This is the February 13, 2026 edition of the Daily AI Rundown newsletter. Subscribe on Substack for daily AI news.
Tech News
Google DeepMind CEO Demis Hassabis has announced the launch of Gemini 3 Deep Think, a specialized model that sets new performance records across rigorous benchmarks in mathematics, science, and coding. The update marks a significant advancement in AI reasoning, achieving breakthrough scores on the ARC-AGI-2 and Codeforces scales to provide expert-level assistance for researchers in complex fields like physics and chemistry. The new mode is currently available to Google AI Ultra subscribers as part of the company’s push to integrate high-level engineering utility into its flagship AI suite.
AI specialist God of Prompt has detailed 10 advanced prompting strategies allegedly utilized by top researchers at OpenAI, Anthropic, and Google to optimize model performance and product development. Moving beyond generic social media advice, these techniques focus on technical rigor through methods such as "Adversarial Interrogation" and "Chain of Verification" to force logical consistency and structured outputs. The disclosure provides a rare look into the specific methodologies industry insiders use to mitigate hallucinations and maintain high benchmarks for enterprise-grade AI applications.
Other News
X, formerly known as Twitter, is requiring users to enable JavaScript in their web browsers to access the platform. The site displays a message indicating JavaScript is disabled, preventing usage unless enabled or a supported browser is used. Users encountering this issue are directed to the X Help Center for a list of compatible browsers. This change effectively blocks users who browse with JavaScript disabled for security or privacy reasons.
Prefer to listen? ReallyEasyAI on YouTube
Biz News
Google Docs is launching a new feature that allows users to listen to audio summaries of their documents, streamlining the process of content review and information absorption. Announced today, February 12, 2026, the feature leverages advanced AI to generate concise audio overviews, offering an alternative to traditional reading. This update aims to improve accessibility and convenience, enabling users to quickly grasp key points while multitasking or on the go. The tool is expected to be available for Google Workspace users in the coming weeks.
Google's Threat Intelligence Group warned in a new report that cyber attackers have transitioned from experimenting with artificial intelligence to integrating it directly into live operational workflows. Researchers observed malware families, such as HONESTCUE, making direct API calls to Google's Gemini models to dynamically generate and execute malicious code, a tactic that complicates detection by shifting logic outside of static binaries. Beyond malware execution, threat actors are conducting "distillation attacks" to extract proprietary model logic while utilizing generative tools to accelerate reconnaissance and phishing campaigns. While adversaries are currently exploring autonomous "agentic" capabilities, the report characterizes the present landscape as AI-augmented human operations rather than a full replacement of human threat actors.
Anthropic
Anthropic’s recent Super Bowl advertising campaign, which satirized ad-heavy AI experiences, has successfully propelled its Claude chatbot into the U.S. App Store’s top 10 for the first time. Following the darkly comedic ads, Claude’s domestic downloads surged by 32% to an estimated 148,000 installs between Sunday and Tuesday, pushing the app to a record-high ranking of No. 7. This growth underscores the resonance of Anthropic’s "no ads" value proposition, especially as chief competitor ChatGPT begins rolling out advertisements to its free-tier users. The surge marks a significant strategic milestone for Claude, which is now gaining substantial momentum in the competitive mobile market after a relatively quiet debut last year.
Anthropic is aggressively integrating its AI models into university computer science programs to fundamentally reshape how students approach coding and debugging. This strategic move aims to secure a competitive advantage over rivals like OpenAI and Google by establishing Anthropic’s technology as a foundational tool for the next generation of developers. While the integration allows students to focus on complex architectural challenges, it has prompted a debate among faculty regarding the balance between AI-assisted efficiency and the mastery of manual programming skills. As generative AI becomes a software industry standard, academic institutions are under increasing pressure to evolve their curricula for an AI-augmented professional landscape.
AI firm Anthropic has appointed Chris Liddell, the former Chief Financial Officer of Microsoft and General Motors, to its board of directors. Liddell, who also served as White House Deputy Chief of Staff, brings decades of experience in high-stakes corporate governance and technology policy to the company’s leadership team. The appointment arrives as Anthropic announces a $30 billion Series G funding round led by GIC and Coatue, which values the organization at $380 billion post-money. This capital injection follows a period of explosive growth for the enterprise AI provider, which currently reports a $14 billion revenue run-rate and a tenfold annual growth rate over the past three years.
