AI is pushing boundaries faster than ever, changing how we think, create, and interact. This week’s highlights showcase breakthroughs in deep reasoning, creative content, and powerful AI hardware.

REER & DeepWriter-8B
A new training method flips AI thinking — starting from answers, it traces the reasoning steps backward. DeepWriter-8B, built on this approach, shows creativity and logic that compete with top models like GPT-4o and Claude 3.5.

Tencent Hunyuan Image
Turn simple prompts into stunning images, cinematic videos, and detailed 3D models — instantly. Tencent’s Hunyuan platform makes AI-driven creativity faster and more accessible than ever.

Arm Lumex CSS
Arm’s latest compute platform boosts AI on phones and PCs, delivering up to 5x faster performance. It powers smarter assistants, real-time translation, and immersive gaming — all with seamless integration.

AI is no longer just a tool. It’s transforming how we create, think, and interact — and the future is unfolding now.

Novel Paradigm for Deep Reasoning in Large Language Models

REverse-Engineered Reasoning (REER) is a new paradigm for training large language models (LLMs) to perform complex, open-ended generation by working "backwards" from high-quality answers to reconstruct the deep, step-by-step reasoning that could have produced them. By leveraging this scalable, gradient-free approach, the team curated and open-sourced DeepWriting-20K, a unique dataset containing 20,000 deep reasoning trajectories for creative writing and open-ended tasks. DeepWriter-8B, their LLM trained on this dataset, demonstrates deliberate, structured thinking and delivers results that outperform top open-source models—often rivaling proprietary giants like GPT-4o and Claude 3.5 in various creative benchmarks.

Tencent Hunyuan Unveils Next-Gen Image, Video & 3D Generation Platform

Tencent’s newly launched Hunyuan Image platform is a breakthrough in AI-powered creative tools, offering high-quality image, video, and even 3D content generation based on simple text or image prompts. Powered by advanced large models, Hunyuan enables users to instantly create stunning visuals, immersive 3D assets, and cinematic videos for everything from marketing and e-commerce to gaming and design workflows. The platform introduces features like image-to-video, interactive 360° views, detailed 3D model export, and a user-friendly interface that streamlines content creation for creators, brands, and businesses. With support for multiple languages and fast, high-fidelity outputs, Hunyuan Image sets a new standard for AI-driven creativity, now accessible to all users through the Tencent Hunyuan website.

Arm Lumex CSS: Next-Gen AI Power for Mobile & PCs

Arm has introduced its most advanced compute subsystem platform, Arm Lumex CSS, designed to accelerate AI experiences on flagship smartphones and next-gen PCs, delivering up to 5x faster AI performance with SME2-enabled CPUs. Lumex unites high-performance CPUs, GPUs like the new Mali G1-Ultra with advanced ray tracing, and system IP optimized for 3nm nodes, enabling real-time AI use cases such as voice translation, smarter assistants, personalized applications, and desktop-class mobile gaming. Deep integration with major AI frameworks like PyTorch ExecuTorch and Google LiteRT ensures seamless acceleration without code changes, supporting broad developer adoption and cross-platform portability from Android to Windows on Arm. With partners including Alibaba, Samsung, Tencent, and Apple, Arm Lumex sets a new foundation for personal, private, and high-performance on-device AI across billions of devices by 2030, fundamentally transforming mobile and consumer technology with unprecedented speed, privacy, and efficiency.

Automate smart, personalized comments, generate viral posts, and track real-time growth all powered by AI. Trusted by 5000+ businesses and creators, Olly helps you skyrocket leads, engagement, and followers across Instagram, LinkedIn, TikTok, and more.

Try Now

Top AI Products from this week

• Cosine CLI – Cosine CLI brings our AI coding agent to your terminal. It writes/refactors/tests code, runs shell, understands your repo, and continues browser tasks. Autonomous or interactive. Full access to familiar local tools. Meeting developers where they work.

• Plander.ai – Plander.ai your AI-powered personal project manager. Break down big goals into actionable steps, integrate seamlessly with your calendar, and adapt dynamically to changes. Plander helps you "Land your plan & Win your game.

• Oboe – Oboe is the world’s first generalized AI-powered learning platform. Oboe lets anyone instantly create educational courses to learn about any topic in a more personalized, efficient, and fun way than ever before.

• Corgi AI - Corgi is your true AI read-later for everything you save — videos, articles, PDFs — a second brain that digests, organizes, and chats with your saved content. No more content graveyards.

• Rocket Journal – Introducing Rocket Journal: An AI-powered voice journal to decode your thoughts, uncover patterns, and unlock clarity — all in just minutes a day. Speak your mind in a safe rant space or reflect deeply with guided prompts. On iOS only. Coming soon to Android.

• Stable Audio 2.5 – Stable Audio 2.5 is a new audio model from Stability AI built for enterprise sound production. It delivers fast, high-quality, structured tracks in seconds, with advanced control features like audio inpainting for professional workflows.

This week in AI

• Stellantis + Mistral AI - Stellantis partners with Mistral AI to launch AI-powered in-car assistant and boost product, manufacturing efficiency, and customer experience with advanced LLMs.

• Nvidia’s AI GPU for Massive Context - NVIDIA Rubin CPX GPU delivers 30 petaflops compute, 128GB GDDR7 memory, & 3x faster attention, built for million-token AI workloads in coding & generative video tasks.

• Optimal Seven Senses for Memory Capacity - A new mathematical model suggests seven senses maximize memory capacity, aiding AI, robotics, and our understanding of human memory beyond our usual five senses.

• Lilly TuneLab AI for Drug Discovery - Google’s NotebookLM creates flashcards, quizzes, reports, and audio overviews from your documents, turning passive reading into active learning and deeper understanding.

• OpenAI gpt-realtime - OpenAI's gpt-realtime is a new speech-to-speech model in the Realtime API that offers natural, expressive voice interaction with image input, SIP calling, and improved instruction following.

Paper of The Day

OM2M is a dual-process AI model combining a fast, habitual graph convolutional network (System 1) with a slower, context-aware meta-adaptive controller (System 2) that dynamically balances intuitive and deliberative reasoning via a learned context gate. It replicates human cognitive biases like anchoring, priming, fatigue, and framing effects without bias-specific tuning, achieving strong generalization across diverse, ambiguous social reasoning tasks. This architecture advances flexible, interpretable, and adaptive Theory of Mind reasoning in AI, enabling efficient belief revision and bias mitigation inspired by human cognition.

To read the whole paper 👉️ here.