgrade on September 10, 2025
The upgrade includes “AI Lens” for interpreting screenshots and forms, enabling workflow initiation based on content
The “AI Experience” features “voice agents” for workflow control via speech, with a demo involving HR and IT ticketing
The company plans to deliver full “AI Experience” capabilities, including voice agents, by the end of 2025
Chief Innovation Officer Dave Wright emphasizes using AI for outcome-driven tasks rather than robotic process automation

Alvaro Lopez Ortega / 2025-10-02 Briefing

▶️ Internet Infrastructure

▶️ Open Source

▶️ Software Development

▶️ Management and Leadership

▶️ Technology