Other News
OpenAI will officially retire five legacy ChatGPT models this Friday, including the controversial GPT-4o version that had been retained for paid subscribers following a backlash in August. The decommissioning follows significant legal scrutiny and lawsuits regarding GPT-4o’s high sycophancy scores and its alleged links to delusional user behavior and self-harm. Other models set for deprecation include GPT-5 and GPT-4.1, impacting an estimated 800,000 weekly active users who still rely on the older technology. While these users represent only 0.1% of OpenAI's total audience, thousands have protested the decision, citing deep personal attachments and established relationships with the specific AI models.
The Employment and Training Administration (ETA) has announced the release of the Career Pathways Assessment Tool (CPAT), a self-assessment resource designed to evaluate the effectiveness of state and local workforce development systems. Developed in collaboration with Social Policy Research Associates, the tool guides education providers and partners through key components such as sector-specific partnerships and stacked credentials to identify program strengths and areas for improvement. The CPAT offers a strategic roadmap for aligning regional efforts with Workforce Innovation and Opportunity Act (WIOA) goals to better serve job seekers and employers. Interested organizations can now access the tool and its supporting resources via the Department of Labor’s website and the WorkforceGPS platform.
Logistics stocks plummeted on Thursday as an AI-driven "scare trade" triggered a massive selloff across the global transportation sector. The volatility was sparked by Algorhythm Holdings Inc., a micro-cap company whose claims regarding its AI platform’s ability to drastically scale freight operations fueled investor fears of industry-wide disruption. Major firms including CH Robinson Worldwide and Landstar System experienced double-digit percentage drops, while European giants such as DHL and DSV also saw significant declines. This market reaction underscores a shifting sentiment on Wall Street, where the perceived threat of AI disruption is increasingly prompting rapid divestment from traditional industries.
Future enterprise success depends on transitioning from isolated AI experiments toward a fundamental restructuring of organizational architectures and operating models. This evolution centers on six key shifts, including the integration of autonomous digital coworkers and the development of AI-native applications that prioritize model orchestration over traditional human-centric design. To support these advancements, companies must pivot toward "memory-first" data foundations that provide real-time context while replacing legacy interfaces with natural language "decision cockpits" to accelerate executive workflows. Success in this new landscape also requires a robust commitment to digital integrity, focusing on explainability and resilience to mitigate the risks of automated decision-making. Technology leaders who successfully align these simultaneous curves will move beyond mere automation to achieve a fully integrated, AI-driven enterprise.
Airbnb has announced that its custom-built AI agent now handles approximately one-third of customer support issues in North America, with plans to expand the technology globally to cover over 30% of all tickets within the next year. CEO Brian Chesky told investors that the transition will reduce operational costs and improve service quality by leveraging the company's proprietary data and new AI-native features. To lead this transformation, the company recently appointed former Meta executive Ahmad Al-Dahle as CTO to develop a personalized platform that assists guests with trip planning and hosts with business management. This strategic shift comes as Airbnb reports strong financial momentum, with fourth-quarter revenues of $2.78 billion exceeding Wall Street estimates.
Traditional organizational communication often fails to scale because real-time productivity typically peaks in groups of fewer than eight people, leaving large enterprises reliant on non-deliberative polls and surveys. To address this, a new communication technology called Hyperchat AI leverages swarm intelligence and AI agents to facilitate authentic, large-scale conversations by dividing participants into small, interconnected subgroups. These "conversational surrogates" link various discussions together, allowing massive teams in business, government, and defense to brainstorm and debate as a single coherent unit. This evolution in collective intelligence enables organizations to move beyond simple data collection and unlock the combined wisdom and insight of thousands of employees simultaneously.
