Today’s tech news highlights include the leak of 410GB of plaintext chats from US officials via TeleMessage, Google’s launch of Project Mariner AI for web browsing, and Waymo’s safer autonomous driving demonstration contrasting Tesla’s FSD failure. Additionally, OpenSource efforts like Microsoft’s open WSL code and France’s adoption of UN open source principles mark significant advancements.
▶️ Internet Infrastructure
DDoSecrets leaks 410GB of TeleMessage data exposing US officials’ plaintext chats
DDoSecrets released 410 GB of TeleMessage heap dumps containing plaintext chat logs and metadata, exposing security flaws and usage by US officials, after multiple hacks and server vulnerabilities.
- DDoSecrets published 410 GB of heap dumps hacked from TeleMessage’s archive server on May 19, 2025
- Data includes plaintext messages, metadata, sender/recipient info, timestamps, and group names from TeleMessage software used by government officials
- TeleMessage, used since at least February 2023, produces software that archives encrypted messaging apps like Signal and WhatsApp, but the dumps reveal plaintext chat logs and server vulnerabilities
Google Unveils Project Mariner Web-Browsing AI for Ultra Subscribers
Google introduced Project Mariner at I/O 2025, an AI agent capable of multi-task web browsing, now available to AI Ultra subscribers, with cloud-based operation enabling concurrent task handling.
- Google announced during Google I/O 2025 the rollout of Project Mariner, an AI agent that browses and interacts with websites, to more users and developers
- Support for nearly a dozen simultaneous tasks, running on cloud virtual machines, significantly updates the initial browser-based design
- U.S. subscribers to the $249.99/month AI Ultra plan gain access, with support for additional countries and integration into Gemini API and Vertex AI planned
Tesla FSD Fails Red Light Test While Waymo Demonstrates Safer Autonomous Driving
Business Insider tested Tesla’s FSD and Waymo’s robotaxi in San Francisco; Tesla made a critical red-light violation, while Waymo demonstrated safer, assertive driving with minor lane changes.
- Waymo’s robotaxis have provided fully autonomous rides in San Francisco since 2024, equipped with fifth-generation Waymo Driver, five lidar sensors, six radars, and 29 cameras.
- Tesla’s Model 3 with Hardware 4 and FSD Supervised software v13.2.8, using eight external cameras, was tested; Tesla plans to release FSD Unsupervised later in 2025.
- During a test ride from Twin Peaks to Chase Center, Tesla’s FSD ran a red light at Twin Peaks Blvd and Portola Drive, while Waymo maintained safety without such errors.
Oracle Cloud Identity Outage in Europe Lasts 6 Hours
OCI’s identity platform in Europe was down for about 6 hours on May 19, 2025, with problems peaking at 111 reports; Oracle did not publicly comment on the outage.
- Oracle Cloud Infrastructure (OCI) experienced a European outage on May 19, 2025, lasting approximately 6 hours
- The outage affected the OCI identity platform in parts of Europe, including the Germany Central (Frankfurt) region
- Failover attempts to the Netherlands Northwest (Amsterdam) region were unsuccessful; Oracle did not publicly acknowledge the issue
Dell Launches Rapid-Deployment Private Cloud Platform for VMware, Nutanix & Red Hat
Dell’s private cloud platform allows deployment of VMware, Nutanix, and Red Hat stacks with 90% fewer steps, enabling rapid setup in 2.5 hours using the Dell Automation Platform.
- Dell introduced a private cloud platform enabling deployment of VMware, Nutanix, and Red Hat stacks on Dell servers and storage
- Uses “validated blueprints” for rapid deployment, reducing setup steps by 90%, with environments operational in 2.5 hours via “Dell Automation Platform”
- Supports co-existence of multiple private clouds across server and storage fleets with centralized management; only VMware vSphere is available initially, with other templates coming later in 2025
AI GPU Utilization Falls Short of Expectations, Calls for Better Metrics
Uptime Institute highlights that AI GPU utilization is often only 35-45%, with no existing perfect metric; MFU shows promise as it correlates with power consumption, but real-world data is needed.
