Google's AI Advancements: Introducing PaLM 2 and Beyond
When reflecting on the most significant breakthroughs in AI over the past decade, Google has consistently been at the forefront. Our pioneering work in foundation models has become the cornerstone of the industry, powering AI-driven products that billions use daily. As we continue to responsibly advance these technologies, there is enormous potential for transformative applications in diverse fields, from healthcare to human creativity.
Throughout our decade-long journey in AI development, we've learned that scaling up neural networks unlocks immense possibilities. We've witnessed surprising and impressive capabilities emerge from larger models. However, research has shown us that it's not just about size; creative approaches in model architecture and training are crucial. Recent advancements have revealed the importance of multimodality, incorporating human feedback, and building models more efficiently. These insights are key as we strive to push the boundaries of AI, creating models that genuinely benefit people in their everyday lives.
Introducing PaLM 2
Building on this foundation, we are excited to introduce PaLM 2, our next-generation language model. PaLM 2 represents a state-of-the-art advancement with enhanced multilingual, reasoning, and coding capabilities.
Multilinguality
PaLM 2 has been extensively trained on multilingual text, covering over 100 languages. This training has significantly enhanced its ability to understand, generate, and translate nuanced text — including idioms, poems, and riddles — across various languages, solving a traditionally difficult problem. PaLM 2 also achieves mastery-level proficiency in advanced language exams.
Reasoning
With a wide-ranging dataset that includes scientific papers and web pages containing mathematical expressions, PaLM 2 demonstrates improved logic, common sense reasoning, and mathematical capabilities.
Coding
PaLM 2 was pre-trained on a vast amount of publicly available source code datasets, excelling in popular programming languages like Python and JavaScript, and generating specialized code in languages like Prolog, Fortran, and Verilog.
As we continue to develop PaLM 2, we will introduce version updates as the model is integrated into various products.
A Versatile Family of Models
PaLM 2 is not only more capable but also faster and more efficient than previous models. It comes in various sizes, making it easily deployable for a wide range of use cases. PaLM 2 will be available in four sizes: Gecko, Otter, Bison, and Unicorn. Gecko is lightweight enough to operate on mobile devices, supporting interactive applications even when offline. This versatility allows PaLM 2 to be fine-tuned to support diverse products, benefiting more people.
Powering Over 25 Google Products and Features
At I/O today, we announced over 25 new products and features powered by PaLM 2, bringing the latest AI advancements directly to our products for consumers, developers, and enterprises worldwide. Here are a few examples:
- PaLM 2’s enhanced multilingual capabilities are expanding Bard to new languages and powering our recently announced coding update.
- Workspace features like Gmail and Google Docs are utilizing PaLM 2 to help users write, organize, and work more efficiently.
- Med-PaLM 2, trained by our health research teams, can answer complex medical questions and summarize dense medical texts. It was the first large language model to achieve "expert" level on U.S. Medical Licensing Exam-style questions. We're adding multimodal capabilities to integrate information like x-rays and mammograms to improve patient outcomes. Med-PaLM 2 will open up to a small group of Cloud customers for feedback later this summer.
- Sec-PaLM, a specialized version of PaLM 2, is designed for cybersecurity use cases, helping analyze potentially malicious scripts and detect threats more effectively. Available through Google Cloud, it represents a potential leap in cybersecurity analysis.
- Since March, we've been previewing the PaLM API with a small group of developers. Starting today, developers can sign up to use the PaLM 2 model, or customers can use it in Vertex AI with enterprise-grade privacy, security, and governance. PaLM 2 also powers Duet AI for Google Cloud, a generative AI collaborator designed to help users learn, build, and operate faster than ever before.
Advancing the Future of AI
PaLM 2 demonstrates the impact of versatile, highly capable models that bring real benefits to everyone. While we are committed to releasing the most helpful and responsible AI tools today, we are also focused on creating the best foundation models for the future.
Our Brain and DeepMind research teams have achieved many defining moments in AI over the past decade, and we are bringing these world-class teams together into a single unit, Google DeepMind, to accelerate progress. Supported by Google's computational resources, Google DeepMind will deliver incredible new capabilities in everyday products while responsibly paving the way for the next generation of AI models.
We are already working on Gemini, our next model designed from the ground up to be multimodal, highly efficient at tool and API integrations, and built to enable future innovations like memory and planning. Gemini is still in training, but it's already exhibiting multimodal capabilities never before seen in previous models. Once fine-tuned and rigorously tested for safety, Gemini will be available in various sizes and capabilities, just like PaLM 2, ensuring it can be deployed across different products, applications, and devices for everyone's benefit.