AI Weekly Update - October 25

Claude can now control your computer, OpenAI's new model is imminent (maybe)

what to know for now

🖱️ Claude 3.5’s new computer use skill. Anthropic’s Claude 3.5 Sonnet can now navigate computers, follow screen-based commands, and emulate user actions through virtual interaction, a step that greatly broadens AI application scope. This capability is currently in public beta, with built-in safety measures addressing misuse and potential cyber threats like prompt injection. Read more

💼 Copilot’s new agent capabilities boost productivity. Microsoft 365 Copilot expands its reach with autonomous agents in Dynamics 365, aimed at streamlining workflows in sales, supply chain, and customer service. These agents automate complex tasks, enabling users to transform business processes and maximize impact. Read more

🚀 xAI Launches Grok API to Developers. xAI’s Grok API, featuring its beta model “grok-beta,” provides minimal capabilities for now but offers function-calling integrations with databases and external tools. At $5/million input tokens and $15/million output tokens, xAI is positioning itself against OpenAI and Anthropic. Read more

🖥️ Claude gains coding prowess advantage. Claude now writes and runs JavaScript code, leveraging a new analysis tool to deliver mathematically precise, reproducible answers by analyzing complex data and creating interactive visualizations. Available on the web, this feature positions Claude against Google’s Gemini and OpenAI’s Advanced Data Analysis. Read more

🎨 Newly released Stable Diffusion 3.5 upgrades model capability. Stable Diffusion 3.5 debuts in powerful variants suited for researchers, creators, and enterprises, supporting both fine-tuning and high-quality image generation. Large, Large Turbo, and Medium models offer scalable performance, optimized for consumer hardware and open use under Stability AI’s license. Read more

🎭 Runway launches model for character animation performance. Act-One enables character animations from simple video and voice inputs, preserving complex eye-lines, micro-expressions, and realistic pacing. This single-camera setup replaces traditional multi-step facial animation, supporting lifelike emotions across diverse character designs. Read more

🧪 AI Research of the Week

Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models
From OpenAI

Jake’s Take: OpenAI introduces a new way to train Consistency Models (CMs) — a type of fast, efficient generative AI model. By simplifying the model’s design and stabilizing the training process, the authors have scaled it up successfully to handle massive datasets, producing high-quality images with fewer steps than older methods. Their tweaks, like improved model structures and smarter training techniques, help these models achieve almost the same quality as state-of-the-art methods without the typical complexity.

If these models gain traction, they could affect how quickly and effectively AI generates realistic images, impacting industries from gaming to design.

what to know for later

🧑‍💻 OpenAI teases Orion amid denial. OpenAI hints at releasing its powerful new model, Orion, by December, limited to business partners via API access. Despite Altman calling reports "fake news," insiders suggest Orion aims to be 100 times more powerful than GPT-4. Read more

🔍 AGI readiness lead resigns at OpenAI. Miles Brundage, OpenAI's AGI readiness lead, resigned, citing the company's unpreparedness for AGI's impact. Brundage, who helped define a five-step AGI maturity scale, warns that neither OpenAI nor society is equipped for advanced AI capabilities approaching human-level reasoning. Read more

📱 Apple launches AI preview with ChatGPT integration. Apple introduced beta features in iOS 18.2, integrating ChatGPT for advanced responses and adding Genmoji, Image Playground, and Image Wand. While Siri lacks in-app action control, its ChatGPT access enhances Visual Intelligence by recognizing and translating text in real time. Read more

🛡️ Biden doubles down on AI security. Biden’s new memo reinforces national security with AI, targeting China competition and securing AI chip supply chains. Civil rights groups criticize the move for potentially intensifying surveillance capabilities. Read more

🔍 Meta unveils self-checking AI model. Meta’s new “Self-Taught Evaluator” model, which leverages the "chain of thought" technique for structured problem-solving, promises to autonomously validate other AI outputs. Using exclusively AI-generated data, Meta aims to minimize human intervention in AI evaluation, potentially reducing the dependency on human feedback in future AI development. Read more