OpenAI thinks, Google executes

The race to achieve artificial general intelligence (AGI) has intensified. After Google’s DeepMind recently launched Veo and Imagen3 on Google Cloud, this time it has also introduced Genie2. Genie2 is a large model that can create many types of interactive 3D environments.

Artificial worlds

Genie2 simplifies the process of developing visual AI agents. By turning an image into an interactive virtual world, this model provides the possibility of exploring these spaces for both humans and artificial intelligence through keyboard and mouse.

In a blog post, Google DeepMind announced: “Genie2 could enable AI agents in the future to be trained and evaluated in an endless world of new environments. “This research also paves the way for new and innovative processes for prototyping interactive experiences.”

“The model of the world of artificial intelligence is taking shape,” said Demis Hessabis, CEO of DeepMind.

The new version of Genie2 is based on its predecessor, Genie1. The previous version was limited to 2D environments, but now Genie2 has added new capabilities in 3D environments. This model has the ability to simulate physical interactions, make complex animations and create environments with real physics, proper lighting and real interactions between objects. Trained on a large video dataset, Genie2 uses a new technology called “Autoregressive Latent Diffusion” to generate frames sequentially based on user reactions.

DeepMind has stated that games have always played an important role in the company’s AI research, from early experiments with Atari games to breakthroughs such as AlphaGo and AlphaStar.

However, one of the main problems in training more general AIs has been the limitation in diverse training environments. Genie2 solves this problem by providing “long-term horizon stability”, enabling the simulation of changing scenarios and allowing AIs to dynamically explore environments.

Google DeepMind also introduced an artificial intelligence agent called SIMA that performed various tasks in the world created by Genie2 by following natural language commands. The company stated that SIMA’s performance demonstrates the model’s ability to create unique test environments for artificial intelligence and demonstrates the ability to generalize to new and unknown tasks.

Another feature of Genie2 is its ability to generate new content on the fly, so it can maintain a static and identical world for up to a minute. The model is also capable of generating different perspectives, such as first-person and isometric views, and can simulate real-world environments using images.

Weather forecasting and AGI

In addition, Google DeepMind has introduced a new model called GenCast. This AI model improves weather forecasts, providing faster and more accurate forecasts up to 15 days ahead, while also removing uncertainties and risks associated with the weather.

Google recently introduced its experimental AI model, Gemini-Exp-1121, which competes with OpenAI’s GPT-4 model. The company is also preparing to launch Google Gemini-2, which is expected to compete with the upcoming OpenAI model called o1.

Artificial intelligence critic Gary Marcus said in an exclusive interview with AIM that DeepMind seems to be on a better path to AGI than its competitors. However, he noted that no company has yet found a definitive path to AGI. “Of the big companies working on this, DeepMind is probably on the right track,” he added.

“DeepMind has done the most interesting work in this field,” Gary Marcus continued. He also added: “Although AlphaFold is not as designed for general applications as GPT-4, it demonstrates the success of the application of neural-symbolic AI in the field of protein folding.”

Christmas has come early for OpenAI

After celebrating ChatGPT’s second anniversary, OpenAI has announced plans to introduce new models and features over the next twelve days. The company wrote on its official account on X: “12 days. 12 live broadcasts. A collection of new things, big and small. The 12 days of OpenAI will start tomorrow.”

Sam Altman, CEO of OpenAI, wrote on X: “Starting tomorrow at 10am PT, we’re kicking off 12 days of OpenAI. Every weekday we’ll have a live stream with an unveiling or demo, some big and some smaller. We have some great stuff to share, hope you enjoy! Merry Christmas”

According to Altman, ChatGPT has more than 300 million weekly active users. The company plans to almost quadruple this number within the next year and aims to reach 1 billion users. Additionally, according to OpenAI statistics, 1 billion messages are sent through ChatGPT every day, and 1.3 million developers in the United States have worked on OpenAI.

According to sources, OpenAI seems to be planning to release its long-awaited model for video generation, called Sora, and its full reasoning model, called o1. It’s also possible that the company will eventually launch its own audio engine and show off its own web browser. Recently, OpenAI hired Darrin Fisher, one of the original developers of the Google Chrome browser and former vice president of engineering at Google.

Also, reports indicate that OpenAI is planning to release its own artificial intelligence agent called “Operator”, which is expected to be released in January 2025. Speculations suggest that this feature will help ChatGPT to monitor and observe desktops as an agent.

This year, OpenAI introduced many new features including ChatGPT search for up-to-date answers, Canvas UI for writing and coding, advanced audio version with different accents, lower cost GPT-4o mini model, o1 preview and o1 mini version for reasoning. introduced

Greg Brockman, co-founder and president of OpenAI, recently returned from leave. However, in a recent post on X, he revealed that during this time he was teaching basic DNA models in collaboration with the Arc Institute.

Greg Brockman wrote in his post: “I have long believed that deep learning should provide major changes in the field of medicine and health for humans and animals. In my opinion, this project is an important step towards this goal and is considered a fundamental step for modeling virtual cells.”

Season of transfers

In other news, Xiaohua Zhai, a longtime Google DeepMind researcher, announced that he is joining OpenAI. In a post on X, he wrote: “After an amazing decade at Google DeepMind, I’m excited to announce that I’ll soon be joining OpenAI! I am excited about the opportunity to collaborate on building the OpenAI office in Zurich.” Zhai will work with OpenAI researchers Jeff Hinton and Alexander Klesnikov to establish and expand the Zurich office.

OpenAI recently hired Keith Roach as its first Chief Marketing Officer. “His experience in consumer and business marketing will shape how we approach the transformative power of AI,” the company said.

Interestingly, in the last 18 months, OpenAI has recruited 85 employees from Google, half of whom are engineers. This shows that Altman believes in hiring experienced talent and prefers to use experienced people instead of hiring 24-year-old programmers.

source