AI has turned intelligence into a commodity

What was discussed

The first session had Choi Yong-ho of AWS walk through how agents are actually used inside Amazon. The delivery example stood out. Vague addresses were causing a lot of failed first deliveries, so they had an agent pull together internal databases, public address records, geographic data, and even business hours to make the call. First-attempt delivery failures dropped 74%.

What creates scale, he said, is not the individual agent but the platform. Agent Z, Amazon's internal repository, now holds more than 23,000 reusable agents, and you don't need to be a developer to build one: plain language is enough. The way software gets built has changed too: they write a spec document first, have people review and revise it, and only then start development. That process has produced over 15,000 specs. The Bedrock framework, originally scoped at 30 developers and 18 months, shipped with 6 people in 76 days.

He boiled it down to three lessons. Give agents a goal and let them act on it rather than issuing step-by-step commands. Have several agents collaborate rather than relying on one. And let them run long, hours or days, rather than in short repeated bursts. He was also candid that productivity barely moved for the first few weeks; the gains only showed up after teams rebuilt the way they worked.

In the second session, Akimasa Omori of Classmethod Japan surveyed adoption across Asia-Pacific. Malaysia and Thailand are still pre-adoption because of cost and security concerns, Japan is moving gradually, and Korea, in his view, is the fastest. In Japan the main use cases were development, security review, reverse-engineering legacy systems like COBOL, and company-wide rollout.

The problem he singled out was "work slop." AI churns out documents that look plausible but say nothing, and employees end up spending an average of two hours a day interpreting and fixing that output. He added that in societies like Korea and Japan, where much of the context goes unsaid, answers drift further off target unless you put that context into the prompt properly. His conclusion: pair AI not with artificial intelligence but with actual intelligence, human judgment.

The fireside chat put lawyer Lim Hyun-seo across from Park Ha-eon of AIM Intelligence. Lim said his workload has grown 2.5 times, and that having AI draft the long, courteous replies for him has cut the emotional toll considerably. He also noted that court rulings are not published in Korea, which limits what legal-domain models can learn from. Copyright in AI-generated work remains unsettled territory, and he shared a case where someone claimed to have drawn a piece by hand and got caught when the figure turned out to have six fingers.

Park, the CTO, talked about AI security. His team finds vulnerabilities through red teaming and blocks them with guardrails, and they take part in pre-release testing for Anthropic, Meta, and OpenAI. He described finding a NASA chatbot about the Parker Solar Probe that had no filters and was producing dangerous answers; the service was taken down. His warning was that attacks that once required a specialist are now within reach of ordinary users.

Asked who survives in the age of AI, the two split. Lim picked the diligent person who keeps clicking through whatever changes come. Park picked the verifier with deep domain knowledge: AI is nowhere near taking responsibility for its output, so the experts who can judge that output are the ones who stay. The roundtable landed in a similar place: generalists have the edge right now, but once the market saturates, the specialists who set direction matter again, and in the end it is understanding people that makes AI more effective.