Global AI Weekly Issue 80

AI Model Development & Capabilities:
• Deliberative Alignment for Safer LLMs: OpenAI has introduced a new alignment strategy called "deliberative alignment" for its o-series models. This approach directly teaches reasoning LLMs about safety specifications, allowing them to analyze user prompts, identify relevant safety policies, and generate safer responses. The o1 model outperforms other state-of-the-art LLMs on safety benchmarks using this technique.

• Vocal Imitation AI: MIT researchers have developed an AI model that can produce and understand human-like vocal imitations of everyday sounds without prior training on human vocal impressions. This model could lead to new sonic interfaces and more human-like AI characters.

• Synthetic Data for AI Training: With human data for AI training being exhausted, AI companies are turning to synthetic data generated by AI models for training purposes. However, concerns about "model collapse" and "hallucinations" are also emerging with this approach.

• GPT Model Evolution: GPT-2 and GPT-3 models shifted the AI paradigm from "pre-training plus fine-tuning" to "pre-training only," using task-agnostic learning, the scale hypothesis, and in-context learning. The models showed that larger models trained on larger datasets could develop new capabilities automatically, including the ability to perform tasks with no fine-tuning.

AI in Business & Industry:
• Deterministic GenAI Chatbots: A new approach to developing GenAI chatbots in regulated industries focuses on ensuring accuracy by only providing pre-approved answers. This method uses subject matter experts to review and approve answers before they are given to customers.

• AI-Powered Business Tools: AI is becoming essential for businesses, with tools like ChatGPT, Jasper, Midjourney, Claude, Synthesia, and Runway Gen-2 revolutionizing industries. AI is enhancing marketing strategies through hyper-personalization, visual content creation, and real-time campaign analysis. Customer service is also being transformed by AI through multimodal assistants and 24/7 support.

• AI for Construction and Manufacturing: Multimodal AI is enabling new applications in industries like construction and manufacturing, with AI being able to spot flaws using visual data. For example, drones with AI can inspect buildings to find defects.

Ethical & Regulatory Considerations:
• Ethical Implications of AI: AI raises critical ethical questions regarding bias and fairness, privacy and surveillance, accountability and transparency, autonomy and control, job displacement, human rights, human interaction and long-term risks.

• OpenAI's AI Regulation Blueprint: OpenAI has published an "economic blueprint" for AI, advocating for U.S. government action to support the AI industry, including increased funding for infrastructure. The blueprint also proposes developing best practices for model deployment, streamlining engagement with national security agencies, and establishing export controls.

• AI Regulation and Compliance: Governments worldwide are starting to create AI regulations, such as the EU AI act. Companies must ensure compliance with laws and regulations, and there is a growing focus on responsible AI development practices.

Other Notable Developments:
• Microsoft's New AI Organization: Microsoft has created a new internal organization, CoreAI, focused on accelerating AI infrastructure and software development.

• Knowledge Graphs: Knowledge graphs are becoming an important method for information retrieval, particularly in combination with RAG systems. LLMs are now being used to construct knowledge graphs more efficiently.