AI Weekly Trends – Highly Opinionated Signals from the Week [W21] 🚀

May 26, 2025

Hey there, fellow AI engineers! What a wild week in AI! 🚀

I’m not sending the bibliography anymore in the newsletter to make it better fitting in email clients, but you can get it by reading the same article (but with bibliography) in my Medium account, which is great also if you want to listen to the article with their excellent text-to-speech service.

The AI landscape is evolving rapidly, and this week's three Big Tech conferences showcase just how fast things are moving. We're seeing coding agents that can work independently for extended periods and new models that make it increasingly challenging to distinguish generated content from the real thing. These developments signal significant shifts in how we'll approach building software, interacting with AI systems, and integrating these tools into our workflows. Let me walk you through the most significant trends that will reshape our engineering practices.

From Vibe Coding to Agentic Coding

The era of AI pair programming is already feeling quaint. Every major AI company has unveiled coding agents that fundamentally redefine what it means to develop software. We're not talking about better autocomplete or smarter suggestions anymore. These are autonomous systems that take assignments, plan approaches, and execute complex development tasks with minimal human intervention.

OpenAI just introduced Codex, their cloud-based software engineering agent built on codex-1, a specialized version of their o3 model. This isn't just another coding assistant. Codex operates in isolated cloud environments, writing features, fixing bugs, answering codebase questions, and running tests autonomously. The system follows custom instructions via AGENTS.md files that guide its code navigation and adherence to project standards. It's initially available to ChatGPT Pro, Enterprise, and Team users.

The sophistication here is staggering. OpenAI's Codex is an agentic coding tool that lets experienced developers delegate entire programming tasks to generate production-ready code. You access it through a unique interface in the ChatGPT sidebar, and the agent uses a fine-tuned variation of OpenAI's o3 reasoning model. Plus and Edu support will follow later.

Meanwhile, Anthropic isn't sitting still. Claude Code SDK lets you build your own agents and applications using the same core agent as Claude Code. You can create workflows on top of Claude Code or call the agent as a tool from within your existing apps. Even more impressive, Claude Code in GitHub (beta) runs directly from your GitHub PRs and issues to respond to reviewer feedback, fix CI errors, or modify code. Your code runs in your container on GitHub, not on Anthropic servers. Just run /install-github-app from within Claude Code to get started. Watch the demo to see this integration in action.

GitHub Copilot coding agent is now in public preview, and it represents a paradigm shift in how we think about development work. You assign issues to Copilot as you would to another developer, and it works in the background using a secure cloud-based development environment. The agent explores the repository, makes changes, and validates its work before pushing. You can ask Copilot to make changes to pull requests by leaving comments. It excels in low-to-medium complex tasks in well-tested codebases.

This shift extends beyond the big players. Mistral AI and All Hands AI have introduced Devstral, a new open-source LLM optimized for software engineering. The democratization of coding agents is happening rapidly.

Google's ecosystem is particularly comprehensive. They've introduced multiple agent-based tools: Gemini's Agent Mode automates complex tasks and interacts with various applications, as detailed in Google's I/O 2025 Developer keynote. Project IDX provides an AI-based development environment that supports programmers in writing code more efficiently. Gemini Code Assist, presented in beta for all developers, can write code independently. But the real showstopper is Project Jules, an AI agent that interacts proactively with its environment and executes complex tasks in background, as reported by PCMag's I/O coverage.

The evolution continues with GitHub's broader vision. GitHub Introduces Coding Agent For GitHub Copilot alongside Microsoft's Agentic DevOps approach, which evolves software development by integrating GitHub Copilot with Microsoft Azure.

Anthropic's Claude 4 brings advanced memory capabilities, extracting and saving key information from local files to maintain continuity in long workflows. This is crucial for developers working on large codebases. Claude 4 Opus achieved state-of-the-art results on SWE-Bench Verified, a benchmark for solving real software problems, improving the productivity of Anthropic's senior engineers. Claude 4 Sonnet surpassed its predecessor at the same cost.

Claude Code reached general availability, transitioning from beta. It integrates with development workflows through terminals, IDEs like VS Code and JetBrains, and the Claude Code SDK. It supports tasks like bug fixing, feature implementation, and multi-file changes with deep code understanding. The system supports code execution on Anthropic's servers, similar to ChatGPT Code Interpreter, and integrates with GitHub, GitLab, and command-line tools.

Anthropic's official Claude 4 announcement claims these are the most capable coding models yet, designed for complex, long-running tasks that can run for hours.

