Unleash the Future: Google's Gemini Pro Surpasses GPT-4, Meta's Ambitious Llama 4 Plan

Meta is aiming to develop the most advanced AI model in the industry by 2025. They plan to train their upcoming Llama 4 model on 10 times more data than Llama 3, which they claim is already competitive with the most advanced models.

Zuckerberg stated that Meta would rather build too much compute capacity than not enough, as they are planning for the compute and data needed for the next several years. The amount of compute required to train Llama 4 will likely be almost 10 times more than what was used for Llama 3, and future models will continue to grow beyond that.

This ambitious goal means that Llama 4 will need to outperform the latest models from Google, Anthropic, OpenAI, and others. It remains to be seen if Meta can achieve this, as the AI race continues to heat up with rapid advancements across the industry. However, Meta's willingness to invest heavily in compute and data suggests they are serious about maintaining their position as a leader in large language models.

According to Adam D'Angelo, the CEO of Quora and a board director at OpenAI, artificial general intelligence (AGI) may be achieved within the next 5 to 15 years. D'Angelo made this prediction during a recent event, stating that the advent of AGI will be a very important change for the world.

OpenAI, the company behind the popular language model GPT, has internally developed a new five-level classification system to track its progress towards building AGI. The first three levels include:

1. Chatbots with conversational language abilities.
2. Reasoners and systems with human-level problem-solving skills.
3. Agents and systems that can take actions.

D'Angelo's prediction suggests that even before reaching the full AGI milestone, the achievement of human-level problem-solving and action-taking capabilities will be "game-changing" events that could significantly transform the world.

Given the rapid advancements in AI witnessed in recent years, the prediction of AGI within the next 5 to 15 years, though ambitious, is considered within the realm of possibility by industry experts. The next 5 years, in particular, are expected to see an acceleration in AI development as more of the world's top research labs and companies focus their efforts on this challenge.

However, it remains to be seen if any major roadblocks or technical hurdles will arise on the path to AGI. The race to achieve this milestone is intensifying, and the impact of its realization could be profound, making it a crucial area to monitor in the coming years.

Google's new experimental model, Gemini Pro 0801, has been tested in the chatbot arena over the past week, gathering over 20,000 community votes. For the first time, Gemini has claimed the number one spot, surpassing GPT-4 and CLAUDE 3.5 with an impressive score of 1,300 and also achieving the top position on the vision leaderboard.

Gemini Pro excels in multilingual tasks and delivers robust performance in technical areas, hard prompts, and coding. This is a significant achievement, as Gemini 1.5 Pro has managed to outperform the highly capable GPT-4 and CLAUDE 3.5 models.

Interestingly, Google has not labeled this model as Gemini 2, suggesting that they may have implemented some additional reasoning or post-training techniques to enhance the model's capabilities. This approach is similar to what Anthropic has done with CLAUDE 3.5, where the model demonstrates improved reasoning abilities compared to previous versions.

The performance of Gemini Pro 0801 highlights the ongoing advancements in the chatbot arena, with models continuously pushing the boundaries of what is possible. It will be interesting to see how long Gemini Pro 0801 can maintain its top position and whether OpenAI or other AI companies will respond with even more capable models in the near future.

Nvidia is working to simplify and accelerate the development of humanoid robots with its Project Roo initiative. The company is introducing a set of tools for developers in the humanoid robot ecosystem to build their AI models more efficiently.

The key components of Nvidia's approach include:

1. Synthetic Data Generation Pipeline: Nvidia starts with human-collected demonstrations using mixed reality devices like the Apple Vision Pro. They then multiply this data by a thousand or more using Nvidia's simulation tools like Omniverse, RoboSuite, and MimicGen.

2. Distributed Computing Infrastructure: Nvidia is leveraging its DGX, OVX, and Jetson Thor computing platforms to power the development workflow. The DGX handles the processing of videos and text to train the multimodal foundation model, the OVX runs the simulation stack, and the Jetson Thor is used for testing the model on real robots.

3. Omniverse-Powered Simulation: Nvidia's Omniverse simulation framework, integrated into the Isaac Lab, allows developers to generate a massive number of environments and layouts to increase the diversity of the training data.

4. Generative AI-Enabled Tools: Nvidia's MimicGen tool helps generate large-scale synthetic motion data sets based on the small number of original captures, further expanding the training data.

The goal is to enable developers worldwide to build better AI models for humanoid robot hardware platforms. Nvidia believes the era of "physical AI" is here, where robots can understand and interact with the physical world.

By simplifying the development workflow and providing powerful computing infrastructure, Nvidia aims to accelerate the progress in humanoid robotics and bring us closer to the age of AI-powered humanoid robots.

Researchers at ICML 2024 presented a new prompt engineering technique called "Plan Like a Graph" that can significantly improve the performance of language models on complex, multi-step tasks.

The key insight behind this technique is that current language models struggle with asynchronous planning - the ability to parallelize certain subtasks while sequentially executing others. To address this, the "Plan Like a Graph" method prompts the model to first generate a graph representation of the task, capturing the dependencies between subtasks. The model can then use this graph to devise an optimal plan for completing the overall task.

The researchers found that this approach outperformed baseline methods across a variety of language models. For example, on a task involving making breakfast (e.g. brewing coffee, frying an egg, making toast), the "Plan Like a Graph" method reduced the total time to complete the task by over 20% compared to sequential planning.

This work highlights that there is still significant untapped potential in language models, and that novel prompt engineering techniques can unlock new capabilities. As the researchers note, this is an "off-the-shelf prompt engineering method" that requires no additional training, making it an accessible way to boost model performance.

Overall, the "Plan Like a Graph" technique represents an important advance in language model capabilities, particularly when it comes to complex, multi-step reasoning. As language models continue to evolve, we can expect to see more innovative prompt engineering approaches that push the boundaries of what these systems can achieve.