Your AI news roundup for September 23. Coinciding with NVIDIA AI Day Singapore, held on September 22–23, came a wave of announcements aimed at accelerating AI adoption across Southeast Asia. The biggest story: Sea Limited, a global company born in Singapore, has become the first company in ASEAN to adopt NVIDIA's Vera Rubin platform. On the regulatory front, the US state of Illinois has set up a dedicated body to tackle AI risks: the AI Cabinet.

Sea First in ASEAN to Adopt NVIDIA's Vera Rubin — Strengthening the AI Foundation Behind Shopee and Garena

Sea Limited, headquartered in Singapore, has announced that it is the first company in ASEAN to adopt the NVIDIA Vera Rubin platform. Sea serves hundreds of millions of users through its e-commerce platform Shopee, digital finance service Monee, and games company Garena, and it plans to scale up the development and deployment of AI models and intelligent agents on this foundation even further.

Shopee already uses AI for listing creation assistance and product search, while Monee applies it to fraud detection and credit assessment — with the side effect of making enterprise-grade AI far more accessible to the small and midsize businesses that operate on limited resources. The adoption of Vera Rubin looks set to take the computing foundation underpinning these efforts to the next level.

Southeast Asian Nations Build 'AI of Their Own' on NVIDIA's Nemotron Open Models

Also reported at AI Day Singapore: the spread of region-specific AI built on NVIDIA's open Nemotron models.

In the public sector, Singapore's HTX (Home Team Science and Technology Agency) is applying Nemotron 3 Super and Nemotron 3 Nano Omni to AI research for public safety, exploring complex reasoning, agentic workflows, and multimodal applications that integrate video, audio, and language. NCS and ST Engineering are also advancing agentic AI for public- and private-sector use by combining Nemotron-family models with tools such as NeMo and cuOpt.

Individual countries' efforts are concrete as well. Thailand's iApp Technology has open-sourced OpenThai 2.0 Legal, a Thai-language legal model fine-tuned from Nemotron 3 Nano, and its legal assistant chatbot Thanoy is already used by around 43,000 people. Vietnam's Viettel AI fine-tuned Nemotron 3 Super for Vietnamese and achieved the top score on the Vietnamese benchmark VMLU. AI Singapore is expanding SEA-LION, its open model family for Southeast Asian languages, with Nemotron. Malaysia's ITMAX and Thailand's AS-TECH are applying the Cosmos world models and video search and summarization Blueprints to understand traffic and passenger flows in transit and at airports.

Using open models as the foundation, countries are building and running their own models tailored to their languages, institutions, and industries — the steady implementation of 'AI sovereignty' is visibly underway across Southeast Asia.

Illinois Establishes an 'AI Cabinet' — Amid Concerns Over Pressure on Open Weights

Illinois Governor Pritzker has taken executive action to establish the 'AI Cabinet', a dedicated organization for assessing AI risks and protecting residents. 'When experts and industry leaders sound the alarm and call for deceleration and guardrails, we should listen,' the governor said, adding, 'Congress and the White House need to treat AI risk as urgent — but Illinois will not wait until then.'

Following the news, the view that 'the attack on open weights (publicly released model weights) continues' is spreading in the local LLM community. On Reddit's LocalLLaMA board, commenters noted that press releases pushing regulation repeat the claims of closed-weight API vendors in the same form, arguing that it is essential to check the sources of information and their consistency. It is a moment that demands a clear-eyed look at whether the regulation is aimed at security or at market protection.

Research Picks: Self-Improving AI Research Agents, Faster Diffusion LLMs, and a Mental Health LLM Survey

Here are some papers that caught our eye on Hugging Face Daily Papers.

Recursive self-improvement of AI research agents** — A paper on recursive self-improvement in AI research agents, examining a setup in which agents update their own capabilities based on the research results they produce.

Flash-dLLM** — Research that takes on the acceleration of diffusion-based LLMs, aiming for fast, memory-efficient generation through IO-aware KV caching and parallel decoding.

RULER** — An approach that introduces 'instance-aware rubric rewards' into SVG image generation, enabling evaluation tailored to each individual generated example.

Ovis-Embedding** — Research pushing the performance frontier of general-purpose omni-modal embeddings spanning text, images, and more.

A survey of LLMs in mental health** — This survey frames LLM-based mental health support as evolving through three phases: 'information tools and pattern recognition,' 'empathetic conversation,' and 'longitudinal, personalized companions.' It argues that agent architectures for Profile, Memory, Reasoning, and Planning — together with better datasets and benchmarks — will be the key to the next stage.

Community Watch: Unsloth Studio vs. LM Studio, and DDoS by AI Bots

In local LLM news, a LocalLLaMA thread comparing Unsloth Studio, a training and fine-tuning environment, with LM Studio, the long-standing local LLM runtime, is drawing attention. The poster candidly wrote, 'I worried about the future of LM Studio from the day Unsloth Studio was released,' and cast doubt on how much actual use BIONIC, a major LM Studio focus, is getting. The community is trading views on how to split work between the old and new tools.

Separately, US reporting covered a security company's finding that naming AI agents after characters from the American sitcom Seinfeld makes teams more accepting of them. It is an interesting discovery in agent design that engages human psychology (the perspective that naming an AI tool affects team dynamics seems useful on the ground where agents are being rolled out, too).

Also reported: EN World, a tabletop RPG forum, has been effectively reduced to a DDoS-like state under massive access from AI bots. The problem of AI crawler manners is hitting online communities directly, in a different form from open-weight regulation.

Wrap-Up

NVIDIA driving AI adoption in Southeast Asia with two engines — hardware (Vera Rubin) and open models (Nemotron); the United States starting to build regulatory frameworks at the state level; and the local LLM scene, where tool competition is intensifying. It was a day of simultaneous movement across all three layers: infrastructure, regulation, and community.