- GPU accelerators in AI are costly, with Nvidia H100 priced between $27,000 and $40,000; cloud rental on Azure costs $6.98/hour per GPU
- GPU utilization is often overestimated by AI teams; actual usage during training is around 35-45%, with GPUs operational about 80% of the time
- Current metrics like server operational time or hardware load are insufficient; Model FLOPS Utilization (MFU) measures ratio of observed to theoretical performance, with typical values between 35-45%, indicating low efficiency
HPE Enhances Storage Guarantee Program with Cyber Resilience and Energy Efficiency SLAs
HPE’s Alletra Storage MP B10000 now includes SLAs for cyber resilience, energy efficiency, and zero data loss, featuring ransomware detection, immutable snapshots, and active site fail-over guarantees.
- HPE expands storage guarantee program with new SLAs for cyber resilience, energy efficiency, and zero data loss on the Alletra Storage MP B10000
- Introduces a cyber resilience guarantee with real-time ransomware detection, immutable snapshots, and a 30-minute support response guarantee
- Adds energy consumption guarantee reducing power usage by 45%, decreasing storage footprint by 30%, and lowering TCO by 40%; also offers a zero RTO/RPO guarantee for continuous data availability
Virgin Media O2 Fixes Location Leak in 4G Calling Protecting User Privacy
Virgin Media O2 patched a VoLTE metadata leak enabling callers to locate users within 100 meters by analyzing IMS, IMEI, and cell ID headers, which has now been addressed to protect user privacy.
- Virgin Media O2 fixed a vulnerability in its 4G Calling feature that allowed callers to locate recipients within approximately 100 meters using IMS, IMEI, and cell ID data.
- Researcher Daniel Williams demonstrated that IMS signaling messages contained detailed headers, including IMSI, IMEI, and cell ID, which could be used to identify user location and device information.
- The issue, first reported on May 17, was resolved after VMO2 confirmed the fix was implemented and tested, with Williams noting the headers should be removed from all IMS/SIP messages to ensure privacy.
▶️ Open Source
Solving the Emoji Problem: Elliptic Curves and Rational Point Generation
The article explores solving a complex elliptic curve problem related to the “Emoji Problem,” using line methods, rational point generation, and computational algebra to find large integer solutions.
- The article discusses the “Emoji Problem,” a challenging mathematical puzzle involving rational points on elliptic curves, inspired by viral emoji math problems.
- It explains the use of line and tangent line techniques to generate new rational points on elliptic curves, including formulas for point addition and duplication.
- The post details a step-by-step process of finding rational solutions, employing computational tools like Mathematica to handle large algebraic expressions, ultimately producing enormous positive integer solutions.
Microsoft Opens WSL Source Code for Community Collaboration in 2025
Microsoft released WSL as open source on May 19, 2025, enabling community development; WSL 2.5.7 includes core components like command line tools, WSL service, Linux processes, and plan9 file sharing, with some proprietary components remaining.
- Microsoft announced the open-source release of the Windows Subsystem for Linux (WSL) on May 19, 2025, available on GitHub.
- WSL codebase is now open source, allowing community contributions, bug fixes, and feature development; current version is 2.5.7.
- WSL architecture includes command line executables (wsl.exe, wslconfig.exe, wslg.exe), WSL service (wslservice.exe), Linux init and daemon processes, and file sharing via plan9 server; some components like Lxcore.sys and filesystem redirection remain proprietary.
Creating a Safe YAML Parser That Mimics Complex Behaviors Across Languages
The article details creating a “safe” YAML parser that reproduces complex parser differentials across six languages using only safe features like merge keys, !!binary, anchors, aliases, and BOM, avoiding unsafe methods.
- The article explores constructing a “safe” YAML parser that reproduces the “YAML monster” behavior using only safe parsing methods across multiple programming languages.
