Gemma 4：逐字节对比，性能最强的开源模型Gemma 4: Byte for byte, the most capable open models

1. OpenAI has introduced flexible, pay-as-you-go pricing for Codex-only seats on ChatGPT Business and Enterprise plans, allowing teams to access Codex without fixed seat fees. This enables smaller groups to pilot the tool, test its value in specific workflows, and scale usage incrementally. Codex-only seats operate without rate limits and are billed based on token consumption, offering greater cost transparency and easier budget tracking across teams and projects. In contrast, standard ChatGPT Business seats—now reduced from $25 to $20 annually per seat—include Codex access but with usage limits. To support adoption, eligible workspaces can receive up to $500 in credits ($100 per new Codex-only user, capped at five users) for a limited time. The Codex app for macOS and Windows, enhanced with Plugins and Automations, facilitates integration with existing team systems. These updates come as Codex usage grows rapidly: over 2 million developers use Codex weekly, and Codex adoption within ChatGPT Business and Enterprise has increased sixfold since January. Companies like Notion, Ramp, Braintrust, and Wasmer are leveraging Codex to improve engineering efficiency, enabling faster execution and more consistent workflows.
   Key Takeaways:
   Flexible Codex pricing supports scalable team adoption
   Token-based billing improves cost visibility and control
   ChatGPT Business price reduced to $20 per seat annually
   Adoption incentives include up to $500 in credits
   Source: Original Article

1. OpenAI has acquired The Business of Politics Network (TBPN), a media platform known for facilitating discussions on AI, technology, business, and culture. The acquisition was announced by Fidji Simo, who emphasized TBPN’s editorial strengths, audience engagement, and ability to convene influential voices in the tech space. According to Simo, TBPN has become a key venue for ongoing, real-time conversations about AI and its impact, particularly among builders and users of the technology. OpenAI stated that traditional corporate communication strategies are insufficient for its mission, given the transformative nature of artificial general intelligence (AGI). The company views TBPN as already fulfilling the role of fostering constructive dialogue around AI’s societal implications. As part of the acquisition, TBPN will retain editorial independence, continuing to manage its programming, guest selection, and content decisions. OpenAI also highlighted TBPN’s marketing and communications expertise, noting their ability to help brands navigate digital spaces and anticipate industry trends. The integration aims to enhance OpenAI’s outreach and public understanding of AI’s broader impact. TBPN will operate within OpenAI’s structure, though specific reporting lines or departmental placement were not detailed.
   Key Takeaways:
   OpenAI acquires TBPN to strengthen AI discourse and public engagement
   TBPN retains editorial independence under OpenAI ownership
   Acquisition leverages TBPN’s marketing and audience engagement expertise
   Move reflects need for specialized communication in AI development
   Source: Original Article

1. Google has introduced new advanced features in Google Vids, including high-quality video generation, custom music creation, and AI avatars, aimed at enhancing user-generated video content. The platform, designed for intuitive video editing, now allows users to transform ideas into polished videos for personal or professional use, such as tutorials, travel recaps, or celebratory montages. These updates expand accessibility and creative control for a broad user base.

As of this week, all Google account holders can generate up to 10 high-quality video clips per month at no cost using Veo 3.1, Google’s latest video generation model. Users can create content from simple text prompts or photos, enabling quick production of promotional materials, event flyers, or greeting cards. Additionally, Google AI Pro and Ultra subscribers gain access to custom music generation via the Lyria 3 and Lyria 3 Pro models, allowing creation of original soundtracks ranging from 30 seconds to three minutes tailored to a video’s mood. Subscribers can also use AI avatars powered by Veo 3.1, offering directorial control over appearance and behavior to maintain consistency and engagement across video frames.

