The Week in AI: Microsoft, Anthropic, Google Deep Mind, Cognition and More

Nvidia made a ton of announcements last week at its GTC Conference, and you can see our coverage on this site, but they weren't the only AI vendors with news to share. This edition of our weekly roundup of AI product and services news includes an autonomous AI software engineer, an open-source Python library for intervention-based research on ML models, an automated, AI-driven software development framework, a football (soccer) tactics assistant, and more.

Anthropic announced that Claude 3 Haiku and Claude 3 Sonnet are now available on the Google Cloud Vertex AI machine learning (ML) platform. Vertex offers access to Gemini multimodal models from Google, and lets users train and deploy ML models and AI. Enterprises will be able to use the Claude family of foundational models to optimize intelligence, speed, and cost with Google Cloud's robust infrastructure and tools, the company said. This collaboration enables businesses to prototype and scale generative AI solutions quickly with enterprise-grade data privacy and security. The benefits for Google Cloud customers include the ability to keep their data within their existing cloud environment, simplify data governance, reduce operational costs and complexities, and more easily manage access permissions. "The launch of Claude 3 models on Google Cloud means more organizations can now adopt safer, more efficient AI solutions within their preferred cloud environment," the company said.

Microsoft introduced AutoDev, a fully automated, artificial intelligence-driven software development framework. AutoDev was designed to allow AI agents to handle a wide range of software engineering tasks independently, including complex code edits, thorough testing, and advanced Git operations. The framework was engineered for autonomy, efficiency, and security, and it employs Docker containers to streamline development while protecting user privacy and maintaining project integrity, according to Microsoft researchers. AutoDev relies on AI agents with the ability to take on complex software engineering tasks using a variety of tools and processes that enhance their autonomy. These agents can edit files, compile code, and conduct tests. This type of automation positions AI as a central figure with the aim of freeing developers to focus on more strategic tasks. (More here.)

A group of Google Deep Mind researchers proposed an AI football (soccer) tactics assistant developed and evaluated in close collaboration with domain experts from Liverpool FC. Dubbed TacticAI, it focuses on analyzing corner kicks "as they offer coaches the most direct opportunities for interventions and improvements," the researchers wrote in a paper published on Nature Communications. TacticAI uses predictive and generative A and employs a geometric deep learning approach to help create generalizable models. The tactics assistant was developed and evaluated with experts from Liverpool Football Club as part of a multi-year research collaboration. TacticAI’s suggestions were preferred by human expert raters 90% of the time over tactical setups seen in practice, the company said.

Cognition, an artificial intelligence startup backed by Peter Thiel's venture capital firm Founders Fund and tech industry notables, such as former Twitter executive Elad Gil and DoorDash co-founder Tony Xu, unveiled "Devin," a software development assistant the company is billing as "a fully autonomous AI software engineer." Unlike existing coding assistants (such as GitHub Copilot), Devin reportedly has the ability to manage an entire development project from start to finish, including writing code, debugging, and executing projects. It's even capable of handling projects on Upwork, the startup has demonstrated. The launch signifies a pivotal change in AI-assisted development by providing engineers with a comprehensive AI collaborator, beyond the limited assistance of existing tools.

Stanford University researchers have introduced pyvene, an open-source Python library to support intervention-based research on machine learning models. Pyvene aims to enable detailed interventions on PyTorch models, surpassing existing tools that often lack flexibility, expandability, and ease of use. At its core, pyvene employs a configuration-based approach to interventions, straying from the usual code-executed methods and providing a user-friendly, adaptable means of altering model states. The library manages a variety of intervention types, such as static and modifiable parameters, to meet diverse research demands. Notably, pyvene supports complex intervention schemes, including sequential and parallel interventions. It can apply these at different points in a model's decoding process, enhancing the library's utility for generative model research where the dynamics of model output generation are of special interest.

Sakana AI, an AI startup based in Tokyo, today announced a new method for creating generative AI models, which the company says is similar to evolutionary processes. The company calls the method "evolutionary model merging." This approach employs so-called evolutionary algorithms to find ways to build new foundational models by merging a wide range of open-source models with diverse capabilities, the company said. "We are taking on the challenge of creating new methods to streamline, sophisticate, and automate the development of basic models," the post stated. This method can automatically create a new underlying model that is good at the capabilities specified by the user, the company explained in a blog post. Models can be created very efficiently because it leverages the vast collective intelligence of existing open models. Company researchers published a paper describing this method, "Evolutionary Optimization of Model Merging Recipes."

The University of California, Irvine, is extending the use of its new, tailored generative artificial intelligence tool, ZotGPT Chat, to students this year. Launched in January for all faculty and staff, ZotGPT Chat's rollout accompanies a university-wide initiative to enhance AI literacy among the campus community. Developers are enhancing ZotGPT Chat to include web-enabled responses, image creation, and custom chatbots tailored to departmental data or websites. Additionally, the tool, accessible via mobile phones and voice chat, will soon allow faculty and staff to incorporate it into their programs and research efforts. "ZotGPT Chat gives UC Irvine researchers a powerful platform to explore the potential of generative AI in ways that protect the integrity of our data while simultaneously providing us access to ‘industrial-strength’ tools," university Chief Innovation Officer Errol Arkilic said in a statement. "With its access, we can cost-effectively run comprehensive experiments across a wide range of applications."