[DigitalToday reporter Hwang Chi-gyu (황치규)] Government AI policy is likely to be a key issue in this year’s state audit, running from Oct. 6 to Oct. 30. In particular, the outcome and future direction of the independent AI foundation model project, dubbed Dokpamo, have drawn attention. Details of a frontier AI development programme that appears set to proceed separately from Dokpamo have also emerged as a major focus.

The government’s position is to proceed with Dokpamo as scheduled and to disclose a specific plan for the frontier AI development programme later. But with Dokpamo not yet completed, the government has put frontier AI development on the agenda. That makes the division of roles between Dokpamo and frontier AI, and the sustainability of the Dokpamo project, likely issues.

ㆍAs rumours spread that Dokpamo could be halted, the science ministry’s minister and vice ministers made consecutive public appearances... Two-track push with frontier AI ㆍWill investment-style R&D programmes expand... Government reviews a public-private SPC for AI model development ㆍDeputy Prime Minister Bae Kyung-hoon: "Dokpamo will continue... We will also take on frontier AI" ㆍMinistry of Science and ICT: "Dokpamo will proceed as planned"... "Need to focus resources on frontier AI"

Unlike large language models (LLMs) that generate text, the competitive landscape for decision AI models focused on probability-based judgments has been rapidly reshaped into a multi-party contest.

ㆍAs 'Jev' rises, the judgment and decision AI model space quickly becomes a battleground... many major companies join ㆍAWS also launches open-source decision model 'Strands Decider 2B' ㆍOpenAI unveils 'Decisions API' to support faster judgments... targeting software automation

TypesafeAI, which opened the decision model market by unveiling Jev in mid-September, said it expected copycat products to appear. Still, it expressed confidence that Jev’s performance would be superior.

ㆍTypesafeAI CEO: "25 percent of Fortune 500 companies use 'Jev'"... confident about competing with copycat models

The article also summarised moves by companies at home and abroad around AI.

OpenAI held its annual developer conference, DevDay 2026, in late September. It unveiled more than 20 new products and features across ChatGPT, Codex, models and APIs.

ㆍChatGPT evolves into an agent platform... OpenAI DevDay 2026 roundup ㆍOpenAI unveils GPT-6.1 Sol... Astra-level performance at one-fifth the price ㆍOpenAI unveils 'Dots' aimed at Meta Muse... it moves on its own before being told ㆍOpenAI adds collaboration space, document and slide features to ChatGPT ㆍOpenAI adds $500-a-month plan 'Pro 500'... halves existing Pro limits

Google unveiled a new AI model, Gemini 4 Argon, designed to improve performance for specialised tasks such as coding, cybersecurity, finance and law. Google plans to roll out Argon in phases, starting with cybersecurity partners. Meta expanded the user base of its AI agent Muse from individuals to small businesses. Meta unveiled "Muse for Small Business," which links Muse with Shopify, Dropbox and Slack, among others.

ㆍGoogle releases new AI model 'Gemini 4 Argon'... phased rollout starting with security partners ㆍMeta expands AI agent 'Muse' beyond individual users to small businesses

Microsoft unveiled three voice AI models, including one for real-time transcription and two for speech synthesis. Mustafa Suleyman (무스타파 술레이만), CEO of Microsoft’s AI division, said, "By handling dictation and voice generation together, we can achieve high quality and speed." He added, "Going forward, new MAI models will replace existing voice models used across the company."

ㆍMicrosoft unveils real-time transcription and speech synthesis models... supports development of voice AI agents

Nvidia released its open foundation model for tabular data prediction, Kumo Tabular, on the open-source AI model-sharing platform Hugging Face. IBM launched a self-hosted deployment version of its agentic software development platform, IBM Bob, so companies can use it in on-premises, private cloud, sovereign cloud and air-gapped environments they can control directly. Adobe announced integration with Google Gemini and also significantly expanded features within Claude.

ㆍNvidia unveils Kumo Tabular... momentum for expansion into open-source structured-data foundation models ㆍIBM releases self-hosted deployment version of agentic software development platform 'IBM Bob' ㆍAdobe to offer key creative tools in Gemini as well... expands features within Claude

Chinese AI company DeepSeek is developing programming tools for Huawei Ascend chips in cooperation with Huawei. Alibaba Group unveiled its full-stack agentic solution, 'Qwen Intelligence,' to help smartphone manufacturers implement agentic AI features on their devices. The OpenKlo Foundation, which manages the open-source-based AI agent tool OpenKlo, released its enterprise AI agent management platform, OpenKlo Enterprise (OCE), as open source.

ㆍDeepSeek unveils development tools for Huawei AI chips... targeting Nvidia CUDA ㆍAlibaba unveils 'Qwen Intelligence'... "supporting implementation of agentic AI on smartphones" ㆍOpenKlo Foundation open-sources enterprise agent platform 'OpenKlo Enterprise'

Six Samsung companies, including Samsung Electronics, Samsung C&T, Samsung SDS, Samsung SDI, Samsung Life Insurance and Samsung Fire & Marine Insurance, decided to invest a total of $1 billion in U.S. AI infrastructure company Helix Digital Infrastructure.

ㆍSix Samsung companies to invest $1 billion in U.S. AI infrastructure startup Helix

Douzone Bizon will supply its generative AI platform, ONE AI Private Edition (PE), to a Seoul city project to build an administrative support AI agent. LG CNS signed a co-investment partnership with General Atlantic and will begin investing in global enterprise technology companies in earnest. Upstage unveiled its self-developed small large language model (LLM), Solar Mini 4.

ㆍDouzone Bizon to provide 'ONE AI PE' for Seoul’s 'administrative support AI agent' project ㆍLG CNS to expand enterprise software business in earnest... invests in global companies ㆍUpstage unveils 35 billion-parameter sLLM 'Solar Mini 4'

Amazon Web Services enabled AI models, including Anthropic’s Claude Opus 5 and Claude Sonnet 5, to be used on Amazon Bedrock based on the AWS Asia Pacific (Seoul) region.

ㆍAWS expands access to Anthropic Claude in Seoul region

As AI advances so quickly, governments around the world are grappling with how to respond. Even when AI laws are enacted, their effective lifespan is short, keeping challenges both large and small unresolved. AI-related laws enacted two years ago already have limited effectiveness. Many bills introduced to regulate AI fail to reach a vote and drift.

ㆍFrightening pace of AI development... governments worldwide at a loss over regulatory response

Meta and OpenAI are stepping up efforts again to target the consumer AI market with personal AI assistants, but there are still concerns about limits to profitability. The dilemma of operating costs that exceed demand has emerged as a key issue in consumer AI businesses.

ㆍPaid conversions are slow and operating costs are high... consumer AI revenue models put to the test ㆍ[Tech Inside] Can it sustain its early surge... the cold reality of Meta's 'Muse'