Key Takeaways:
Google Vids now offers free high-quality video generation for all users
Custom music and AI avatars available to Google AI Pro and Ultra subscribers
Veo 3.1 enables avatar direction and dynamic storytelling in videos
Personal accounts receive 10 free video generations monthly

Source: Original Article

1. Google has introduced two new service tiers—Flex and Priority—for the Gemini API, designed to help developers balance cost and reliability based on workload requirements. Flex Inference targets latency-tolerant, high-volume background tasks such as data enrichment, CRM updates, and agentic reasoning, offering a 50% cost reduction compared to the Standard API. It operates through synchronous endpoints, eliminating the need for asynchronous job management required by the Batch API. This simplifies integration while maintaining performance for non-time-sensitive operations. Priority Inference, meanwhile, is optimized for interactive, user-facing applications like chatbots and copilots that demand high reliability and low latency. Both tiers use the same unified API interface, allowing developers to route requests based on criticality by setting the service_tier parameter. This integration reduces architectural complexity by enabling both background and interactive workloads to run through standard synchronous calls. Flex is available across all paid tiers and supports GenerateContent and Interactions API requests. The launch reflects Google’s effort to support evolving AI use cases, particularly autonomous agents, by providing scalable, cost-efficient infrastructure. Detailed information on Priority Inference’s technical specifications and availability is limited in the provided content.
   Key Takeaways:
   Flex tier cuts costs by 50% for non-urgent AI tasks
   Priority tier ensures high reliability for user-facing applications
   Unified API simplifies architecture for diverse workloads
   Synchronous endpoints reduce need for async job management
   Source: Original Article

1. The AI news digest highlights a series of developments in World Models over the past quarter, featuring insights from leading researchers and companies. Yi Tay introduced the topic, followed by discussions with Fei-Fei Li and Justin Johnson on Marble at World Labs, and Pim de Witte of General Intuition on gaming-data-trained models. Andrew White discussed the Cosmos World Model on the Science pod, while the hosts explored Adversarial World Models. Major players like Nvidia, Waymo, Tesla, and Google (with Genie 3) have released their own approaches. Yann LeCun secured $1 billion for AMI and introduced LeWorldModel. In contrast, Moonlake AI presents a fundamentally different approach—emphasizing multiplayer interactivity, indefinite simulation duration, and diverse world modeling through environment simulation and long-horizon planning. Unlike Genie 3, which suffers from terrain clipping, lack of physics, and limited 60-second immersion, Moonlake’s system supports continuous, interactive experiences. Chris Manning and Ian Goodfellow argue in Towards Efficient World Models that efficiency through structured, causal reasoning—rather than mere scaling—is critical, noting that state-of-the-art models still exhibit spatial and physical inconsistencies. They suggest high-resolution pixel modeling is often unnecessary for economically valuable tasks, drawing parallels to human perception.

Key Takeaways:
Moonlake AI offers a scalable, interactive alternative to current World Models
Efficiency via causal structure outperforms blind scaling in world modeling
Genie 3’s limitations highlight flaws in non-interactive, short-duration simulations
Human-like perception prioritizes relevance over pixel-level detail in planning

Source: Original Article

1. Google has significantly upgraded its AI-powered video editing tool, Google Vids, by integrating its latest video and audio models—Veo 3.1 and Lyria. Veo 3.1, previously introduced in Gemini, enhances realism and consistency in generated video content. Unlike OpenAI, which appears to be scaling back on video generation, Google is expanding its AI video capabilities within consumer and business applications. Vids is positioned not for professional filmmaking but for casual and commercial uses such as animated party flyers, business sizzle reels, and video greeting cards. Users can select from controllable AI avatars to appear in generated videos, increasing customization. Sharing directly to YouTube has also been streamlined. Access to video generation is tiered: free users receive 10 generations per month, AI Pro subscribers get 50, and AI Ultra subscribers (personal or enterprise) receive up to 1,000. Generated videos are capped at eight seconds and 720p resolution, consistent with other Veo implementations. The Lyria music model now enables users to generate background music by describing a desired mood, producing 30-second or three-minute tracks without requiring lyrics. Music generation limits also vary by subscription tier.