- Techniques include leveraging YAML features such as merge keys (
<<:),!!binarytags, anchors, aliases, and UTF-8 BOM to create parser differentials without unsafe methods. - Examples demonstrate how different languages (Python, Go, Ruby, Java, Node.js, Rust) interpret complex YAML constructs, highlighting inconsistencies and parser quirks, with detailed YAML snippets and language-specific behaviors.
Bilibili’s AniSora: Open-Source AI for High-Quality Anime Video Creation
AniSora, developed by Bilibili, is an open-source AI model for high-quality anime-style video generation from images or prompts, supporting diverse styles and applications like series, manga, and VTuber content.
- AniSora is an open-source animated video generation model developed by Bilibili, supporting one-click creation across anime styles including series episodes, Chinese animations, manga, VTuber content, and PVs.
- It leverages IJCAI'25-accepted research to enable high-quality, expressive animation from images or text prompts, with a focus on motion consistency and visual fidelity.
- Users can upload images, select from multiple AI models (e.g., Kling 2.0, Vidu Base, AniSora), and generate videos via the AniSora AI Video Generator.
La France adopte officiellement les principes open source de l’ONU avec 19 organisations
La France est la première nation à adopter officiellement les principes open source de l’ONU, avec 19 organisations supplémentaires, renforçant l’engagement international pour la transparence et la collaboration logicielle.
- France devient le premier gouvernement à endosser les principes open source de l’ONU, rejoint par 19 organisations, selon unite.un.org
- Les 8 principes de l’ONU #OpenSource incluent : ouverture par défaut, contribution, sécurité, inclusion, réutilisabilité, documentation, reconnaissance, durabilité
- La déclaration officielle date du 18 mai 2025, soulignant un soutien croissant pour la collaboration open source dans le système des Nations Unies
Deep Learning and Topology: Unveiling Data Manifolds and Neural Network Transformations
Deep learning models operate on high-dimensional manifolds, with topology revealing how neural networks manipulate data and embeddings to classify, reason, and generate models.
- The article explains how deep learning models relate to topology, viewing data and neural networks as high-dimensional manifolds or subspaces.
- It discusses the role of topological transformations in classifying complex data, emphasizing the importance of high-dimensional spaces for separability.
- Neural networks are described as topology generators, embedding data into mathematical spaces where semantic properties are represented as points and surfaces.
Microsoft Revives DOS-era Edit as Lightweight Open-Source Rust TUI for Windows 11
Microsoft revived the DOS-era text editor Edit as an open-source Rust TUI app, integrating modern features and planning to include it in Windows 11, enhancing command line editing.
- Microsoft open-sourced the DOS-era text editor Edit, weighing less than 250 KB, announced at Build, available on GitHub.
- The modern version of Edit is a Rust-based TUI application with features like mouse support, Unicode, menu keybindings, word wrap, Find & Replace, Match Case, and Regex.
- The new Edit is intended to ship as part of Windows 11, with a preview for Windows Insiders, providing a lightweight, stable, and familiar text editing tool for command line use.
High-Severity OpenPGP.js Flaw Lets Attackers Spoof Signed Messages
A high-severity vulnerability in OpenPGP.js (CVE-2025-47934) enables message spoofing by bypassing signature verification, affecting versions 5.0.1–5.11.2 and 6.0.0-alpha.0–6.1.0, with a proof-of-concept imminent.
- Vulnerability CVE-2025-47934 (severity 8.7 – high) affects OpenPGP.js versions 5.0.1–5.11.2 and 6.0.0-alpha.0–6.1.0
- Allows spoofing of signed and encrypted messages by passing maliciously modified messages to
openpgp.verifyandopenpgp.decrypt - Exploit can make a message appear legitimately signed without actual signature verification; full proof of concept coming soon
▶️ Software Development
Claude Code SDK Launches CLI Tools for Seamless Application Integration
The Claude Code SDK provides command line tools for integrating Claude Code into applications, supporting JSON and streaming outputs, multi-turn conversations, custom prompts, MCP configuration, and session management.
