Introducing k0rdent: The Open-Source Distributed Container Management Environment | Learn More

The State of AI: Key Developments, March 2025

Justin Winter - March 17, 2025

- AI Ethics, Embodied AI, Open Source AI, Advanced AI Research, Diffusion Models, AI Voice Technology, AI Agents, AI Developments, AI Trends, AI

Welcome to a new series of regular blog postings rounding up AI developments around the world. As you know, things in the AI space are changing fast and furiously. At Mirantis, we have a very active group of AI/ML researchers who are keeping up with the latest developments, and we want to make sure that we are sharing them with you!

Today’s roundup is the usual update of models and agents, but also has some more interesting updates around enabling accessibility and robotics. There’s a little bit here for everyone.

Let’s get to it …

Next-Generation Multimodal AI Models

The competitive landscape for multimodal AI has intensified dramatically in early 2025, with Google's Gemma 3 surpassing DeepSeek V3 in Elo rankings while maintaining open-source accessibility, OpenAI's GPT-4.5 (nicknamed "Chonky Orion") deploying approximately 5 trillion parameters to establish new reasoning capabilities, Anthropic's Claude 3.7 Sonnet introducing enhanced "thinking" features for complex reasoning tasks, and smaller models like QwQ-32B achieving performance parity with much larger systems like DeepSeek R1-671B through innovative architecture and training approaches.

AI Agents and Developer Platforms

The focus on agentic AI is growing, with new systems now "thinking multiple steps ahead" by integrating native tool use — such as search, code execution, and real-time action—with advanced reasoning capabilities. Google's Gemini 2.0 update introduces prototypes like Project Astra and Project Mariner, which are being trialed for autonomous web tasks and real-world interactions. At the same time, third-party ecosystems, such as OpenAI's Agents SDK and LangChain's frameworks, are fostering new applications that automate complex workflows.

AI Voice Technology Breakthroughs

Recent advancements in AI voice synthesis have led to significant breakthroughs, enhancing the naturalness and expressiveness of machine-generated speech. Notably, Sesame, a startup led by Oculus co-founder Brendan Iribe, has introduced an AI voice assistant named "Maya," offering human-like interactions that surpass existing offerings like ChatGPT's voice mode and xAI's Grok 3. Additionally, Moshi, developed by Kyutai Labs, is a speech-to-speech model capable of transforming input speech into different voices or languages while preserving the original speaker's emotional tone and intent. Similarly, ElevenLabs has made significant strides in AI voice technology, including the release of Eleven v2 Turbo, an English-exclusive model combining high-quality speech synthesis with low latency, and Multilingual v2, supporting 28 languages with emotionally rich speech generation.

Sources:

Diffusion Models and Advanced AI Research

The research landscape has seen remarkable innovations with LLaDA (Large Language Diffusion Models) combining the structured reasoning of language models with the generative power of diffusion processes, Process Reinforcement through Implicit Rewards (PRIME) enabling models to learn from implicit feedback derived from task execution rather than explicit reward signals, "Titans: Learning to Memorize at Test Time" addressing the fundamental limitation of static knowledge bases by enabling dynamic information incorporation during inference, and OpenAI's Inverse Optimal Inference (IOI) with Nvidia's specialized Kernels producing models with "near-superhuman" coding abilities.

Open Source Ecosystem and Accessibility

The open-source AI ecosystem continues to democratize advanced capabilities, with Allen AI's OLMoCR, representing a significant breakthrough in Optical Character Recognition (OCR) technology by combining the power of large language models with visual processing to accurately extract and understand text from complex documents, screenshots, and images with challenging layouts. This innovative open-source tool excels at detecting text in diverse formats including tables, diagrams, and multi-column documents, while offering impressive accessibility benefits through its ability to interpret handwritten notes and convert visual text to machine-readable formats that work seamlessly with screen readers and other assistive technologies.

OLMoCR has been made available across multiple platforms, including a dedicated macOS version that enables local processing of sensitive documents without requiring cloud connectivity, demonstrating how the open-source community continues to deliver specialized AI tools that rival or exceed proprietary alternatives while prioritizing user privacy and cross-platform availability.

Robotics & Embodied AI

The integration of advanced AI with physical systems has accelerated dramatically, with platforms like NEO Gamma representing significant steps forward in home robotics. Building upon foundations established by agentic AI systems like Project Astra and Project Mariner, which bridge virtual reasoning with physical action, improvements in object detection, multi-step dexterity tasks, and autonomous decision-making are enabling robots to operate more effectively in unstructured environments.

Notably, Tesla's Optimus robot has evolved through multiple generations, with the latest version, Gen 3, demonstrating enhanced capabilities such as advanced dexterity and full autonomy. This model features hands with increased degrees of freedom, allowing for precise movements, and is designed to integrate with over 1,000 tasks applicable to both home and industrial settings. Additionally, Boston Dynamics' Atlas robot has transitioned to a fully electric model, showcasing advanced mobility and manipulation skills.

These developments represent the convergence of large language models, computer vision, and specialized robotics training, poised to transform industrial automation, personal assistance, and healthcare applications throughout 2025.

Infrastructure and Computational Resources

The scale of infrastructure investment for AI development has reached unprecedented levels, exemplified by Project Stargate — a proposed $500 billion datacenter representing approximately 1.7% of US GDP — while Anthropic's successful $61.5 billion Series E funding round demonstrates the financial sector's confidence in AI's transformative potential, providing resources for long-term research agendas and potentially challenging larger competitors through focused innovation, all while research into computational efficiency like Google's Donut Method (GDM) aims to maximize performance without proportional increases in infrastructure scale.

Competitive Dynamics and Market Evolution

The AI market has entered a phase of intense competition characterized by rapid-fire model releases and capability announcements creating a constant state of technological leapfrogging, with Google's Gemma 3 and 2.0 Flash outperforming competitors in key benchmarks, Chinese AI companies like DeepSeek briefly claiming the #1 position on the US App Store while causing a 17% decline in Nvidia's stock value, and open-weight models challenging commercial systems by offering comparable capabilities at a fraction of the deployment cost – creating a highly dynamic ecosystem where today's leader may quickly become tomorrow's follower and rewarding continuous innovation over static technological advantages.

Ethical and Practical Implications

The acceleration of AI capabilities raises profound questions about ethical, social, and economic implications, with increasingly autonomous agent systems that can operate web browsers, navigate digital environments, and potentially control physical systems introducing new challenges around responsibility and oversight, while the concentration of computational resources in massive projects raises questions about environmental impact and technological inequality between organizations – all while systems like Bespoke-Stratos and Sky-T1 represent a "Vicuna+Alpaca moment for reasoning" that could democratize advanced reasoning capabilities previously confined to the largest models.