March 25 AI News.
Google releases Gemini 2.5 Pro experimental AI image generation hits new heights with startup Reve and OpenAI’s 4o image generation and Figure Robotics makes a training breakthrough. Here’s today’s AI news. Google has released Gemini 2.5 Pro experimental, a thinking model that Google of course claims is their most intelligent model ever.
Featuring a 1 million token context window, Gemini 2.5 Pro is already available in Google AI Studio and the Gemini app. We’ve already seen some absolutely amazing outputs from the model and Incredibly, it scored 18.8% in the humanities last exam benchmark.
Google seems to be continuing its amazing winning streak. The battle for the world’s greatest AI image generation really heated up today. This morning, all the talk was about Reve, launched by a startup of the same name. From out of nowhere, this model immediately jumped to the top of the pile with many proclaiming it as the new king of image generation.
Its ability to produce high quality images, accurate text rendering and adherence to complex prompts had everyone absolutely stunned. It felt like an out of the blue DeepSeek moment for image generation. What’s more, you can use Reve image completely free right now.
But wait. This afternoon, OpenAI launched 4o image generation integrated into ChatGPT and Sora. This model is not only producing some amazing quality images, it’s also leveraging OpenAI’s LLMs to help understand the exact nature of the user’s request.
It also seems to be able to go some way to achieving the holy grail of image generation, character consistency. Both of these models really seem to raise the standard for image generation in their own slightly different ways. And we haven’t even seen Gemini 2.5 image creation yet.
Remember when AI images struggled with fingers? Yeah. Those days are long gone. Finally, in the battle of the bots, Figure Robotics has revealed that they have made a significant breakthrough with their figure 02 robot, which now walks with a gait that they say closely resembles a human.
This was achieved through reinforcement learning in a high fidelity physics simulator where extensive data was processed rapidly and then transferred directly to real hardware without additional tuning. So this represents both a breakthrough in the walking mechanics and and a breakthrough in the training methods.
It looks like we can expect rapid progress from this point.