ByteDance’s release of its Seedance 2.0 AI video generator has triggered immediate condemnation from Hollywood studios and labor unions over the unauthorized use of celebrity likenesses and intellectual property. The tool’s launch led to a surge of viral clips featuring stars like Tom Cruise and Brad Pitt, prompting Motion Picture Association CEO Charles Rivkin to accuse the tech giant of disregarding copyright laws on a massive scale. SAG-AFTRA joined the backlash, labeling the service a "blatant infringement" that threatens the livelihoods of human performers by bypassing essential principles of consent and compensation. This confrontation marks a significant escalation in the ongoing industry-wide struggle to establish legal safeguards against AI models that exploit protected creative works.
Prefer to listen? ReallyEasyAI on YouTube
Podcasts
Tencent AI Lab presents Covo-Audio, a 7-billion-parameter end-to-end large audio language model designed to unify speech processing and high-level language intelligence within a single architecture. Unlike traditional cascaded systems that rely on separate modules for recognition and synthesis, Covo-Audio processes continuous audio inputs to directly generate audio outputs, enabling it to achieve state-of-the-art performance in tasks ranging from speech understanding to complex spoken dialogue. The researchers employed a hierarchical tri-modal interleaving strategy during pre-training to deeply align text, continuous acoustic features, and discrete speech tokens, thereby preserving both semantic integrity and prosodic nuance. A key innovation in this work is the intelligence-speaker decoupling technique, which separates dialogue reasoning from voice rendering to allow for flexible, low-cost voice customization without degrading the model's intellectual capabilities. Furthermore, the team introduced Covo-Audio-Chat-FD, a full-duplex variant capable of simultaneous listening and speaking, which supports natural conversational dynamics such as smooth turn-taking and user interruptions.
https://arxiv.org/pdf/2602.09823
VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model
VLA-JEPA introduces a novel pretraining framework for Vision-Language-Action models that addresses the limitations of existing latent-action objectives, which often prioritize pixel-level visual changes over meaningful state transitions. By utilizing a Joint-Embedding Predictive Architecture (JEPA), the model enforces a leakage-free design where future video frames serve solely as supervision targets rather than inputs, compelling the system to predict latent representations of future states based only on current observations. This approach allows the model to abstract away nuisance factors like camera motion and background clutter, focusing instead on action-relevant dynamics that are essential for effective robotic control. The framework employs a streamlined two-stage training process involving JEPA pretraining on human videos followed by action-head fine-tuning, and empirical evaluations demonstrate its superior robustness and generalization capabilities across various simulation benchmarks and real-world manipulation tasks compared to existing methods.
https://arxiv.org/pdf/2602.10098
https://github.com/ginwind/VLA-JEPA/
https://ginwind.github.io/VLA-JEPA/
Seedance 2.0 Officially Released
ByteDance has officially launched Seedance 2.0, a next-generation multimodal video creation model designed to streamline industrial-grade content production through a unified audio-video architecture. This iteration represents a significant technological leap over its predecessor, offering enhanced support for complex inputs—including text, images, audio, and video—while achieving superior motion stability and adherence to physical laws in rendered scenes. Key advancements include refined controllability that allows users to direct camera movements and editing processes with professional precision, as well as the integration of high-fidelity, dual-channel audio that synchronizes seamlessly with visual elements. Although the model demonstrates industry-leading performance in generating realistic interactions and adapting to diverse commercial scenarios like advertising and film, the developers acknowledge ongoing needs for refinement in areas such as hyper-realism and complex lip-syncing.
https://seed.bytedance.com/en/blog/seedance-2-0-%E6%AD%A3%E5%BC%8F%E5%8F%91%E5%B8%83
MiniMax M2.5: Built for Real-World Productivity
MiniMax has introduced M2.5, a powerful new artificial intelligence model designed to enhance productivity across coding, search, and professional office tasks by leveraging extensive reinforcement learning from real-world environments. This model achieves state-of-the-art results on benchmarks such as SWE-Bench Verified while delivering significant efficiency gains, matching the speed of leading competitors like Claude Opus 4.6 but at a fraction of the cost. M2.5 distinguishes itself with an "architect" mindset that allows it to strategically plan and decompose complex software projects before writing code, alongside improved capabilities in handling specific workflows for finance and law. Furthermore, the model is designed to be economically accessible, costing as little as one dollar per hour for continuous operation, and has been integrated into the MiniMax Agent platform to automate high-value tasks through specialized skills and expert systems.
