Twice a week, Data Points brings you the latest AI news, tools, models, and research in brief. In today’s edition, you’ll find:
- A new automated paper-writing system from Sakana AI
- GitHub’s Autofix uses GPT-4 to find and plug security holes
- Google’s new AI voice assistant
- Anthropic’s prompt caching may reduce developers’ bills
But first:
xAI’s Grok-2 challenges AI leaders with limited release
xAI unveiled Grok-2 and Grok-2 mini, new AI models that it claims outperform competitors like Claude 3.5 Sonnet and GPT-4-Turbo on several benchmarks. The models excel in areas such as graduate-level science knowledge, general knowledge, math competition problems, and visual reasoning tasks. xAI launched beta versions of both models for Premium and Premium+ subscribers on the X platform (formerly Twitter). Grok-2 also employs the new Flux.1 models for image generation. xAI plans to make Grok-2 and Grok-2 mini available through an enterprise API later this month. (xAI)
Google unveils Imagen 3, its top text-to-image AI model
Google released Imagen 3, a new text-to-image AI model with improved image quality, prompt understanding, and versatility across various styles and formats. The model features enhanced text rendering capabilities, better detail capture, and optimized versions for different tasks, from quick sketches to high-resolution images. Imagen 3’s development incorporated Google’s latest safety and responsibility innovations, including data filtering, red teaming, and the SynthID watermarking tool. (Google DeepMind)
Sakana AI develops automated scientific paper production system
Sakana AI unveiled “The AI Scientist,” an automated system that leverages large language models to conduct end-to-end machine learning research. The system generates research ideas, implements them by modifying existing codebases, runs experiments, analyzes results, and produces full scientific papers with citations. It incorporates an automated peer review process that evaluates papers based on top-tier conference standards, providing feedback for iterative improvement. Applied to areas like diffusion models and transformers, The AI Scientist produced papers rated as “Weak Accept” at top machine learning conferences, at a cost of approximately $15 per paper. While still facing limitations like occasional critical errors in result interpretation, the system demonstrates the potential to accelerate scientific discovery by automating the entire research lifecycle. (Sakana AI)
GitHub’s AI assistant combines GPT-4 and code analysis to speed up security fixes
GitHub introduced Copilot Autofix, a new feature in GitHub Advanced Security that uses artificial intelligence to help developers fix code vulnerabilities faster. The tool employs large language models, specifically GPT-4, combined with GitHub’s CodeQL code analysis engine to analyze security issues, explain them, and generate suggested fixes. Copilot Autofix can address various types of vulnerabilities, including SQL injection and cross-site scripting, in both new and existing code. During testing, the tool helped developers fix vulnerabilities more than three times faster than manual methods. This feature aims to make it easier for developers to address security problems, potentially transforming how teams manage code security and reduce their backlog of security issues. (GitHub)
Google first AI giant to release new mobile voice assistant
Google launched Gemini Live, a new conversational AI experience for mobile devices. The feature allows users to have continuous, interruptible conversations with the AI assistant, even when the phone is locked or the app is running in the background. Gemini Live initially rolls out to Gemini Advanced subscribers on Android in English, with iOS and additional language support planned. Google also added 10 new voice options for users to customize their interaction with the AI assistant. (Google)
Anthropic announces prompt caching for its API
Anthropic launched prompt caching in public beta for its Claude 3.5 Sonnet and Claude 3 Haiku models, allowing developers to cache frequently used context between API calls. The feature can reduce costs by up to 90% and latency by up to 85% for long prompts, enabling more efficient use of large context windows in various applications. Prompt caching is particularly useful for conversational agents, coding assistants, and processing large documents, offering significant improvements in speed and cost-effectiveness for AI-powered services. (Anthropic)
Still want to know more about what matters in AI right now?
Read this week’s issue of The Batch for in-depth analysis of news and research.
This week, Andrew Ng discussed how open-source models are helping companies all over the world build momentum for their AI projects:
“Seeing the momentum behind AI in Thailand — where the per capita GDP is around one fifth that of Japan, and one tenth that of the United States — left me feeling that any country, company, or person has a shot at doing meaningful work in the field.”
Read Andrew’s full letter here.
Other top AI news and research stories we covered in depth: AI model prices drop as competition heats up, Black Forest Labs' Flux.1 outperforms top text-to-image models, OpenAI faces financial growing pains, spending double its revenue, and all about TransAgents, a system that boosts literary translation with a multi-agent workflow.