The AI Arms Race: Exploring the Latest Advancements in Large Language Models and AI-Generated Content

The real story this week is the news around new large language models becoming available or soon to be released.

Google announced that Gemini 1.5 is now available in over 180 countries, featuring a 1 million token context window - the equivalent of around 750,000 words. This expanded context window is a major improvement over previous models. Developers can now access Gemini 1.5 via the API to build applications.

In response, OpenAI announced an improved GPT-4 Turbo model is now available through the API and rolling out in ChatGPT. While details are limited, it is reportedly better at coding and math compared to previous versions.

The open-source world is also heating up. Stability AI released Stable LLM2, a 12 billion parameter model. Meanwhile, Anthropic quietly released a new 22 billion parameter Mixture of Experts model, Mixr 8X 22B, as a torrent download.

Google also announced new open-source Gemma models - one fine-tuned for coding, and another designed for efficient research. Additionally, Meta is expected to release the highly anticipated LLaMA 3 model, which is rumored to be on par with GPT-4 in capability but open-sourced.

This flurry of large language model announcements demonstrates the rapid pace of innovation in this space, with both closed-source and open-source models continuously pushing the boundaries of what's possible.

It seems that all the major tech companies building large language models are trying to reduce their reliance on Nvidia GPUs. Nvidia currently dominates the market for GPUs used in AI training, but Google, Intel, and Meta are all introducing their own custom AI chips.

Google introduced their Axion processors at the Google Cloud Next event. Intel unveiled their Gaudi 3 AI chip, which they claim has 40% better power efficiency than Nvidia's H100 GPUs. Meta announced their second-generation MTI (Meta Training and Inference) accelerator chip, which they say has 3 times improved performance over the first generation.

Meanwhile, at Nvidia's GTC event earlier this year, they announced their next-generation Nvidia Blackwell chips, which are supposedly 4 times more powerful than the current H100 GPUs. This shows that Nvidia is still far ahead in terms of raw compute power for AI training.

While these new custom chips from Google, Intel, and Meta aim to reduce reliance on Nvidia, it remains to be seen whether they can truly catch up to Nvidia's latest advancements. The race is on to develop the most powerful and efficient AI hardware.

During the Google Cloud Next event, the tech giant unveiled several exciting AI-powered video creation tools that are poised to transform the industry.

Image-In 2: Google's answer to tools like Dolly and Firefly, Image-In 2 can generate not just static images, but also short animated GIFs and clips. These text-to-live-image capabilities allow users to create engaging, looping visuals with ease.

Google Vids: Described as a "PowerPoint-style" video generator, Google Vids uses AI to create slide-based videos from scripts or prompts. The resulting videos mimic the aesthetic of professional presentation software, making it a valuable tool for creating polished, AI-driven video content.

Magic Time: Developed by a research team, Magic Time is a specialized video generator focused on creating high-quality time-lapse footage. By simply providing a prompt, users can generate visually stunning time-lapse videos of scenes like growing plants or construction projects. The open-source code and Hugging Face demo make Magic Time accessible for experimentation and integration into various video workflows.

These AI-powered video tools demonstrate the rapid advancements in generative capabilities, empowering creators to streamline their video production processes and explore new creative avenues. As the technology continues to evolve, the impact of these AI-driven video generators on the content creation landscape is poised to be significant.

According to the transcript, a new bill has been introduced to the U.S. Congress that aims to force artificial intelligence companies to reveal the copyrighted material they use to train their generative AI models. The key points are:

• The bill would require AI companies to file a report about the copyrighted material they used to train their models, at least 30 days before releasing the AI model.

• This is seen as a move to increase transparency, as some of the biggest tech companies like Google, Microsoft, and Meta may not want to reveal the data they used for training.

• There are concerns that these powerful companies may lobby against the bill to prevent it from being passed.

• The proposed legislation comes amid growing scrutiny over the training data used by large language models, with reports that OpenAI may have used over a million hours of YouTube videos to train GPT-4.

• The bill is intended to address the lack of transparency around the training data used by AI companies, which could include copyrighted material from various sources.

A card game developer recently made a significant investment in AI-assisted art, paying an AI artist $90,000 to generate card art. While the term "AI artist" may be debated, this approach highlights the potential of AI to assist artists in creating high-quality content at scale.

The developer found that no human artists were able to match the quality of the AI-generated images. However, the process involved more than just pressing a button and letting the AI do the work. The developer then went on to touch up and refine the AI-generated images using Photoshop and other image editing tools, ensuring the colors, consistency, and overall style matched the desired aesthetic.

This approach demonstrates the power of AI-assisted art, where the AI generates the initial concept, and the human artist then polishes and refines the output to achieve the desired result. By leveraging AI, the developer was able to create a large number of card images efficiently, while still maintaining the artistic touch and quality control required for their project.

The success of this endeavor highlights the growing role of AI in the creative industries, where it can be used as a powerful tool to augment and enhance the work of human artists, rather than replace them entirely. As AI technology continues to advance, we can expect to see more examples of this type of AI-assisted art, where the collaboration between humans and machines leads to innovative and high-quality creative outputs.

The rapid advancements in large language models and AI technology are truly remarkable. This week saw a flurry of exciting announcements, from the availability of Gemini 1.5 with its impressive 1 million token context window, to the release of GPT-4 Turbo and the open-source Stable LM2 and Mixr 8X 22B models.

The competition between tech giants to develop their own AI chips and reduce reliance on NVIDIA is another fascinating development, with Google, Intel, and Meta all unveiling new AI-focused processors. The ability to generate animations and videos using AI, as showcased by Google's Image-in-2 and the Magic Time project, is also a significant step forward.

The potential implications of these advancements, both positive and concerning, are vast. The introduction of bills to force AI companies to disclose their training data sources is an important step in addressing transparency and accountability. Meanwhile, Adobe's approach of directly purchasing video content from creators to train their models could be a model for the future.

Overall, the AI landscape continues to evolve at a breakneck pace, with new capabilities and challenges emerging on a weekly basis. As an AI enthusiast, it's an exciting time to follow these developments and consider the ethical and practical implications for businesses, creators, and society as a whole.