The Week in AI: OpenAI Enhances Fine Tuning, JetBrains Integrates Local AI Models, Replit Unveils Coding Assistant, More

This edition of our weekly roundup of AI product and services news includes OpenAI's enhancements to its fine-tuning API; Stability AI's release of Stable Audio 2.0: Cohere's launch of its latest large language model, Command R+;  CodiumAI's new AI support for software developers; and more.

OpenAI launched new enhancements of its fine-tuning API, which was introduced last August, to streamline the optimization of its large language models (LLMs) for specific tasks. The updates are designed to improve the fine-tuning process, which allows customers to augment OpenAI's language models with proprietary data, enhancing the model's ability to handle specialized inquiries. One key upgrade enables developers to save iterations of an AI model throughout the fine-tuning stages, or epochs, facilitating recovery from errors without starting over. Another addition is the Playground UI, where developers can assess the performance of fine-tuned models at various stages. OpenAI also refined its AI fine-tuning dashboard to offer greater customization of models' hyperparameters—settings crucial for response accuracy. This revamped dashboard now integrates with third-party AI tools, starting with Weights and Biases, for more in-depth data analysis.

Stability AI has released Stable Audio 2.0, an enhanced AI system for sound-clip generation. Incorporating a Transformer neural network, developed by Google in 2017, Stable Audio 2.0 recognizes and replicates large-scale structures for higher quality musical output. It comes with new features built on its predecessor, which was introduced last September. Stable Audio 2.0 can create audio files up to 180 seconds long, incorporating, not just text prompts, but also user-supplied sound clips to refine the generated audio's style to match specific needs. It can also produce structured compositions with intros, development, and outros, as well as sound effects. The upgrade stems from advancements in the AI's diffusion model architecture, particularly in the generation of latent spaces, which are efficient data structures that streamline training by focusing on key details and omitting extraneous information. Currently available for free on a dedicated website, Stability AI plans to make Stable Audio 2.0 accessible through an API for integration into other applications, the company said.

AI startup Cohere launched its latest large language model (LLM), "Command R+," which the company said is tailored for real-world business applications. Command R+ is the most potent and scalable LLM Cohere has developed, the company said, surpassing its predecessor, Command R, in both cost efficiency and performance. This version of the LLM was designed to enhance enterprise workflows through features such as categorization, automation, and data analysis. The new model boasts advanced retrieval-augmented generation (RAG) for improved accuracy and reduced errors, and supports more than 10 languages. It's also capable of complex, multi-step reasoning and tool use, crucial for tasks like customer relationship management and order fulfillment. Cohere has partnered with Microsoft to make Command R+ available on Microsoft Azure now, with plans to extend to Oracle Cloud Infrastructure and other platforms, soon.

GenAI-powered code-testing platform provider CodiumAI launched a new AI tool to aid developers in improving software quality. The "proactive AI Agent," similar to GitHub's Copilot, not only suggests tests for newly written code but also generates fixes for errors and inefficiencies on the fly. Leveraging OpenAI's GPT-4, CodiumAI's solution addresses costly software issues, which, according to a 2020 study, have impacted U.S. businesses significantly. The Israeli startup's tool, unveiled after an $11 million seed funding round a year ago, enhances traditional coding processes by integrating quality testing into the development workflow, offering real-time analysis and recommendations. CodiumAI's co-founder and CEO Itamar Friedman envisions a future where developers and AI agents collaborate closely to ensure high-quality, correct code.

Developer tools startup Replit unveiled an AI coding assistant, "Replit Teams," aiming to take on  GitHub with efficiency-boosting tech for software developers. This new tool within Replit's browser-based IDE suggests real-time coding fixes and improvements without requiring prompts from devs. It's designed to enhance productivity, allowing developers to accept or dismiss suggested code edits. Replit, known for its accessible web-hosted IDE, enables collaboration and immediate coding startup with no setup. Building on its Ghostwriter Chat, the new feature is part of its broader strategy to compete in the AI coding market, with its platform already serving over 20 million developers. The beta version of Replit Teams is now available, offering collaborative coding with limited access.

Software development toolmaker JetBrains announced the integration of local AI models for code completion into its IDEs, allowing developers access to immediate full-line code completion and prediction, with the new AI models operating entirely on-device. JetBrains' latest offering is an advanced iteration of its full-line code completion technology available in its paid IDE subscriptions. The new local AI feature supports a range of programming languages including Java, Kotlin, Python, JavaScript, TypeScript, CSS, PHP, Go, and Ruby across JetBrains' IDEs. Upcoming updates will add languages such as C#, Rust, and C++. The company also provides an AI assistant for more extensive coding tasks, including test generation and code refactoring, which requires cloud connectivity due to its reliance on larger language models.

Facebook parent company Meta announced that, starting in May, it will implement "Made with AI" labels on AI-generated videos, images, and audio across its platforms, broadening an initiative that initially targeted a limited range of doctored videos, said Monika Bickert, vice president of content policy, in a blog post. The company said it will introduce distinct and more conspicuous labels for digitally altered content deemed to pose a considerable risk of significantly misleading the public on important matters, irrespective of whether the content was produced using AI or other methods. This new strategy will alter Meta's handling of manipulated content, shifting from removing select posts to maintaining the material online while providing users with details about its creation process.