- The Claude Code SDK enables programmatic integration of Claude Code into applications, supporting command line usage with upcoming TypeScript and Python SDKs.
- Basic usage includes running prompts in non-interactive mode with options for JSON, stream-json output, and session continuation; advanced features include multi-turn conversations, custom system prompts, and MCP configuration.
- CLI options include
--print,--output-format,--resume,--continue,--system-prompt,--allowedTools,--mcp-config, among others; output formats supported are text, json, and stream-json with strict message schemas.
Overcoming Go Interface Limitations for Zero-Copy Image Decoding
The article discusses overcoming Go’s interface indirection by implementing custom Peek and Bytes methods, enabling zero-copy data access in image decoding workflows.
- The article explores inefficiencies in Go’s handling of
io.Readerinterfaces, especially when extracting byte slices without copying. - It highlights that
bytes.Readerdoes not implementPeek, complicating zero-copy access, and thatbufio.Readerwraps readers without exposing underlying data. - The author implements custom
PeekandBytesmethods via shadow APIs and unsafe pointer conversions to enable zero-copy image decoding.
Jules: AI-Powered Automation for Seamless Coding and GitHub Integration
Jules automates coding workflows with GitHub integration, cloud VM testing, and Gemini 2.5 Pro AI, streamlining tasks like version updates and PR creation.
- Jules is an asynchronous coding agent designed to automate coding tasks such as bug fixing, version bumping, and feature building
- Integrates with GitHub to import repositories, clone code in cloud VMs, run tests, and create pull requests
- Utilizes the latest Gemini 2.5 Pro model (DeepMind) for planning and code development
Making Games in 2025: Noel Berry’s Open-Source, Engine-Free Approach
Noel Berry advocates making games in 2025 without large engines, using C# with SDL3, FMOD, Dear ImGui, and open-source libraries, emphasizing control, portability, and longevity.
- Noel Berry has been making games for 20 years, emphasizing development without commercial engines like Unity or Unreal.
- Uses C# with dotnet’s hot reload, reflection, and modern features for cross-platform, high-performance game development.
- Relies on libraries such as SDL3 for windowing, input, rendering, and FMOD for audio; writes custom layers and tools, often utilizing Dear ImGui for editors.
- Supports porting to major platforms via SDL3’s console ports and SDL3’s GPU abstraction, with C# code compiled using Native-AOT for console compatibility.
- Has transitioned to Linux for development, citing frustrations with Windows and advocating open-source, cross-platform tools.
- Prefers small, custom tools over large engines, citing control, simplicity, and longevity; uses open-source frameworks like Love2D, FNA, and SDL.
- Implements asset management by loading files at startup or on demand, with simple preprocessing scripts; creates custom level editors with Dear ImGui.
- For 3D or advanced tech, considers rolling own frameworks or using Unreal, depending on project needs; highlights making stylized, less tech-dependent games.
- Advocates for smaller teams or solo developers to build tailored workflows, reducing reliance on proprietary engines and promoting open-source solutions.
- Uses tools like Aseprite for assets, integrating them directly into custom engines for animation and game data management.
Zod 4 Launches with Major Speed Boost and Compact Size Improvements
Zod 4, released after a year of active development, enhances performance, reduces bundle size, and adds features like metadata, JSON Schema conversion, recursive objects, and improved error handling.
- Zod 4 is now stable after a year of development, offering improved speed, reduced size, and new features
- Introduces significant performance benchmarks: 14x faster string parsing, 7x faster array parsing, 6.5x faster object parsing
- Reduces core bundle size by approximately 57% (from 12.47kb to 5.36kb gzipped) and introduces
zod/v4-miniwith an 85% size reduction (to 1.88kb gzipped)
▶️ Management and Leadership
GitHub Copilot Coding Agent Launches in Public Preview for Issue Delegation
GitHub announced Copilot coding agent in public preview, enabling issue delegation and automated code changes via secure cloud environment, available for Pro+ and Enterprise plans, with usage tracking and gradual platform expansion.