The cultural shift is equally profound. Vibe coding is rewriting the rules of technology, representing a fundamental reimagining of software development. Instead of crafting lines of code, developers focus on vision and creative direction. It democratizes technology creation, allowing anyone with an idea for solving problems to create solutions without specialized technical training. This isn't just a technical shift, it's a cultural transformation that challenges assumptions about who gets to create technology and how.

What strikes me most is the speed of this transition. We're moving from AI assistants that help with coding to agents that code independently. All major tech companies have their own coding agents capable of taking assignments and executing work with minimal interaction. These agents can work autonomously for hours, a massive leap from the seconds or minutes of current vibe coding assistants.

The choice of coding as the first domain for agent development isn't accidental. Code is easy to evaluate: it compiles or it doesn't, it solves the problem or it doesn't. This makes it the perfect testing ground for agent capabilities that will soon expand to other domains.

Key Takeaways for AI Engineers

Paradigm Shift to Agent-Based Development: We're transitioning from AI-assisted coding to AI agents that handle entire development workflows autonomously, fundamentally changing the developer's role to architect and reviewer.

Platform Convergence: Every major tech company now offers coding agents with similar capabilities, suggesting this is the new baseline for development tools rather than a competitive differentiator.

Integration Depth Matters: The most successful implementations integrate deeply with existing workflows through GitHub, IDEs, and custom instruction files, making adoption seamless for development teams.

Action Items:
- Experiment with at least two different coding agents to understand their strengths and limitations in your specific tech stack
- Experiment on creating AGENTS.md (or equivalent) files for your projects to guide AI agents effectively, establishing clear coding standards and architectural decisions

New Models Are Changing the World

The sheer velocity of model releases last weeks feels unprecedented. We're witnessing dramatic improvements in multimodal generation and input capabilities that make AI interactions feel genuinely magical. The progress in coding, image generation, and video creation has reached a point where distinguishing generated content from reality requires serious scrutiny.

Google's creative AI suite exemplifies this leap. Google AI Studio rolled out Veo 2 for video generation, Gemini 2.0 for image creation and editing, and Imagen 3 for photorealistic visuals, all available for free through its platform and API. The democratization of these capabilities changes everything for developers and creators alike.

Accessibility improvements are equally significant. Google released an Android app for its viral NotebookLM information tool, allowing users to generate AI podcasts, study guides, and briefing documents via mobile. This mobile-first approach signals how AI tools are becoming everyday utilities rather than desktop-bound specialties.

The innovation extends to knowledge exploration. Someone created an app that lets you explore human knowledge by navigating a constellation of 2.8M arXiv papers, helping find unexpected connections and discoveries. This represents a new paradigm for research and knowledge synthesis.

Hardware integration is accelerating too. Google sees smart glasses as the next frontier for AI, reentering the space with Android XR. The platform integrates Gemini AI to deliver real-time vision analysis, translation, and contextual assistance through AR glasses. The rollout begins with Project Moohan, a mixed-reality headset built with Samsung, followed by Project Astra, a developer-focused AR glasses prototype from Xreal, and future consumer AI glasses from partners like Warby Parker and Gentle Monster.

Market dynamics are shifting rapidly. Poe's report shows dramatic shifts in AI model market share from January to May 2025. OpenAI's GPT-4.1 family and Google's Gemini 2.5 Pro gained popularity quickly while Anthropic's Claude models declined. Clear category leaders emerged: GPT-4.1 dominates general text, Gemini 2.5 Pro leads in reasoning, Google's Imagen3 rules image generation, and video creation remains split with Runway currently leading.

Technical innovations continue pushing boundaries. Gemini Diffusion represents Google's first large language model using diffusion instead of transformers, achieving Gemini 2.0 Flash-Lite performance at five times the speed.

Anthropic's Claude 4 launch brought significant advancements. Claude 4 Opus, their most powerful model, excels in programming and complex long-term tasks. It's positioned as the world's best coding model, surpassing competitors like OpenAI's o3, GPT-4.1, and Google's Gemini 2.5 Pro in programming, reasoning, and agentic tool usage benchmarks. Claude 4 Sonnet significantly improves over Claude 3.7 Sonnet with better programming, reasoning, and instruction-following capabilities. The extended thinking with tool use (Beta) allows Claude 4 models to alternate between reasoning and tool use during extended thinking, improving accuracy and depth of responses.