Key Takeaways:
Google advances AI video tools while OpenAI scales back
Veo 3.1 improves realism in short-form video generation
Tiered subscriptions limit free access to AI features
Lyria enables mood-based music creation without lyrics

Source: Original Article

1. An AI State of the Union: We’ve Passed the Inflection Point, Dark Factories Are Coming, and Automation Timelines
   In a recent episode of Lenny Rachitsky’s podcast, the guest discussed the current state of artificial intelligence, emphasizing that the field has passed a critical inflection point in November. The conversation highlighted how software engineers are serving as early indicators of broader trends affecting information workers, as AI tools increasingly assist in coding tasks. The guest demonstrated coding on a mobile device, underscoring the growing accessibility of AI-powered development tools. The discussion also touched on “vibe coding,” a term describing intuitive, AI-assisted programming, and stressed the importance of responsible use to avoid errors. The concept of “dark factories”—fully automated manufacturing environments—was introduced, with StrongDM cited as an example of infrastructure enabling secure, automated access in enterprise settings. A key insight was that while AI accelerates code generation, testing has become the new bottleneck in software development. The guest acknowledged the mental fatigue associated with rapid AI advancements, noting the emotional and cognitive toll on professionals.
   Key Takeaways:
   AI has reached a pivotal moment in adoption and capability
   Software engineers are leading indicators of AI’s impact on knowledge work
   Testing is now the main constraint in AI-driven development
   Source: Original Article

1. Google has enhanced its video editor app Vids with several new AI-powered features, including text-prompt-driven avatar direction, Veo 3.1 video generation, direct YouTube export, and a Chrome extension for screen recording. Users can now direct avatars using natural language prompts to perform actions such as interacting with products or equipment, while maintaining character consistency across scenes. Avatars can also be customized through prompts by adjusting appearance, clothing, and backgrounds based on video themes. The Veo 3.1 model enables users to generate up to eight-second video clips within the app, with 10 free generations per month for all users and up to 1,000 monthly generations for Google AI Ultra and Workspace AI Ultra subscribers. Additionally, finished videos can now be exported directly to YouTube as private uploads, allowing creators to review content before publishing. A new Chrome extension enables screen recording with audio or video capture. These updates build on previous integrations, such as Lyria 3 music models and multilingual voice-over support added earlier in 2025. Google continues to expand Vids’ capabilities since its 2024 launch, targeting enterprise and consumer content creation.
   Key Takeaways:
   Google enhances Vids with AI avatars and video generation
   Veo 3.1 enables short video clips with free monthly limits
   Direct YouTube export streamlines content publishing workflow
   Screen recording Chrome extension expands capture options
   Source: Original Article

1. SpaceX has accused Amazon’s Project Kuiper (formerly Kuiper Systems) of violating orbital debris mitigation requirements by launching satellites into initial altitudes 50–90 km higher than authorized, increasing collision risks with other spacecraft, including SpaceX’s Starlink constellation. SpaceX claims this action disregards safety protocols and endangers operational systems and crewed missions. The accusation follows two recent Starlink satellite failures that generated new space debris. Amazon denies the allegations, stating its launch parameters comply with FCC approvals and that SpaceX previously supported a similar altitude for a Kuiper launch when it served as Amazon’s launch provider. Amazon argues that adjusting the altitude now would cause significant delays. Both companies have a history of using FCC regulatory processes to challenge each other’s satellite deployment plans. SpaceX alleges Amazon failed to coordinate properly and is creating “unmitigable collision risks.” Amazon maintains it disclosed the launch altitudes to the FCC and suggests SpaceX’s objections are retaliatory, especially given FCC Chairman Brendan Carr’s public criticism of Amazon’s efforts to block SpaceX’s expansion requests. Detailed information on specific collision probability data or FCC regulatory responses is limited.