- GitHub Copilot coding agent enters public preview on May 19, 2025, supporting issue delegation via github.com, GitHub Mobile, and CLI.
- Operates in a secure cloud environment powered by GitHub Actions, exploring repositories, making changes, validating with tests and linters before pushing.
- Available for Copilot Pro+ and Copilot Enterprise subscribers; usage consumes GitHub Actions minutes and premium requests, with rollout to mobile (iOS, Android) and CLI starting May 19.
- From June 4, 2025, each model request by the agent will use one premium request; UI is subject to change during preview.
- Supports low-to-medium complexity tasks such as feature addition, bug fixes, test extension, refactoring, and documentation improvements, with multiple issues assignable simultaneously.
InventWood to Launch Superwood: A Lightweight, Strong, Fire-Resistant Material
InventWood will mass-produce Superwood, a high-strength, lightweight, fire-resistant material derived from treated cellulose, with 50% more tensile strength than steel, starting summer 2025.
- InventWood plans to commercialize Superwood, a material with 50% higher tensile strength than steel and a strength-to-weight ratio ten times better, starting summer 2025
- The process involves treating and compressing cellulose in timber, densifying it by up to 4x, creating stronger hydrogen bonds, and resulting in durable, fire-resistant, rot- and pest-resistant wood
- The technology originated from a 2018 Maryland lab breakthrough by Liangbing Hu, who refined the process from over a week to a few hours, licensing it to InventWood; initial products target facade materials for buildings
Use Browser Language Settings Instead of IP Geolocation for Accurate User Language Detection
Rely on the browser’s Accept-Language header for accurate user language detection, avoiding IP geolocation, which is unreliable and causes user dissatisfaction.
- Using IP geolocation for language detection is unreliable due to VPNs, travel, and multilingual regions, leading to user frustration.
- Countries like Belgium, Switzerland, India, and Canada have multiple official languages, making country-based language assumptions inaccurate.
- Browsers send an
Accept-Languageheader indicating user language preferences, which should be used instead of IP-based methods. - The
Accept-Languageheader example:Accept-Language: en-US,en;q=0.9,de;q=0.8. - Recommended approach: read and respect
Accept-Language, allow user adjustments, and use GeoIP only for currency, shipping, or legal purposes.
Apple to Launch SDK for Developers to Build on Its AI Models at WWDC
Apple will open its AI models via an SDK at the June 9 WWDC, enabling developers to create new apps and features based on Apple Intelligence’s large language models.
- Apple plans to release a software development kit (SDK) and frameworks enabling third-party developers to build applications using its large language models (LLMs) for Apple Intelligence.
- The initiative aims to enhance app creation and device functionality, with the plan expected to be announced at the Worldwide Developers Conference on June 9.
- The move involves opening access to Apple’s AI models to foster new applications and improve user engagement on Apple devices.
Big Tech Firms Flatten Structures to Boost Efficiency and Decision-Making
Big Tech firms like Microsoft, Intel, Amazon, Google, and Meta are reducing middle management layers to boost efficiency and agility, with some managers overseeing up to 60 direct reports, despite potential risks to leadership development and employee morale.
- Microsoft plans to cut approximately 6,000 employees to increase managers’ span of control
- Intel, Amazon, Google, and Meta are also flattening organizational structures for efficiency
- Flattening management can speed decision-making but risks overloading remaining managers with too many direct reports
Lawyers’ Use of ChatGPT and AI in Legal Brief Sparks Judicial Criticism
Burke’s lawyers used ChatGPT and Westlaw AI to draft a legal brief with multiple inaccuracies, prompting judicial criticism; Burke faces hacking charges related to leaked Fox News footage.