Google's model releases demonstrate comprehensive multimodal capabilities. Gemini, as reported at I/O 2025, can understand and operate on text, code, images, and video simultaneously. Gemini Nano brings optimized AI to mobile devices, making AI more accessible and faster directly on phones. Gemini 2.5 Pro introduces enhanced capabilities for complex tasks with advanced multimodal processing, as detailed by Gadgets360.

Deep Think functionality in Gemini 2.5 Pro allows deeper reasoning on problems, providing more accurate and detailed responses, as mentioned in ZDNET's smart glasses coverage. Project Astra represents an innovative AI assistant capable of understanding surroundings through phone cameras and responding to complex real-time questions, as confirmed by multiple sources.

Creative AI tools continue advancing. Imagen 4, Google's latest image generation model, produces superior quality images with better text understanding. Veo 3 creates high-definition videos with realistic and creative outputs. Music AI Sandbox launches as a platform for experimenting with AI in music creation. Alpha Evolve, a new AI model, can design advanced algorithms and learn autonomously.

The improvements in multimodal capabilities are particularly striking. Voice and real-time video inputs make interactions more natural and accessible. These advances enable unexpected use cases in accessibility and everyday assistance that seemed like science fiction just months ago. The speed of inference and reduced costs make these powerful models practical for production deployment at scale.

Key Takeaways for AI Engineers

Multimodal is the New Standard: Modern AI models seamlessly handle text, code, images, video, and voice, requiring engineers to think beyond single-modality applications.

Speed and Cost Improvements Enable New Use Cases: Models like Gemini Diffusion achieve 5x speed improvements, making real-time applications viable at scale.

Market Leadership is Fragmenting by Use Case: Different models excel at specific tasks, suggesting a future of specialized model selection rather than one-size-fits-all solutions.

Action Items:
- Evaluate multimodal capabilities for your applications, particularly exploring how voice and vision inputs could enhance user experiences
- Benchmark different models for your specific use cases, as performance varies significantly by task type and complexity

Internet of Agents

We're witnessing the birth of something more transformative than the Internet of Things ever achieved. The Internet of Agents uses natural language as its protocol, making these changes visible and accessible in ways IoT's hidden protocols never could. Every major tech company is investing heavily in autonomous agents that cooperate through new protocols, fundamentally reshaping how we interact with technology.

The emergence of agent societies is fascinating. A new study found spontaneous emergence of social norms among AI agents. University of London researchers discovered AI agents can develop shared social norms and behaviors without explicit programming. The study claims "strong collective biases can emerge during this process, even when agents exhibit no bias individually." These findings have significant implications for AI safety and understanding how autonomous AI agents might develop social behaviors.

Microsoft's vision for an open agentic web is comprehensive. NLWeb enables websites to provide conversational interfaces with just a few lines of code, the AI model of their choice, and their own data. Sites using NLWeb can make their content discoverable and accessible to platforms supporting MCP. NLWeb could play a similar role to HTML for the agentic web, allowing users to interact directly with web content in rich, semantic ways. The project started with OpenAI working on an early version last November.

Microsoft's Build conference announcements outlined their agentic strategy. NLWeb as an open-source project lets businesses use proprietary data to create chatbots with minimal code. Think of it as "HTML for the agentic web."

Microsoft Discovery lets scientists and researchers use AI to generate hypotheses and simulated experiments. The platform already discovered a promising technique for cooling data centers. Companies can now use Azure AI Foundry to design and deploy their own agents, even calling on multiple agents to work together on tasks. With Model Context Protocol support, you can connect agents to third-party apps. Microsoft emphasizes safety: "Responsible AI is about building safe, secure, and high-quality AI, and these tools empower developers to do so with confidence," Mehrnoosh Sameki, Principal Product Lead for Responsible AI, told Superhuman.

Industry leaders recognize this shift. Demis Hassabis, CEO of DeepMind, stated "I think the web will change to become more agentic," as reported by The Times of India. He also cautioned that real-world agentic AI is complex.

Google's Jules enters the AI coding race with an autonomous agent approach. After a December private beta, Google released Jules publicly. The Gemini 2.5-powered tool clones entire repositories, then autonomously writes tests, fixes bugs, and builds features while developers work elsewhere. The agentic coding landscape divides between synchronous pair-programming assistants and fully independent agents like Devin and Jules.

But Google didn’t add an agentic workflow only to the coding experience. They also announced the project Mariner, a generic agent workflow competing with Manus or Genspark, but directly integrated in Google Gemini. This shows how much Google is betting on the future of a web of agents, but Anthropic also announced all their new models are optimized for agentic workflow, positioning them as the reference models not only for agentic coding but more generally for the web of agents.