Key Takeaways:
SpaceX accuses Amazon of unsafe satellite launch altitudes violating debris rules
Amazon denies violations, cites prior coordination and FCC disclosures
Both companies have history of regulatory disputes over LEO satellite deployments
Collision risk concerns highlight growing tensions in crowded low-Earth orbit

Source: Original Article

1. A newly discovered fossil site near Fuxian Lake, south of Kunming, China, has yielded over 700 species from the late Ediacaran period, approximately 7 million years before the Cambrian explosion. The fossils, collectively termed the Jiangchuan Biota, were found in the Dengying Formation and include both Ediacaran-era organisms and early representatives of groups that later flourished in the Cambrian. This coexistence challenges the long-held view that Ediacaran life forms disappeared abruptly in a mass extinction before the Cambrian diversification. Instead, the findings suggest a more gradual ecological transition between the two periods. The fossils are small—typically one to two centimeters—and preserved as carbon-rich impressions in a single sedimentary layer, with some showing detailed internal structures. Researchers from Yunnan University and Oxford University propose the organisms were rapidly buried in a shallow marine environment slightly below the low tide mark, aiding exceptional preservation. The site was previously known for abundant algae, but this discovery significantly expands understanding of late Ediacaran biodiversity. Detailed information limited on specific species and ecological interactions.

Key Takeaways:
Fossils show Ediacaran and Cambrian species coexisted before Cambrian explosion
Jiangchuan Biota reveals gradual transition between geological periods
Exceptional preservation in carbon-rich sediment enables detailed study
Site near Fuxian Lake expands knowledge of late Ediacaran life

Source: Original Article

1. Researchers led by molecular biologist Pablo S. Villar at Harvard University have conducted the first detailed scientific study of octopus mating behavior, focusing on the California two-spot octopus (Octopus bimaculoides). The study, published in Science, addresses long-standing gaps in understanding how these solitary, deep-ocean cephalopods reproduce. Male octopuses use a specialized arm called the hectocotylus to transfer sperm to females, but prior knowledge was largely based on anecdotal observations. To investigate further, the team placed a male and female octopus in a controlled tank environment with a physical barrier to prevent aggression, given the species’ solitary and potentially combative nature. The experiment aimed to observe initial interactions and mating mechanics without risking injury to the animals. While full mating was not observed due to the barrier, the study provided new insights into courtship behaviors and sensory cues used during early contact. The decentralized nervous system of octopuses—where two-thirds of neurons reside in the arms—may play a role in how individuals assess potential mates through tactile and chemical signals. This research marks a foundational step in understanding cephalopod reproduction, with implications for marine biology and neurobiology. Detailed information on the full mating sequence remains limited due to experimental constraints.
   Key Takeaways:
   First controlled study of octopus mating behavior reveals new courtship insights
   Male octopuses use hectocotylus for sperm transfer in solitary encounters
   Decentralized nervous system may influence mate assessment through arm-based sensing
   Experimental design prioritized safety due to octopus aggression risks
   Source: Original Article

1. Google has expanded its Gemma 4 family with new open models designed for on-device AI applications, emphasizing local execution over cloud dependency. These models are small, fast, and capable of handling multimodal tasks, enabling real-time processing directly on everyday devices such as smartphones and laptops. By leveraging local context, the models aim to deliver more relevant and actionable insights without relying on constant internet connectivity. This shift supports privacy-preserving AI and reduces latency, making AI more responsive and efficient in consumer applications. The development reflects a broader industry trend toward decentralized AI, where computation occurs closer to the user. While specific performance benchmarks and availability details are not provided, the focus on efficiency and omni-capability suggests Google is targeting broad deployment across personal devices. This move positions Google competitively in the open model space, challenging proprietary alternatives while promoting accessibility and innovation in edge AI.

