Multiverse Playground for Robots and Actual General Intelligence
The field of artificial intelligence is growing and expanding on a weekly basis. In the first few weeks of 2025 we have seen some impressive announcements from Sam Altman (OpenAi) and last week Jensen Huang (Nvidia) at CES 2025. This is setting the stage for some amazing innovation towards actual general intelligence and autonomous robots.

Are we on the verge of a major breakthrough with robotics? It certainly seem so. At CES 2025 in Las Vegas, Jensen Huang unveiled exciting new chips and software capabilities designed to power the next generation of autonomous vehicles and robots.
“The ChatGPT moment for General Robotics is just around the corner.”
Jensen Huang
More on that later in this article but lets take a look at how we got here.
Evolution of Graphic Processing Unit (GPU)
Nvidia has long been at the forefront of computer graphics, and this technology has played a pivotal role in the rise of AI and large language models.
- 1999 Initially the GPU was designed for computer graphics in video games (GeForce 256)
- 2009 GPUs were repurposed for cryptocurrency mining and encryption, accelerating blockchain technology.
- Nov 2022 GPUs became integral in training large language models, such as GPT, which use neural networks to analyze data, answer questions, and generate coherent, creative content.
From its early days as a graphics tool, the GPU has evolved into the backbone of today’s cutting-edge AI chipsets.
The challenge of AI Training: “Attention is all you need.”
Most people have likely tried ChatGPT and been impressed by its results. ChatGPT (Chat Generative Pre-trained Transformer) has been trained by processing vast amounts of content. To put this in perspective, the data used to train ChatGPT-3 is roughly equivalent to a human reading nonstop for 2,600 years. Computers clearly have the advantage when it comes to processing large datasets, provided they have the resources.
Large language models like ChatGPT predict the next word in a sentence based on context. During training, the model compares its predictions to the actual outcomes and refines itself through numerous iterations, resulting in the sophisticated version of ChatGPT we know today.
In 2017, Google engineers introduced the paper “Attention Is All You Need,” which proposed a new method for training massive neural networks more efficiently. This led to the development of the Transformer architecture, allowing AI to scale by processing parallel streams focused on producing the most accurate answers. This innovation has been essential for creating efficient large language models.
The race to Actual General intelligence (AGI)
In September 2024, Sam Altman introduced OpenAI’s O1 model. O1 excels in complex reasoning tasks, such as those in science and mathematics, by taking additional time to generate a “chain of thought” before producing an answer. This marks a significant leap forward, as previous models often “hallucinated” answers or struggled with intricate problems.
By December 2024, Altman announced the O3 model, which had advanced so rapidly in just three months that he claimed it was approaching AGI (Artificial General Intelligence). This remarkable progress—overcoming benchmarks previously thought to take decades—signals that the path to AGI is accelerating faster than anyone expected. This leads to a discussion on what benchmarks that are solvable by humans but challenging for AI should be used.
Training AI models with synthetic data
This brings us to Nvidia’s recent announcement, which highlights an emerging solution to a significant challenge in AI: the need for training data. As Elon Musk pointed out at the end of 2024, the available data for training AI models has already been exhausted. The AI systems have consumed all publicly available human-created data, yet more is needed. To address this, Nvidia is leveraging synthetic data—data generated by AI itself.
Jensen in his talk explained that we need three computers to deliver safe AI:
- DGX – (Deep learning GPU training system) A computer to train AI models
- Omniverse with Cosmos – A computer to simulate the data in simulated real world environments
- AGX – A computer to actually drive the autonomous vehicel or robot in the real world.
For text-based models like ChatGPT, LLMs (Large Language Models) can rely on vast datasets from world literature. However, autonomous vehicles require far more specialized data. Even though there is a wealth of real-world data from test-driving cars, it’s still insufficient.
What’s particularly impressive about Omniverse with Cosmos is its ability to generate entirely new training data by simulating different virtual realities. This innovation allows AI models to train in environments that closely replicate real-world conditions, which is critical for autonomous systems.

Autonomous Vehicle safety application
Testing autonomous vehicles in real-world environments requires the creation of artificial data representing various scenarios. Using Omniverse with Cosmos, Nvidia has tested its autonomous vehicle code to the equivalent of 15,000 engineer years and 2 million test cases. This level of testing meets the highest safety standards, including the ASIL D classification.
ASIL D, an abbreviation of Automotive Safety Integrity Level D, refers to the highest classification of initial hazard (injury risk) defined within ISO 26262 indicating the strictest safety measures to avoid potential harm in hazardous situations.

The next wave of Physical AI driven robot models
The future of AI is rapidly taking shape, moving beyond theoretical concepts into tangible, real-world applications. From tools that help with spelling, grammar, and translation to solving the most complex global challenges, AI’s potential is limitless. We are now on the cusp of a robotic revolution powered by AI—one that will reshape industries and everyday life.
Further reading/listening:
Greater detailed explanation on O3