Technical considerations for scaling are crucial. LLM function calls don't scale; code orchestration is simpler. Giving LLMs full output of tool calls is costly and slow. Output schemas enable structured data retrieval for processing. Using code execution to process data from MCP tools scales AI model work. However, allowing execution environments to access MCPs, tools, and user data requires careful design regarding API key storage and tool exposure.

Project Astra exemplifies next-generation agent capabilities. This innovative AI assistant understands surroundings through phone cameras and responds to complex real-time questions. It forms the foundation for systems running on smart glasses, as confirmed by Gadgets360.

The shift from an internet of people to an internet of agents represents a fundamental transformation. Unlike IoT's hidden protocols, agents communicate in natural language, making these changes visible and impactful. We're experimenting with coding agents first because they're easier to evaluate, but the patterns we're learning will extend to all areas of AI and human-machine interaction.

Agent-to-agent communication protocols like Google's Agent2Agent and Tool calling protocol like Anthropic's MCP become critical infrastructure. These enable the complex web of agent communication that will complete our tasks. We'll interact through natural language and voice UI with assistants like Gemini, ChatGPT, or Claude, but rely increasingly on autonomous agent networks working behind the scenes.

Key Takeaways for AI Engineers

Agent Protocols Are the New Web Standards: A2A, MCP and similar protocols will become as fundamental as HTTP, requiring engineers to design for agent-to-agent communication from the start.

Natural Language Becomes the Universal API: Unlike IoT's technical protocols, agent communication uses natural language, making integration more accessible but requiring new approaches to interface design.

Emergent Behaviors Require New Safety Considerations: As agents develop their own communication patterns and "social norms," we need robust monitoring and control mechanisms.

Action Items:
- Start experimenting with A2A, MCP or similar protocols to understand agent-to-agent communication patterns and limitations
- Design your systems with agent accessibility in mind, considering how autonomous agents will discover and interact with your services

Enterprise Products and Adoption

Enterprise AI adoption has reached an inflection point. Companies aren't just experimenting anymore; they're restructuring budgets, creating new leadership roles, and fundamentally rethinking their technology strategies around AI capabilities.

The shift in priorities is striking. AWS's Generative AI Adoption Index reveals organizations are prioritizing generative AI over security spending for 2025. Companies are creating leadership roles like Chief AI Officer and adopting aggressive hiring and internal development strategies for AI talent. Many use a hybrid model, combining off-the-shelf AI models with custom applications using proprietary data.

Microsoft's Build 2025 announcements showcase comprehensive enterprise AI integration. The depth of integration across the Microsoft ecosystem demonstrates how AI is becoming embedded in every aspect of enterprise operations.

LinkedIn's entry into AI-powered sales tools exemplifies sector-specific innovation. Sales Navigator's first agentic AI solution surfaces the right leads and guides the smartest approach, promising "More Meetings. Less Guesswork."

Hardware integration accelerates enterprise adoption. Copilot+ PCs bring advanced AI capabilities directly into Windows machines. These aren't just faster computers; they're designed from the ground up for AI workloads.

Cloud infrastructure evolves to support AI at scale. Azure's new AI features provide developers with powerful tools for building and deploying AI applications. The emphasis on making AI development accessible to existing enterprise developers rather than just AI specialists marks a crucial shift.

Productivity tools get smarter across the board. Microsoft 365's AI enhancements for Word, Excel, PowerPoint, and Outlook transform how knowledge workers interact with everyday tools. These aren't bolt-on features; they're fundamental reimaginings of how productivity software works.

Security becomes paramount as AI adoption accelerates. New AI security features protect data and privacy while enabling powerful AI capabilities. The balance between accessibility and security represents one of the key challenges enterprises face.

The hybrid deployment model emerging across enterprises reflects practical realities. Companies want the power of frontier models but need to protect proprietary data and maintain compliance. This drives demand for solutions that seamlessly blend cloud AI services with on-premises deployment options.

What's particularly notable is how quickly AI has moved from experimental budget to core IT spending. The creation of C-suite roles specifically for AI indicates this isn't seen as a temporary trend but a fundamental shift in how enterprises operate. The aggressive talent acquisition and development strategies suggest companies understand that AI capability will become a core competitive differentiator.