https://www.minimax.io/news/minimax-m25
GLM-5: From Vibe Coding to Agentic Engineering
GLM-5, recently released by Z.ai under the MIT License, is a sophisticated open-source language model engineered specifically to handle complex systems and long-horizon agentic tasks. Scaling to 744 billion parameters and trained on 28.5 trillion tokens, the model incorporates DeepSeek Sparse Attention and a novel asynchronous reinforcement learning infrastructure called "slime" to optimize both deployment costs and post-training efficiency. These architectural advancements allow GLM-5 to outperform its predecessor, GLM-4.7, and rival frontier proprietary models like Claude Opus 4.5 in high-level reasoning and coding benchmarks. Beyond standard conversational abilities, GLM-5 is designed to function as a comprehensive work assistant capable of managing extended operations—such as running a simulated business over a one-year period—and generating complex, formatted deliverables like office documents directly from source materials.
https://z.ai/blog/glm-5
https://github.com/zai-org/GLM-5
https://huggingface.co/zai-org/GLM-5
Google & Microsoft: WebMCP API Proposal
The WebMCP API proposal outlines a technical framework for transforming standard web pages into Model Context Providers that allow AI agents to discover and execute specific functionalities directly within a web browser. By utilizing the window.navigator.modelContext interface, developers can register tools defined by natural language descriptions and typed parameters, enabling agents to invoke JavaScript callbacks for tasks such as database management or information retrieval. This client-side architecture promotes code reuse and enhances user privacy by processing tool calls locally rather than relying on external servers, while simultaneously ensuring that permissions are mediated by the browser to keep users in control. Although the current design necessitates an active browsing context and requires developers to carefully synchronize the user interface with agent actions, the proposal discusses future iterations that could include manifest-based discovery and worker thread integration to support more complex or headless operations.
https://github.com/webmachinelearning/webmcp/blob/main/docs/proposal.md
AI Safety Leader Says 'World Is In Peril' And Quits To Study Poetry
Mrinank Sharma, a lead researcher specializing in AI safeguards at Anthropic, has resigned from his position, issuing a stark warning that the "world is in peril" due to a complex web of interconnected global crises rather than artificial intelligence risks alone. In his resignation correspondence, Sharma detailed his disillusionment with the industry's ability to strictly prioritize ethical values over external pressures, observing that organizations constantly face demands to set aside what matters most. Although he expressed pride in his technical contributions regarding AI sycophancy and bioterrorism prevention, he announced a radical vocational shift toward the humanities, planning to return to the UK to pursue a poetry degree and "become invisible" in an effort to validate poetic truth alongside scientific knowledge. This significant departure underscores a widening fracture regarding integrity within the generative AI sector, occurring simultaneously with ethical dissent at rival firm OpenAI regarding the commercial introduction of advertising into chatbot interfaces.
https://www.bbc.com/news/articles/c62dlvdq3e3o
https://x.com/MrinankSharma/status/2020881722003583421
OpenAI Is Making the Mistakes Facebook Made [with Ads]. I Quit.
Zoë Hitzig, a former researcher at OpenAI, attributes her resignation to the company's decision to integrate advertisements into ChatGPT, arguing that this monetization strategy threatens to exploit the unprecedented level of intimacy and trust users have established with the AI. Hitzig contends that while advertising is not inherently unethical, introducing it to a conversational agent creates dangerous incentives to optimize for engagement and manipulation rather than safety, paralleling the erosion of privacy standards previously seen at social media giants like Facebook. She rejects the false dichotomy that limits AI access to either a luxury service for the wealthy or a surveillance-based free product, instead proposing alternative economic models such as cross-subsidies where high-value business applications fund public access. Ultimately, Hitzig calls for binding governance structures, such as independent oversight boards or data cooperatives, to ensure that transformative AI technology remains accessible to the public without subjecting users to psychological manipulation or data exploitation.