- Timothy Burke’s lawyers admitted to using ChatGPT and Westlaw AI features without verification in a legal brief containing nine false quotes and six errors.
- Judge Kathryn Kimball Mizelle criticized the brief for “significant misrepresentations” and “miscellaneous problems,” demanding an explanation.
- Rasch’s research involved Westlaw’s Quick Check, Google, Google Scholar, and ChatGPT 4.5’s deep research feature; errors stemmed from unvetted drafts and limited review.
Madhu Gottumukkala Appointed CISA Deputy Amid Leadership Gaps and Cyber Threats
Madhu Gottumukkala became CISA’s deputy director amid ongoing budget cuts, workforce attrition, and threats from foreign governments and cybercriminals, with no confirmed agency head.
- Madhu Gottumukkala appointed as CISA deputy director, announced on May 19, 2025
- Previously oversaw South Dakota’s cybersecurity efforts as CIO and CTO
- CISA remains without a Senate-confirmed director; Sean Plankey’s nomination blocked by Senator Ron Wyden
UK NHS Urges Vendors to Sign Voluntary Cybersecurity Pledge to Combat Ransomware
UK NHS cybersecurity leaders call on vendors to sign a voluntary security pledge to combat endemic ransomware threats, with system patching, MFA, monitoring, and incident reporting as key requirements.
- UK government and NHS cybersecurity officials urge NHS suppliers to sign a public security pledge via a voluntary charter.
- The NHS has experienced multiple ransomware incidents over the past year, including attacks on Synnovis, Wirral University Hospitals, and Alder Hey Children’s Hospital.
- The charter requires vendors to fully patch systems, comply with NHS Data Security and Protection Toolkit (DSPT), implement MFA, deploy 24/7 cyber monitoring, maintain immutable backups, conduct incident response exercises, and report incidents promptly.
- The NHS plans to launch a self-assessment form in autumn for suppliers to sign the charter; signing is voluntary and not legally binding but complements existing legal obligations like GDPR.
- The initiative aligns with upcoming UK Cyber Security and Resilience Bill, which will extend protections for NHS supply chains and enforce cybersecurity standards in NHS contracts.
- The NHS emphasizes that future contractual and regulatory processes will incorporate the charter’s expectations, aiming to improve cybersecurity culture and resilience across healthcare services.
Linus Torvalds Returns to Cherry Blue Keyboard and Calls for Better Keyboard Standards
Linus Torvalds reverted from quiet keyboards to Cherry Blue switches, emphasizing the need for better keyboard engineering and standardized tactile language, amid ongoing hardware experimentation.
- Linus Torvalds experimented with quiet, low-profile keyboards but returned to a Cherry Blue mechanical switch keyboard, citing its superior feel.
- He previously used an Ampere Altra Max system for Arm64 Linux builds in 2024, while still using an Apple M2 laptop for weekly test builds.
- Torvalds criticizes keyboard design, highlighting the lack of a standardized language to describe tactile and auditory feedback, and advocates for better-engineered, personalized keyboards.
Trump’s AI Moratorium Faces Opposition Over Legal Chaos and Consumer Risks
Trump’s proposed bill introduces a 10-year federal moratorium on state AI regulation, risking legal chaos and consumer harm by preventing enforcement of substantive AI laws, amid bipartisan opposition.
- President Trump’s FY 2025 budget reconciliation bill, the “One Big Beautiful Bill Act,” includes Section 43201(c), proposing a 10-year moratorium on state regulation of AI.
- The moratorium restricts enforcement of any state law or regulation regulating AI models, systems, or automated decision systems, with few exceptions allowing only laws that remove legal impediments to AI adoption.
- 40 state attorneys general and over 140 organizations oppose the provision, warning it would prevent states from regulating AI harms, undermine consumer protections, and cause legal chaos.
AI in Science Overhyped: Unreliable Results and Weak Baselines
Author’s experiments and reviews reveal AI in science often overhyped, with unreliable results, weak baselines, and biases, suggesting AI’s role is more incremental than revolutionary.
