AI Daily Roundup: December 4, 2025

TL;DR


🏢 Company Announcements

GeForce NOW added 30 new games including Hogwarts Legacy to its cloud streaming service. The update coincides with a “Half-Price Holiday” sale.

Google Photos launched 2025 Recap, allowing users to explore and customize year-end photo summaries. The feature is now available to all users.

NVIDIA announced that top 10 most intelligent open-source models use mixture-of-experts (MoE) architecture, running 10x faster on Blackwell NVL72. Models include Kimi K2 Thinking and Mistral Large 3.

Hugging Face published a blog on fine-tuning open-source LLMs using Claude. The guide covers technical implementation steps.

OpenAI researchers introduced “confessions” as a method to train models to admit mistakes and undesirable behavior. The approach aims to improve transparency in model outputs.

OpenAI acquired Neptune to deepen visibility into model behavior and strengthen experiment tracking tools. The acquisition supports research operations.

OpenAI announced $40.5M in unrestricted grants to 208 nonprofits through the People-First AI Fund. The fund supports community innovation and opportunity.

OpenAI and Mirakl partnered to integrate ChatGPT Enterprise into commerce workflows. The collaboration aims to enable agent-native commerce through Mirakl Nexus.

OpenAI awarded up to $2M in grants for research on AI and mental health. The program supports projects studying real-world AI risks and benefits.

OpenAI and NORAD launched three ChatGPT holiday tools for “NORAD Tracks Santa”. Features include custom Christmas stories and toy coloring pages.


📰 Top Stories

Anthropic signed a $200M deal to provide its LLMs to Snowflake’s 12,600 customers. The agreement expands access to Snowflake’s enterprise clients.

EU Commission launched an antitrust investigation into Meta’s policy banning rival AI chatbots on WhatsApp. The policy restricts competitors from using WhatsApp’s business tools.

Apple’s top apps of 2025 include AI features without a dedicated AI app or chatbot. AI was highlighted across multiple 2025 award-winning applications.

Nexus Venture Partners allocated $350M from its new $700M fund to Indian startups. The firm manages $3.2B across multiple funds.

Amazon Web Services launched Kiro Powers with integrations for Stripe, Figma, and Datadog. The tool enables AI coding assistants to access specialized workflow expertise.

Gong reported sales teams using AI generate 77% more revenue per rep. The study found 70% of enterprise revenue leaders now trust AI for business decisions.

MIT Technology Review published a report on security implications of AI strategy adoption. The report highlights increasing data, velocity, and variety of security threats.

OpenAI’s LLM confession research was featured in MIT Technology Review’s daily newsletter. The newsletter covered testing methods for exposing model processes.

Startup Zanskar announced AI tools to uncover hidden geothermal energy resources. The technology identifies underground hot spots obscured by surface geology.

Nuclear reactors provide 20% of U.S. winter electricity demand according to MIT Technology Review. The report details seasonal trends in U.S. power generation.


🔧 Tools & Development

OpenAI launched a new tool for AI-assisted coding with Stripe, Figma, and Datadog integrations. The Kiro Powers system targets developer workflow bottlenecks.

Flowchart2Mermaid is a vision-language model system converting flowcharts to editable Mermaid.js code. The tool enables editing of process diagrams shared as images.

STRIDE is a framework for selecting AI modalities including agentic AI, AI assistants, or LLM calls. The framework helps determine when multi-step agent functionality is necessary.

The 4/$\delta$ Bound paper proposes a framework for predictable LLM-verifier systems. The approach aims to improve reliability in formal verification with LLMs.


📚 Further Reading