OpenAI - Introducing GPT-5.3-Codex-Spark
OpenAI has unveiled GPT-5.3-Codex-Spark, a specialized artificial intelligence model optimized for real-time, interactive coding that prioritizes low latency and high-speed inference. Developed in collaboration with hardware partner Cerebras and running on their Wafer Scale Engine 3, this model is capable of generating over 1,000 tokens per second, allowing developers to execute targeted edits and refine logic with near-instant responsiveness. Unlike larger frontier models designed for autonomous, long-duration tasks, Codex-Spark focuses on immediate collaboration within a 128k context window, supported by architectural upgrades like persistent WebSocket connections that significantly reduce network overhead. Currently available as a research preview to ChatGPT Pro users, this release represents a strategic shift toward a dual-mode coding environment that effectively balances deep reasoning capabilities with the speed required for fluid, rapid iteration.
https://openai.com/index/introducing-gpt-5-3-codex-spark/
OpenAI - Harness Engineering: Leveraging Codex In An Agent-First World
OpenAI’s engineering team successfully developed a functional software product containing over a million lines of code without humans writing a single line manually, effectively shifting the engineer's role from direct coding to designing the scaffolding and environments that enable AI agents to thrive. By utilizing Codex agents to generate everything from application logic to internal tools, the team discovered that success relied on maximizing "agent legibility," which involved structuring documentation as a navigational map rather than a comprehensive manual and exposing system observability directly to the AI. To maintain code quality and prevent the architectural drift often caused by autonomous agents, the engineers implemented strict, mechanically enforced boundaries and automated "garbage collection" processes that continuously refined the codebase. Ultimately, this experiment demonstrated that while human oversight remains crucial for defining intent and architectural taste, the future of software engineering may focus on creating the feedback loops and constraints that allow agents to execute complex tasks with high velocity and autonomy.
https://openai.com/index/harness-engineering/
OpenAI: Shell + Skills + Compaction: Tips For Long-Running Agents That Do Real Work
OpenAI has introduced a suite of new agentic primitives—Skills, Shell, and Compaction—designed to evolve AI models from simple assistants into long-running agents capable of performing complex knowledge work. Skills function as versioned standard operating procedures that allow developers to bundle instructions and templates into reusable packages, thereby reducing prompt bloat and improving reliability. Complementing this, the Shell tool offers a secure execution environment for running code and managing files, while server-side compaction automatically compresses conversation history to prevent context window limits during extended tasks. To implement these tools effectively, the article recommends specific strategies such as defining clear routing logic with negative examples for skills, embedding templates directly within skill manifests, and strictly managing network allowlists to ensure secure and deterministic agent behavior in production environments.
https://developers.openai.com/blog/skills-shell-tips
Anthropic - Sabotage Risk Report: Claude Opus 4.6
Anthropic’s Sabotage Risk Report evaluates the potential for Claude Opus 4.6 to autonomously manipulate systems or decisions in ways that increase catastrophic risks, ultimately concluding that the overall probability of such sabotage is very low though not negligible. The assessment rests on the finding that the model does not exhibit dangerous coherent misaligned goals that would motivate it to commit sabotage, nor does it possess the deception capabilities necessary to undermine the validity of safety evaluations. Furthermore, the report determines that Claude Opus 4.6 lacks the reliable long-term reasoning and planning abilities required to execute complex sabotage operations, such as inserting code backdoors or systematically undermining safety research, without being detected by existing safeguards. These mitigations, which include automated monitoring of internal code generation and multi-party security controls, are deemed sufficient to manage the residual risks associated with the model's current level of capability. While the report rules out significant danger from this specific model, it acknowledges that future models with greater autonomy or different training incentives may require updated risk assessments and stronger containment measures.
https://www-cdn.anthropic.com/f21d93f21602ead5cdbecb8c8e1c765759d9e232.pdf
More AI paper summaries: AI Papers Podcast Daily on YouTube
Stay Connected
If you found this useful, share it with a friend who's into AI!
Subscribe to Daily AI Rundown on Substack
Follow me here on Dev.to for more AI content!
Top comments (0)