- Author, a Princeton plasma physics PhD, tested AI methods (PINNs) on PDEs and found them unreliable compared to advanced numerical baselines.
- Systematic review shows 79% of studies claiming AI outperforms traditional methods used weak baselines; positive results often overstate AI potential.
- AI adoption in science increased from 2% in 2015 to nearly 8% in 2022, driven more by researcher incentives than proven scientific benefits.
▶️ Technology
Google DeepMind Launches New AI Models for Video, Image, and Music Creation
Google DeepMind unveiled Veo 3, Imagen 4, Flow, and Lyria 2 on May 20, 2025, advancing AI media creation with new capabilities for video, image, and music generation, emphasizing responsible use.
- Google DeepMind announced Veo 3, Imagen 4, Flow, and Lyria 2 models on May 20, 2025
- Veo 3 generates videos with audio, excelling in physics, lip-sync, and prompt understanding; available in Gemini app and Vertex AI
- Imagen 4 produces high-resolution (up to 2k), photorealistic images with superior typography; faster variant up to 10x in development; accessible via Gemini, Whisk, and Workspace
- Lyria 2 offers advanced music composition tools, including Lyria RealTime for interactive, real-time music generation; expanded access via YouTube Shorts and Vertex AI
- Flow enables cinematic filmmaking through natural language descriptions, integrating Veo, Imagen, and Gemini models; available to Google AI Pro and Ultra subscribers in the U.S.
- All models incorporate SynthID watermarks; SynthID Detector launched for content verification to reduce misinformation risks
Google Unveils Project Astra for Real-Time AI Video Streaming and New Features
Google announced during Google I/O 2025 that Project Astra, a low latency, multimodal AI platform, will enable new Search features, Gemini app capabilities, and third-party developer tools, including real-time video streaming with minimal latency.
- Google announced during Google I/O 2025 that Project Astra, a low latency, multimodal AI experience, will power new features in Search, Gemini app, and third-party products
- Project Astra powers a new Search Live feature enabling real-time video and audio streaming for visual queries via AI Mode, Lens, and the “Live” button
- The technology streams live video/audio into AI models with minimal latency; it originated from Google DeepMind and was showcased via a smart glasses demo at Google I/O 2024; partnerships include Samsung and Warby Parker; launch date pending
- Project Astra enhances the Gemini Live API with low-latency voice, visual input, native audio output, improved emotion detection, and reasoning capabilities from Gemini AI models
- Visual input and real-time video sharing are expanding to all Gemini app users; previously limited to paid subscribers
- Google envisions Project Astra powering future smart glasses, though no launch date has been set; details on glasses design remain limited
Google Gemini 2.5 Pro Debuts with One-Million-Token Context and ‘Deep Think’ AI
Google’s Gemini 2.5 and 2.5 Pro models, with a one-million-token context window and ‘Deep Think’ reasoning, achieve top benchmark scores, supporting advanced multimodal and conversational AI capabilities.
- Google announced Gemini 2.5 and 2.5 Pro at Google I/O 2025, enhancing reasoning, coding, and long-context benchmarks
- Gemini 2.5 Pro features a one-million-token context window and an experimental ‘Deep Think’ reasoning mode, scoring highly on USAMO, LiveCodeBench, and MMMU benchmarks
- Updates include native audio output, text-to-speech with multiple speakers, thought summaries, thinking budgets, and SDK support for Model Context Protocol (MCP)
Google’s Flow AI might disrupt Hollywood with personalized, cost-effective film creation
Google’s Flow AI tool, demonstrated at IO 2025, creates high-quality cinematic videos from text prompts, potentially disrupting Netflix and traditional studios through cost-effective, personalized content.
- Google unveiled Flow, an AI moviemaking tool, at the IO conference, built on Imagen 4 and Veo 3 models.
