Mar 28, 2025: AI updates from the past week — Gemini 2.5, OpenAI 4o image generation, new reasoning agents from Microsoft, and more

Software companies are constantly trying to add more and more AI features to their platforms, and AI companies are constantly releasing new models and features. It can be hard to keep up with it all, so we’ve written this roundup to share several notable updates around AI that software developers should know about. 

Google releases reasoning model Gemini 2.5, its “most intelligent AI model” yet

Gemini 2.0 Flash Thinking was the company’s first reasoning model, and Gemini 2.5 builds on that with a better base model and improved post-training. In its announcement, Google revealed that all of its future AI models will have reasoning capabilities built in.

The first Gemini 2.5 model is Gemini 2.5 Pro Experimental, and it leads in LMArena benchmarks over other reasoning models like OpenAI o3-mini, Claude 3.5 Sonnet, and DeepSeek R1.

“Gemini 2.5 models are thinking models, capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy. In the field of AI, a system’s capacity for “reasoning” refers to more than just classification and prediction. It refers to its ability to analyze information, draw logical conclusions, incorporate context and nuance, and make informed decisions,” Koray Kavukcuoglu, CTO of Google DeepMind, wrote in a blog post. 

OpenAI announces 4o Image Generation

The latest image generation model improves on text rendering, has the ability to refine images through multiple follow-up prompts, and offers better instruction following, with the ability to handle up to 10-20 different objects in a prompt. 

It can also perform in-context learning to analyze and learn from user-uploaded images, and the model also links its knowledge between text and images to generate better results. 

4o image generation has begun rolling out for Plus, Pro, Team, and Free users as the default image generator, and access will soon be available for Enterprise and Edu users. 

Microsoft unveils new reasoning agents in Microsoft 365 Copilot

The two agents, Researcher and Analyst, can help users analyze vast amounts of information, spanning emails, meetings, files, chats, and more. 

Researcher is ideal for multi-step research, such as building a go-to-market strategy based on both the context of a company’s work and broader competitive data found online. Beyond data in Microsoft 365, it can also leverage third-party connectors to bring in data from sources like Salesforce, ServiceNow, and Confluence.

Analyst is designed for complex data analysis, such as turning raw data from multiple spreadsheets into a demand forecast for a new product or a visualization of customer purchasing patterns. 

These two agents will begin rolling out to Microsoft 365 Copilot subscribers starting in April as part of the Frontier early access program. 

Microsoft Security Copilot gets several new agents

The new agents include a Phishing Triage Agent in Microsoft Defender, Alert Triage Agents in Microsoft Purview, Conditional Access Optimization Agent in Microsoft Entra, Vulnerability Remediation Agent in Microsoft Intune, and Threat Intelligence Briefing Agent in Security Copilot.

The company also announced five additional agents from its Microsoft Security partners: Privacy Breach Response Agent by OneTrust, Network Supervisor Agent by Aviatrix, SecOps Tooling Agent by BlueVoyant, Alert Triage Agent by Tanium, and Task Optimizer Agent by Fletch.

The agents will be available in preview starting in April.

“Building on the transformative capabilities of Security Copilot, the six Microsoft Security Copilot agents enable teams to autonomously handle high-volume security and IT tasks while seamlessly integrating with Microsoft Security solutions. Purpose-built for security, agents learn from feedback, adapt to workflows, and operate securely—aligned to Microsoft’s Zero Trust framework. With security teams fully in control, agents accelerate responses, prioritize risks, and drive efficiency to enable proactive protection and strengthen an organization’s security posture,” Vasu Jakkal, corporate vice president of Microsoft Security, wrote in a blog post.

Red Hat AI offers new capabilities across Red Hat OpenShift AI

Red Hat OpenShift AI 2.18 adds new features such as distributed serving that allows IT teams to split model serving across multiple GPUs, an end-to-end model tuning experience across InstructLab and Red Hat OpenShift AI data science pipelines, and model evaluation.

This release also includes a preview of AI Guardrails, which offer additional methods for identifying and mitigating “potentially hateful, abusive or profane speech, personally identifiable information, competitive information or other data limited by corporate policies.”

Akamai launches new platform for AI inference at the edge

Akamai has announced the launch of Akamai Cloud Inference, a new solution that provides tools for developers to build and run AI applications at the edge.

According to Akamai, bringing data workloads closer to end users with this tool can result in 3x better throughput and reduce latency up to 2.5x.

Akamai Cloud Inference offers a variety of compute types, from classic CPUs to GPUs to tailored ASIC VPUs. It offers integrations with Nvidia’s AI ecosystem, leveraging technologies such as Triton, TAO Toolkit, TensorRT, and NVFlare.

Due to a partnership with VAST Data, the solution provides access to real-time data so that developers can accelerate inference-related tasks. The solution also offers highly scalable object storage and integration with vector database vendors like Aiven and Milvus.

AlexNet source code is now open source

AlexNet is a neural network for recognizing images that was created in 2012 by University of Toronto graduate students Alex Krizhevsky and Ilya Sutskever and their advisor Geoffrey Hinton.

“Before AlexNet, almost none of the leading computer vision papers used neural nets. After it, almost all of them would. AlexNet was just the beginning. In the next decade, neural networks would advance to synthesize believable human voices, beat champion Go players, model human language, and generate artwork, culminating with the release of ChatGPT in 2022 by OpenAI, a company cofounded by Sutskever,” wrote Hansem Hsu, curator of the Computer History Museum Software History Center, the organization that is releasing the source code, in partnership with Google.  

The source code can be found here. 

---

Read last week’s AI announcements roundup here.

The post Mar 28, 2025: AI updates from the past week — Gemini 2.5, OpenAI 4o image generation, new reasoning agents from Microsoft, and more appeared first on SD Times.

from SD Times https://ift.tt/BY76KTC