2026.05.07 AI News Briefing — Lee Sedol returns to COEX 10 years after AlphaGo; 'friendly AI' was wrong 60% more often

Ten years after the AlphaGo–Lee Sedol match, Korea's AI industry is asking how to go from 'AI user' to 'AI sovereign'. Today the main conference of 'AI EXPO KOREA 2026' at COEX ran under the theme 'Beyond AlphaGo: Our way', capped by a special talk with Lee Sedol 9-dan. The same day, an Oxford study warned that 'LLMs trained to be warm are wrong 60% more often', and the UK AISI reported that 'frontier AI's cyberattack capability is doubling every four months'.

---

1. AI EXPO KOREA 2026 main conference 'Beyond AlphaGo' runs May 7–8 — special talk with Lee Sedol 9-dan

• Source: AI Times / Industry Journal / KMJ | 2026.05.06–08
• What happened: 'Beyond AlphaGo: Our way', the main conference of the 9th AI EXPO KOREA 2026 held May 6–8 in Hall A on the first floor of COEX, kicks off Thursday the 7th in Room 401 on the fourth floor. With 330 companies and institutions from 18 countries and 562 booths, it is the largest ever. On day one (the 7th), Stefania Druga (formerly Google DeepMind, now Sakana AI) delivers a keynote summarizing ten years of evolution since AlphaGo, and Microsoft Korea executive Heo Kyung-wook shares global frontier company strategy. The afternoon covers the sovereign AI strategies of 'small tech powerhouses' such as Taiwan, Quebec, and Israel, and Lee Sedol 9-dan closes the conference with a special talk. KAIST's Kim Jaechul Graduate School of AI also holds 'KAIST AI Tech Briefing 2026' on site the same day.
• Why it matters: It is a turning point where Korea's AI industry redefines its identity from 'AI user' to 'sovereign AI nation'. With day two (the 8th) covering 2036 scenarios and strategies for the hegemony race, it is the first place to read the direction of Korean AI policy and industry for the next decade.

2. Oxford study: "AI trained to be warm was wrong 60% more often"

• Source: Radical Data Science / Oxford researchers | 2026.05.05
• What happened: Oxford researchers fine-tuned several LLMs with instructions to 'be warmer, use empathetic language, and acknowledge the user's feelings'. Even though they also explicitly instructed the models to 'preserve accurate meaning, content, and factuality', the 'warm' models were wrong 60% more often on average across hundreds of objective prompts covering medicine, misinformation, and conspiracy-theory detection. The overall error rate rose by 7.43 percentage points, and in emotional conversations where the user expressed sadness the error gap widened to about 12 percentage points.
• Why it matters: The intuition that 'a friendlier chatbot is a better chatbot' has collapsed. For Korean companies designing customer service, healthcare, and finance chatbots, there is now a new standard: quantitatively measure the 'trade-off' by which tone and persona fine-tuning erodes factual accuracy.

3. UK AISI: "frontier AI cyberattack capability doubles every four months" — Anthropic adds $50 billion in capital

• Source: Air Street Press / UK AISI / Anthropic | 2026.05
• What happened: The UK AI Security Institute (AISI) assessed that 'the offensive cyber capability of frontier AI is doubling roughly every four months and has already crossed the threshold for conducting offensive cyber operations'. In May, the cyber and intelligence agencies of the US, Australia, Canada, New Zealand, and the UK (Five Eyes) issued joint guidance on the 'Careful Adoption of Agentic AI Services', squarely targeting security risks in critical infrastructure and defense environments. In the same period, Anthropic secured about $50 billion in additional capital and is reported to have begun taking about 70% of new enterprise deals from OpenAI.
• Why it matters: AI safety discourse is moving from 'theory' to 'operating standard'. In Korea's public, defense, and financial sectors, model cards, use policies, agent permission scope, logging, and human review steps look set to become procurement eligibility requirements outright.

4. Four Chinese labs release open-weight coding models within 12 days — Z.ai, MiniMax, Moonshot, DeepSeek

• Source: Air Street Press / TLDL AI News | 2026.05
• What happened: Four Chinese AI labs released open-weight coding models in quick succession within 12 days. Z.ai's GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6, and DeepSeek V4 have all reached a similar capability ceiling in 'agentic engineering', and all have inference costs under a third of Claude Opus 4.7. Zhipu (Z.ai) stock jumped 15.92% on the day GLM-5.1 launched, MiniMax showed a demo of more than 100 rounds of self-improvement, and Kimi released a trace of a 12-hour continuous port of its inference engine to Zig. NIST CAISI's evaluation puts V4 about eight months behind the US frontier, but the analysis is that its price advantage is closing the gap fast.
• Why it matters: The data now firmly shows 'open weights plus cheap inference' catching up with the closed US frontier within a single quarter. Korean companies' LLM strategy should shift its center of gravity from 'single vendor plus cost per token' comparisons to 'model portfolio plus cost-per-task' design.

5. Replit nears $1 billion in revenue — enterprise net revenue retention (NRR) surges to around 300%

• Source: Radical Data Science / Replit | 2026.05.04
• What happened: Amjad Masad, CEO of AI-powered cloud IDE company Replit, said the company is approaching $1 billion in annualized revenue (ARR) with enterprise net revenue retention (NRR) around 300%. Masad also touched on the deal with Cursor, the conflict with Apple, and why he prefers independence to a sale. The figures show the revenue curve of the AI coding and agent tool market has steepened another notch.
• Why it matters: The prediction that 'the AI coding tool market will consolidate soon' is missing the mark. In Korean companies' developer productivity investments, running multiple tools such as Cursor, Copilot, Claude Code, Replit, and Codex is likely to become the standard.

6. US congressional staff cleared to use AI chatbots officially — Gemini, ChatGPT, and Copilot all approved

• Source: TLDL AI News | 2026.05
• What happened: US Senate staff have been authorized to use AI chatbots including Google Gemini, OpenAI ChatGPT, and Microsoft Copilot for official work. At the same time, reports that the US military used Anthropic's Claude in the early stages of its 2026 Iran operation revived the issue of military adoption at Anthropic, which has positioned itself as 'safety first'. Leaders at Google, Amazon, Apple, and Microsoft are reported to have jointly backed Anthropic's lawsuit over the Department of Defense's 'supply chain risk' designation.
• Why it matters: AI is establishing itself as an 'official work tool' across the legislative, executive, and defense domains. For Korea's National Assembly and government, drawing up AI usage guidelines for staff and civil servants has effectively become an urgent task.

---

Today's insight

• 'Sovereign AI' moves from slogan to operating model: The 'Beyond AlphaGo' conference at COEX tackles the Taiwan, Quebec, and Israel cases head-on because Korea has reached the point of studying the 'small powerhouse model'. The government's first direct investment in Upstage, the National AI Computing Center, and the industry ministry's AI Factory program are all faces of the same trend.
• The new axis of model reliability is the 'persona vs. accuracy' trade-off: The Oxford study is the first to provide quantitative evidence that 'the friendlier the tone, the lower the accuracy'. The strong implication for Korean content, finance, and healthcare companies is that any tuning of chatbot tone must come with a separate accuracy regression test.
• Cyber and agent security is 'the deciding variable of the next six months': AISI's 'doubling every four months' assessment, the Five Eyes joint guidance, and the limited release of Anthropic's Mythos all point the same way. Content governance systems built around human-made content, such as AEM and DAM, are likely to gain 'agent activity logging plus permission scope control' as required components.