- Flow generates visuals, sound effects, and dialog from text prompts, enabling scene editing and precise camera control.
- AI-generated content could challenge traditional studios like Netflix by reducing production costs, accelerating timelines, and enabling hyper-personalized films.
Google Launches AI Mode for Conversational Multi-Source Search in the US
Google’s AI Mode, now available in the US, offers conversational, multi-source search with advanced query techniques, integrating future AI features into standard search and developing a universal AI assistant.
- Google is rolling out AI Mode in the US, providing a conversational search experience with multiple source surfacing and follow-up questions.
- AI Mode employs “query fan-out,” running multiple queries simultaneously to enhance complex question answering.
- New features in Labs include Deep Search for fully cited reports, real-time data visualizations, and personalized search via integration with user search history.
Google Expands AI Mode with Gemini 2.5 and New Search Features in US
Google is expanding AI Mode to all US Search users, featuring Gemini 2.5, with new shopping, web task, and visualization tools, previewed at I/O 2025 for broader integration.
- Google has begun rolling out AI Mode to all Search users in the US, announced at I/O 2025.
- AI Mode is a chatbot integrated into Google Search, powered by a custom Gemini 2.5 AI model, replacing Gemini 2.0.
- New features include tools for shopping, ticket price comparison, custom chart and graphic generation, and future integration with Deep Search and web tasks via Project Mariner.
Google Meet Launches Real-Time AI Translation for Natural Multilingual Conversations
Google’s live translation in Google Meet uses AI to provide real-time Spanish translation, enhancing multilingual communication and expected to expand to additional languages soon.
- Google introduces real-time AI-powered live translation in Google Meet, starting with Spanish, available now for Google AI Pro and Ultra plans in beta.
- The feature matches speaker tone, cadence, and expressions, enabling more natural conversations and reducing reliance on live captions.
- More languages are planned to be added, with enterprise availability later this year during early testing phases.
Google Launches Gemini 2.5 Pro and Unveils New AI Innovations
Google unveiled Gemini 2.5 Pro with an Elo score of 1448, enhanced infrastructure, and new AI projects like Beam and personalized services, emphasizing practical AI applications and performance benchmarks.
- Google CEO Sundar Pichai announced the launch of Gemini 2.5 Pro, achieving an Elo score of 1448, surpassing previous versions and leading LLMArena leaderboard
- Google’s infrastructure includes the 7th generation TPU, Ironwood, delivering 42.5 exaFLOPS of compute, available to Cloud customers later in 2025
- New AI features include Gemini API inheriting capabilities from Project Mariner, Project Starline updates with Google Beam for 3D video, and AI-enhanced Google Meet with real-time translation
Google Partners with Warby Parker, Gentle Monster, and Samsung for Stylish XR Smart Glasses
Google’s I/O 2025 revealed collaborations with Warby Parker, Gentle Monster, and Samsung to produce stylish XR smart glasses with AI integration, aiming for mainstream appeal and fashion-forward design.
- Google announced partnerships with Warby Parker, Gentle Monster, and Samsung at I/O 2025 to develop stylish Android XR smart glasses.
- Warby Parker offers accessible, trendy eyewear, while Gentle Monster is a popular, edgy Korean brand favored by celebrities; both aim to target versatile and trendsetting markets.
- The first Android XR glasses will be manufactured by Xreal under Project Aura, featuring cameras, microphones, speakers, and integration with Google’s Gemini AI for functions like photo capture, navigation, and live translation.
Google I/O ‘25 Showcases AI Innovations and Product Launches
Google I/O ‘25 keynote highlighted new AI innovations, product launches, and research, emphasizing advancements in AI tools, accessibility, and user experience, streamed with over 5.78 million views.
- Google I/O ‘25 keynote streamed live for 3 hours, 2 minutes with over 5.78 million views
- Focused on latest AI updates, product announcements, and technological advancements
- Included sessions on AI tools like Flow, Project Astra, and visual interpreters for low-vision navigation