Google and Meta Enhance AI Models to Compete with Rising “AlphaChip” Tech

This week has been bustling with AI developments, largely driven by OpenAI’s activities, including a controversial statement by CEO Sam Altman, the broad deployment of Advanced Voice Mode, speculation about a new 5GW data center, significant personnel changes, and ambitious restructuring initiatives.

However, the AI industry at large continues to operate independently, rapidly producing new AI models and insights. Here’s a recap of some other significant AI developments that unfolded over the past week.

Updates to Google Gemini

This Tuesday, Google revealed enhancements to its Gemini model series, launching two new models ready for production: Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002. These models build on previous versions and deliver better overall performance, particularly in mathematics, extended context management, and vision-based tasks. Google reports a 7 percent improvement in the MMLU-Pro benchmark and a 20 percent boost in math tasks. However, as long-time readers of Ars Technica may realize, AI benchmarks don’t always translate to real-world utility.

Alongside these model enhancements, Google has significantly reduced prices for the Gemini 1.5 Pro, slashing costs for input tokens by 64 percent and for output tokens by 52 percent for prompts under 128,000 tokens. AI researcher Simon Willison highlighted on his blog that, “For comparison, GPT-4o is priced at $5 per million tokens for input and $15 per million for output, and Claude 3.5 Sonnet charges $3 per million for input and $15 per million for output. Gemini 1.5 Pro was already the most affordable among the leading models, and now it’s even more budget-friendly.”

Google has also increased transaction limits, with the Gemini 1.5 Flash now able to handle 2,000 requests per minute and the Gemini 1.5 Pro up to 1,000. The latest versions boast double the output speed and one-third the latency of their predecessors, making it easier and more economical for developers to integrate Gemini into their applications.

Meta Introduces Llama 3.2

Meta announced on Wednesday the launch of Llama 3.2, an important update to its series of open-weights AI models, which we have frequently covered. The update includes vision-capable large language models (LLMs) with 11 billion and 90 billion parameters, and smaller, text-only models with 1 billion and 3 billion parameters, optimized for edge and mobile use. Meta claims these vision models rival the top closed-source models in image recognition and visual understanding tasks, while the smaller models surpass their competitors in various text-based benchmarks.

Simon Willison reported notable results from his tests with the smaller, 3.2 models. Meanwhile, AI researcher Ethan Mollick demonstrated the capabilities of Llama 3.2 on his iPhone using an application named PocketPal.

Meta also rolled out its first official “Llama Stack” distributions, which are designed to streamline development and deployment across different platforms. As with previous versions, Meta is offering these models for free download, albeit with certain licensing restrictions. These new models support extended context windows up to 128,000 tokens.

Google’s AlphaChip Enhances Chip Design

On Thursday, Google DeepMind announced a major breakthrough in AI-driven electronic chip design with AlphaChip. Originating as a research project in 2020, AlphaChip employs a reinforcement learning approach to create chip layouts. Google has utilized AlphaChip to design highly efficient layouts for the last three generations of its Tensor Processing Units (TPUs), which are specialized chips designed to accelerate AI tasks. Google claims that AlphaChip can generate optimal chip layouts within hours, a task that typically takes humans weeks or months to complete. (Nvidia is reportedly also employing AI in chip design.)

Significantly, Google has also made a pre-trained checkpoint of AlphaChip available on GitHub, sharing the model weights with the broader public. Google noted that AlphaChip’s influence has already reached beyond its labs, with companies like MediaTek using and enhancing the technology for their chip designs. According to Google, AlphaChip has initiated a new wave of research in AI for chip design, potentially revolutionizing every phase of the chip design process from computer architecture to production.

While these are just a few of the highlights, they represent the rapid pace and continued innovation within the AI sector. We’ll see what the next week brings as the industry shows no signs of slowing down.