Key Takeaways:
Google introduces efficient on-device AI models in Gemma 4 family
Models prioritize local processing for real-time, context-aware insights
Shift supports privacy and reduces reliance on cloud infrastructure
Open models aim to expand AI accessibility across consumer devices

Source: Original Article

1. A proposed class action lawsuit filed by an anonymous Perplexity user, identified as John Doe, alleges that Perplexity’s AI search engine shares users’ chat conversations—including initial prompts and follow-up questions—with Google and Meta without informed consent. The complaint claims this data sharing occurs regardless of whether users have accounts, and affects both subscribed and non-subscribed users. Non-subscribed users’ initial prompts are shared via URLs that allow third parties to access entire conversations. The lawsuit asserts that personally identifiable information (PII), including email addresses, is transmitted alongside chat data, even when users enable Perplexity’s “Incognito Mode,” which the suit calls a “sham” for failing to prevent data sharing. The complaint further alleges that sensitive information, such as financial and health-related queries, is exposed through embedded ad trackers from Google and Meta, likened to “browser-based wiretap technology.” The plaintiff accuses Perplexity, Google, and Meta of violating state and federal privacy laws by failing to disclose the use of third-party tracking technologies and prioritizing profit over user privacy. Detailed information on Perplexity’s technical infrastructure or specific legal statutes cited is limited.

Key Takeaways:
Perplexity allegedly shares user chat data with Google and Meta without consent
Incognito Mode fails to prevent data sharing, including PII
Sensitive financial and health queries exposed via third-party trackers
Lawsuit accuses companies of violating privacy laws for profit

Source: Original Article

1. TeleGuard, a messaging app marketed as secure and end-to-end encrypted with over one million downloads, has been found to implement encryption so poorly that attackers can easily access users’ private keys and decrypt messages, according to multiple security researchers. The app uploads users’ private keys to its own servers, enabling TeleGuard itself to decrypt messages, contradicting its claims of strong privacy. Additionally, researchers found that private keys can be partially derived by intercepting user traffic, further undermining security. Despite branding itself as “Swiss made” and emphasizing “highly encrypted” communications with “no storage of data,” the app fails to meet basic security standards. This revelation exposes significant risks for users who rely on TeleGuard for confidential communication, particularly in sensitive contexts. The findings highlight broader concerns about the lack of regulation and transparency in the encrypted messaging app market, where misleading marketing can create a false sense of security. Detailed information limited on the specific technical flaws or potential exploitation in the wild, but the implications for user privacy are substantial.

Key Takeaways:
TeleGuard’s encryption is fundamentally flawed and easily bypassed
Private keys are stored on company servers, enabling message decryption
Marketing claims of security are misleading and unsubstantiated
The incident reflects broader issues in encrypted app trustworthiness

Source: Original Article

1. Minnesota photojournalist Rob Levine, represented by the Reporters Committee for Freedom of the Press, has filed a lawsuit against the Federal Aviation Administration (FAA) challenging a January 2025 temporary flight restriction (TFR) that prohibits drone flights within 3,000 feet of Department of Homeland Security (DHS) facilities and mobile assets, including Immigration and Customs Enforcement (ICE) vehicles. The restriction was implemented during heightened ICE activity in Minneapolis under “Operation Metro Surge,” when hundreds of agents operated in both marked and unmarked vehicles, some with altered license plates. Levine argues the rule is unconstitutionally vague and infringes on First Amendment rights by making it impossible for journalists to determine whether their drone operations might violate the restriction. Because DHS vehicles can be unmarked or disguised, drone operators cannot reliably avoid prohibited airspace, creating legal uncertainty and chilling press freedom. The lawsuit contends the FAA failed to provide adequate justification or public notice for the broad restriction. Legal counsel Grayson Clary noted that such restrictions appear increasingly used to limit journalistic access during law enforcement operations. Detailed information on the FAA’s internal decision-making process is limited.

