AI Weekly Briefing – Week Ending 21 Nov 2025)

For the week ending 21 November 2025, AI moved from "chatting" to "contributing to GDP." While Google shipped Gemini 3 and Antigravity to help developers build agents, the real story in Singapore was macroeconomic: AI demand is now visibly lifting national growth.

Below we break it down into Research and Industry, with a specific lens on what matters for organisations in Singapore and Southeast Asia.

🧪 Research – Agentic systems step into sports, science and safety

1. DeepSport: agentic multimodal LLM for sports video reasoning

What happened: A new paper, DeepSport, introduces a multimodal LLM that doesn't just watch video passively—it "thinks with video." Using a tool to actively select and re-visit frames, it uses Reinforcement Learning (RL) to plan its own viewing strategy. On a 6.7k-question benchmark, it outperforms proprietary models by a wide margin. Why it matters (SG/SEA): This is a blueprint for "Video Copilots" in security, manufacturing, and retail. For Singaporean enterprises sitting on petabytes of CCTV footage, this approach turns video into a queryable database where you can ask, "Show me every time safety protocol X was breached," and the agent knows where to look.

2. Agentic Lab: AI “co-embodied” in wet labs & manufacturing

What happened: Researchers released Agentic Lab, a system where AI agents wear AR glasses to guide humans through complex bio-experiments. The system tracks protocols, recognises samples, and gives real-time voice feedback when a human deviates from the SOP. Why it matters (SG/SEA): A huge signal for Singapore's Biomedical & Advanced Manufacturing sectors. This isn't just automation; it's a "Shadow Operating System" for the lab, ensuring compliance and data integrity in real-time.

3. MoRA-RAG: knowledge-grounded agents for disaster resilience

What happened: MoRA-RAG is a new framework designed to read dense, unstructured disaster reconnaissance reports. It uses a "verification loop" to check if its own retrieved evidence is sufficient before answering—reducing hallucinations in safety-critical tasks. Why it matters (SG/SEA): For government agencies handling incident reports, flood risks, or safety audits, this "verify-before-answering" architecture is the key to moving RAG from "experimental" to "deployable."

4. MCP-powered agentic red teaming: “Hiding in the AI traffic”

What happened: A security preprint titled "Hiding in the AI Traffic" shows how attackers can use the Model Context Protocol (MCP)—the same standard used for enterprise agents—to coordinate attacks that look like normal AI traffic. Why it matters (SG/SEA): For CISOs, this is an early warning. If you are deploying agents, your "trusted traffic" channels are now a potential attack vector.

🏗️ Industry & Ecosystem – Gemini 3, National GDP & Sovereign Models

1. Singapore’s "AI Upgrade": GDP Forecast Lifted on Chip Demand

What happened: On Friday (Nov 21), Singapore’s Ministry of Trade and Industry (MTI) upgraded the 2025 growth forecast to ~3.5%, explicitly citing a "surge in demand for AI-related chips and server components." Why it matters: This is the "So What?" for every board member asking about AI ROI. AI is no longer just a cost centre for software trials; it is a macro-economic driver lifting the manufacturing and wholesale trade sectors. The "AI Boom" is now structural to Singapore's economy.

2. Google launches Gemini 3 – and leans fully into agents

What happened: Google launched Gemini 3, its "most intelligent model yet," alongside Antigravity—an agent-first coding environment. Crucially, they emphasized Day-0 integration with open-source tools like LangChain and n8n. Why it matters: Google is finally offering a standardized "Agent Stack." For SG enterprises, this validates the shift away from monolithic chatbots towards modular agents connected to tools.

3. Adobe & Humain: The "Arabic-First" Model Play

What happened: Adobe and Saudi-backed AI firm Humain announced a partnership to build generative AI models specifically for Arabic content, powered by Qualcomm chips. Why it matters: This mirrors the "Sovereign AI" discussions happening in Southeast Asia (like Singapore's "Singlish" model initiatives). It confirms that regional context is a competitive moat. Global models (GPT-5, Gemini 3) are great, but "Culturally Aligned" models are where the specific value lies for government and media.

4. Meta’s “Project Luna”: AI morning brief for 3B users

What happened: Meta is piloting Project Luna, an AI that digests your Facebook feed + external news into a personalised daily brief. Why it matters: This normalises "AI summarisation" as a daily consumer habit. For corporate comms teams, it raises a question: Will your audience read your press release, or will they read the AI's summary of it?

🧭 What this week means for SG & SEA leaders

1. The "AI Economy" is Real (Check the GDP) The MTI announcement is your strongest business case. AI isn't hype; it's physically moving goods and electronics exports in Singapore.

2. Agents are "Leaving the Chat" From DeepSport (video) to Agentic Lab (AR glasses), the research trend is clear: Agents are now sensory and physical. If your AI roadmap is still text-only, you are missing the next wave of operational value.

3. Governance needs to catch "Agentic Traffic" The MCP Red Teaming paper is a wake-up call. As you adopt standards like MCP to connect your tools, remember that bad actors are adopting them too.