Enlarge / There’s so much AI news this week that covering it can feel like running through a hall full of dodgy cathode ray tubes, like this Getty Images illustration.
This week, thanks to OpenAI, we have a blog post from controversial CEO Sam Altman, a broader rollout of Advanced Voice Modes, rumors of 5GW data centers, key staff restructuring, and dramatic restructuring plans. It’s been a very busy week in the news.
But the rest of the AI world doesn’t follow the same rhythm, doing its own thing and churning out new AI models and research by the minute. Here’s a roundup of other notable AI news from the past week.
Google Gemini updates
On Tuesday, Google announced updates to its Gemini model lineup, including the release of two new production-ready models that reiterate past releases: Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002. The company reported noticeable improvements in math, long context processing, and vision tasks, as well as improved overall quality. Google claims 7% performance improvement on the MMLU-Pro benchmark and 20% improvement on math-related tasks. But as anyone who’s been reading Ars Technica for a while knows, AI benchmarks aren’t usually as helpful as we might hope.
In addition to model upgrades, Google introduced significant price reductions for Gemini 1.5 Pro, reducing the cost of input tokens by 64 percent and the cost of output tokens by 52 percent for prompts under 128,000 tokens. As AI researcher Simon Willison stated in his blog: “For comparison, GPT-4o currently costs $5 per input (1 million tokens) and $15 per month of output, compared to Claude 3.5 Sonnet At $3 per million inputs and $15 per 100,000 outputs, the Gemini 1.5 Pro is already the cheapest Frontier model and now it’s even cheaper.
Google has also increased the rate limits, so Gemini 1.5 Flash now supports 2,000 requests per minute and Gemini 1.5 Pro can now handle 1,000 requests per minute. Google reports that the latest model has twice the output speed and three times lower latency than the previous version. These changes may make it easier and more cost-effective for developers to build applications with Gemini than before.
Meta announces Llama 3.2
On Wednesday, Meta announced the release of Llama 3.2. This is a significant update to our lineup of open weight AI models, which we have covered extensively in the past. The new release includes vision-enabled large language models (LLMs) with 11 billion and 90B parameter sizes, and lightweight text-only models with 1B and 3B parameters designed for edge and mobile devices. Meta says the vision model competes with leading closed-source models for image recognition and visual understanding tasks, while the smaller model outperforms similarly sized competitors for a variety of text-based tasks. It is claimed that it will demonstrate.
Willison conducted some experiments with several small 3.2 models and reported impressive results for the size of the models. AI researcher Ethan Mollick showed off running Llama 3.2 on his iPhone using an app called PocketPal.
Meta also introduced the first official “Llama Stack” distribution, created to simplify development and deployment across a variety of environments. As with previous releases, Meta makes its models available for free download with licensing restrictions. The new model supports long context windows of up to 128,000 tokens.
Google’s AlphaChip AI speeds up chip design
On Thursday, Google DeepMind unveiled what appears to be a major advance in AI-driven electronic chip design: AlphaChip. It started as a research project in 2020 and is now a reinforcement learning method for designing chip layouts. Google said it used AlphaChips to create “superhuman chip layouts” in the past three generations of Tensor Processing Units (TPUs), GPU-like chips designed to speed up AI operations. It has been reported. Google claims AlphaChip can generate high-quality chip layouts in hours, compared to weeks or months of human effort. (Nvidia is also reportedly leveraging AI in chip design.)
Notably, Google has also released AlphaChip’s pre-trained checkpoints on GitHub and made the model weights publicly available. The company reported that AlphaChip’s impact has already spread beyond Google, with chip design companies like MediaTek adopting and building on the technology in their chips. Google says AlphaChip sparks a new field of research in AI for chip design, with the potential to optimize every step of the chip design cycle, from computer architecture to manufacturing.
That’s not all that happened, but these are some big highlights. We’ll see what next week holds, as the AI industry shows no signs of slowing down at the moment.