Key Takeaways:
Drone no-fly zone around DHS assets challenged as unconstitutional
Restriction creates legal uncertainty for journalists using drones
Lawsuit claims First Amendment rights are being suppressed
FAA’s vague enforcement undermines press freedom during operations

Source: Original Article

1. NASA’s Orion spacecraft successfully executed a 5-minute, 50-second main engine burn on Thursday, placing the Artemis II mission on a free-return trajectory around the Moon. This translunar injection maneuver occurred approximately 24 hours after launch aboard the Space Launch System (SLS) rocket from Kennedy Space Center. The burn, the final major firing of Orion’s main engine, sets the four astronauts—Reid Wiseman, Victor Glover, Christina Koch, and Jeremy Hansen—on a path to loop around the Moon on Monday, use lunar gravity to return to Earth, and splash down in the Pacific Ocean on April 10. This marks the first time since 1972 that humans have traveled beyond low-Earth orbit, a milestone not witnessed by about 75% of the global population. NASA officials expressed high confidence in the mission’s progress, with Lori Glaze, senior exploration official, stating satisfaction with current operations. Detailed information limited on long-term system performance, but initial checks confirmed effective life support and propulsion systems. A minor issue with the spacecraft’s toilet during initial priming was resolved after additional water was added to activate the pump.

Key Takeaways:
Artemis II completes critical engine burn for lunar flyby
First crewed mission beyond low-Earth orbit since 1972
Life support systems functioning well despite minor toilet issue
Crew on track for Monday lunar pass and April 10 splashdown

Source: Original Article

1. A developer has created MicroSafe-RL, a lightweight reinforcement learning (RL) framework designed to address the "Hardware Drift" problem, where RL agents trained in simulation fail or cause damage when deployed on real-world hardware due to unmodeled physical variations. The system is optimized for microcontrollers and emphasizes safety and adaptability in resource-constrained environments. It achieves a latency of 1.18 microseconds (85 cycles on an STM32 microcontroller running at 72MHz) and uses only 20 bytes of RAM without dynamic memory allocation (no malloc), making it suitable for embedded systems. The model-free approach allows it to adapt to mechanical wear and environmental changes by using exponential moving average (EMA) and median absolute deviation (MAD) statistics. Additionally, the framework includes a Python-based Auto-Tuner that generates optimized C++ parameters from just two minutes of real-world telemetry data, simplifying deployment. The project is open-source and available on GitHub. Detailed information on real-world testing or specific use cases is limited.

Key Takeaways:
MicroSafe-RL reduces simulation-to-reality gaps in reinforcement learning
Ultra-low latency and memory usage enable embedded deployment
Auto-Tuner simplifies parameter optimization from minimal telemetry
Model-free design adapts to hardware degradation over time

Source: Original Article

1. GeForce NOW adds ten new games in April, including Capcom’s PRAGMATA, with a total of a dozen titles available for streaming this week. Arknights: Endfield, an expansion of the popular mobile series into a full 3D action RPG, is among the new additions. The update reflects NVIDIA’s ongoing effort to expand its cloud gaming library with high-profile and niche titles alike. PRAGMATA, long delayed since its initial 2021 announcement, marks a significant release for both Capcom and GeForce NOW users, offering a sci-fi action experience previously unavailable on the platform. The inclusion of Arknights: Endfield signals NVIDIA’s strategy to attract fans of established franchises by bringing mobile-to-console transitions to cloud streaming. These additions enhance GeForce NOW’s value proposition amid growing competition in the cloud gaming sector. While specific performance metrics or user engagement data are not provided, the consistent weekly game additions suggest a focus on content breadth to retain and grow subscribers. The timing aligns with broader industry trends of leveraging exclusive or early-access titles to differentiate cloud platforms.
   Key Takeaways:
   GeForce NOW adds twelve new games this week, including PRAGMATA and Arknights: Endfield
   Capcom’s delayed PRAGMATA finally launches on cloud gaming via GeForce NOW
   NVIDIA continues expanding library to strengthen position in competitive cloud gaming market
   Source: Original Article