The AI world is moving faster than ever , and this week, it’s all about automation, intelligence, and scale.

💡 Google’s Gemini 2.5 now acts like a real user on the web, filling forms, navigating pages, dragging, dropping, and adapting in real time. It’s not just smart, it works like us.

⚙️ Samsung’s Tiny Recursive Model (TRM) proves bigger isn’t always better, with only 7 million parameters, it’s beating massive models in complex reasoning and puzzle-solving.

🧠 OpenAI’s AgentKit makes building AI agents as easy as drag and drop, a no-code toolkit turning workflows that once took months into projects done in hours.

From small-scale brilliance to full-scale automation, these innovations show one thing:
👉 The next wave of AI isn’t just powerful, it’s practical, fast, and in your hands.

Gemini 2.5 Automates the Web Like Magic

Google DeepMind's Gemini 2.5 Computer Use model brings advanced AI agents to web and mobile interfaces, allowing them to fill forms, manipulate elements, and navigate sites just as people do. It features industry-leading accuracy, low latency, and robust operation in complex web and app scenarios, supporting actions like drag-and-drop and handling dynamic content. The model works in a cyclical loop: it views the screen, acts, checks results, and adapts, supporting parallel operations and advanced error recovery. Developers access the model through Gemini API or Vertex AI, customizing actions, adding safety controls, and deploying it for tasks from UI testing to e-commerce automation. Early applications include large-scale workflow automation, enhanced support bots, and automated data collection for business and research needs.

Samsung’s TRM: Small Scale, High Impact AI

The Tiny Recursive Model (TRM) by Samsung SAIL Montreal is a minimalist recursive reasoning neural network with just 7 million parameters, achieving impressive results on complex tasks like ARC-AGI-1 with 45% accuracy and ARC-AGI-2 with 8%, outperforming many large language models with far more parameters. TRM works by recursively improving its predicted answers through repeated updates to latent reasoning features and outputs, using a single tiny 2-layer network without complex hierarchies or biological inspirations, making it simple yet powerful. It significantly advances state-of-the-art performance on benchmarks such as Sudoku-Extreme, Maze-Hard, and ARC-AGI datasets while being computationally efficient. The approach highlights that smaller, well-designed models with recursive reasoning can compete with much larger foundational models, emphasizing effectiveness over size. The project is open-source under MIT license with detailed code, datasets, and training instructions provided on GitHub.

OpenAI AgentKit: The No-Code AI Builder

OpenAI has launched AgentKit, a comprehensive toolkit designed for developers and enterprises to build, deploy, and optimize AI agents more efficiently. AgentKit consolidates fragmented tools into a unified platform with four main components: Agent Builder, a visual drag-and-drop canvas for creating and versioning multi-agent workflows; Connector Registry, a secure admin panel for managing data connections across tools; ChatKit, which enables embedding customizable chat-based agent experiences into apps and websites; and expanded Evals capabilities for rigorous performance evaluation, including automated prompt optimization and support for third-party models. AgentKit reduces development time, enhances collaboration, and improves agent safety and reliability with modular guardrails, making it easier to deploy robust AI workflows at scale. It builds on OpenAI's Responses API and Agents SDK, aiming to replace complex, manual orchestration with a streamlined, end-to-end solution. Early adopters have sped up agent development from months to hours, drastically cutting iteration cycles and boosting productivity. AgentKit is currently in beta with broad availability planned soon under the standard API pricing model.

Olly is your smart AI sidekick for social media, crafting personalized comments, predicting viral hits, and boosting engagement effortlessly. It adapts to your style, saving time while growing your online presence with authentic, meaningful interactions. Perfect for influencers, agencies, and anyone serious about social media impact.

Try Olly and let AI amplify your voice.

Top AI Products from this week

• QA.tech 1.0 - PromptCompose is a visual prompt engineering tool that helps users create, test, and iterate AI prompts efficiently in one unified interface. It offers instant validation, schema suggestions, mustache syntax highlighting, and support for multiple LLM vendors, streamlining the development of AI-powered apps, assistants, and agents.

• ProhostAI - ProhostAI – all‑in‑one AI assistant for guest messaging, maintenance, cleanings, upsells & guidebooks so vacation rental hosts can scale stress‑free. Available on web, iOS and Android for hosts on Airbnb, Hostaway, Guesty, Hospitable and OwnerRez.

• AI AppGen in Retool - Build production-ready apps from natural language using your real data, in your cloud, with enterprise security and governance built in. Start building with AI that knows your stack.

• Promptius AI - Promptius is an AI-powered development environment designed to build and deploy autonomous agents based on user prompts.

• Webstudio Inception - Inception is a new kind of tool for creative web design exploration with AI that lets you experiment with ideas backed by production-grade code. What you see on the canvas is exactly what you get in production - because it’s real code rendering in your browser

• UX TRIAGE - Vibe-coding creates UX debt. UX Triage catches what AI misses in 30 seconds. Upload screenshots, get Nielsen Norman heuristic analysis, export to Markdown, paste into your AI agent. From analysis to implementation instantly. 95% cheaper than consultants.

This week in AI

• Deloitte and Anthropic Expand Enterprise AI Partnership - Deloitte partners with Anthropic to deploy Claude AI to 470,000+ employees globally, creating a Center of Excellence and certifying 15,000 professionals to scale trusted, compliant AI solutions.

• Google’s Secure AI Framework Updates - Google is enhancing AI security with CodeMender, an AI agent that auto-fixes code vulnerabilities, plus a new AI Vulnerability Reward Program and Secure AI Framework 2.0 for safer AI agents.

• Bagel Labs’ Paris AI - Paris is the world’s first decentralized open-weight diffusion model, trained independently across continents with zero synchronization, using 14× less data and 16× less compute.

• LTX Studio AI Upgrade - LTX Studio lets users create unique AI characters with customizable looks, outfits, and personalities, adding lifelike expressions and dialogue for video stories and films.

• TelecomTS Dataset - TelecomTS is a large-scale multi-modal observability dataset from a 5G network, featuring de-anonymized, zero-inflated time series for anomaly detection & multi-modal reasoning.

Paper of The Day

MixReasoning is a framework for large language models that improves reasoning efficiency by dynamically adjusting reasoning depth within a single response. It switches between detailed reasoning on complex steps and concise inference on simpler steps based on token-level uncertainty. This approach reduces unnecessary verbosity and computational cost while preserving or improving accuracy. MixReasoning uses a lightweight adapter to enable seamless mode switching during inference and is shown to produce shorter, more readable, and more accurate reasoning chains across various benchmarks. The key idea is to "think where it matters," focusing detailed reasoning only on pivotal parts of the problem.

To read the whole paper 👉️ here.