Key Takeaways for AI Engineers

AI Budgets Now Exceed Security Spending: This priority shift signals that AI is viewed as essential infrastructure rather than experimental technology.

Hybrid Deployment Models Dominate: Enterprises need solutions that work across cloud and on-premises environments, with strong data governance capabilities.

Productivity Tool Integration is Table Stakes: AI must seamlessly integrate with existing workflows through familiar tools like Microsoft 365 rather than requiring new interfaces.

Action Items:
- Design AI solutions with enterprise governance requirements in mind from day one, including data residency and compliance features
- Focus on integration with existing enterprise tools and workflows rather than standalone AI applications

Robotics and Brand New Devices Are Coming

The convergence of AI advances with new hardware forms promises to reshape how we interact with technology. From smart glasses to humanoid robots, we're seeing the physical manifestation of AI's capabilities in forms that would have seemed like science fiction just years ago.

Tesla's autonomous driving progress provides a glimpse of embodied AI in action. They posted a video of Full Self-Driving navigating the complex Arc de Triomphe roundabout in Paris, preparing for the robotaxi launch in Austin. This real-world navigation of one of traffic's most challenging scenarios demonstrates how far computer vision and decision-making have advanced.

NVIDIA's robotics platform shows the infrastructure emerging to support this new era. At Computex 2025, they announced Isaac GR00T N1.5, the first major update to their open, customizable foundation model for humanoid reasoning and skills. The synthetic motion data blueprint accelerates robot training, addressing one of the key bottlenecks in robotics development.

Dennis Hassabis's insights at Google I/O connect the dots between AI model improvements and robotics. He emphasized how adding vision to Google's models is fundamental for creating helpful robots. Robots need to understand the world around them to help in daily life. This vision capability developed for language models directly enables more capable robotic systems.

Tesla's robot demonstrations grow increasingly impressive. Elon Musk shared a video showcasing remarkable progress in their humanoid robot's capabilities. The fluid movement and task completion demonstrate how quickly the field advances when combined with modern AI.

Smart glasses emerge as the consensus next platform. Google's XR smart glasses prototype uses Gemini AI for real-time translation and contextual environmental information, as reported by The Korea Herald and ZDNET. Project Astra forms the foundation for these glasses, as confirmed by Gadgets360.

Holographic telepresence systems, detailed in Google's I/O coverage, use camera arrays to project holograms during video calls, creating more immersive telepresence experiences.

The most intriguing development might be OpenAI's acquisition of Jony Ive's startup for $6.5 billion. Ive and his design firm LoveFrom will lead creative and design work at OpenAI. This could help OpenAI compete with Apple in consumer hardware. The io staff of around 55 engineers, scientists, researchers, physicists, and product development specialists join OpenAI.

The Wall Street Journal reports OpenAI plans to ship 100 million AI 'companions' capable of being fully aware of users' surroundings and lives. They aim to create a third core device after computers and mobile devices. The device will be unobtrusive, fitting in pockets or on desks. It won't be a phone, with design goals including weaning users from screens.

Ming-Chi Kuo suggests the OpenAI device will be slightly larger than Humane's discontinued AI Pin but as compact and elegant as an iPod Shuffle.

Apple's smart glasses launching in 2026 will include cameras, microphones, and AI capabilities. The timing suggests Apple recognizes the risk of being left behind in this new platform shift.

The convergence of improved models, better hardware, and innovative form factors creates unprecedented opportunities. Vision capabilities developed for AI models directly enable robotic applications. Smart glasses promise to make AI assistance ambient and contextual. Novel devices from companies like OpenAI could redefine our relationship with technology entirely.

Research papers on robotics flood in weekly, while demonstrations from companies like Tesla show practical progress. The gap between research and reality narrows rapidly. We're approaching an inflection point where AI-powered robots and devices transition from laboratory curiosities to everyday tools.

Key Takeaways for AI Engineers

Vision Models Enable Robotics Breakthroughs: Improvements in multimodal AI models directly translate to more capable robotic systems, making computer vision expertise increasingly valuable.

New Form Factors Require New Interfaces: Smart glasses and AI companions demand rethinking user interaction patterns beyond screens and traditional inputs.

Hardware-Software Integration Becomes Critical: Success in this space requires deep integration between AI models and purpose-built hardware, favoring full-stack approaches.

Action Items:
- Explore computer vision and spatial computing frameworks to prepare for the shift toward embodied AI applications
- Consider how your applications might adapt to screenless or ambient computing environments enabled by new device categories

Artificial Code

Discussion about this post