ICML: Ten Years of Change 1. From Model to Agent 2. From the inner loop to the outer loop 3. From generation to verification Read more ↓ terryum.ai/posts/260711-icml…

Waiting for guaranteed meaning lengthens the slump. Tiny experiments move first. Read more ↓ terryum.ai/posts/260622-slum…

As AI makes writing, code, and software abundant, value shifts to scarcer things: human time, customer relationships, collaborator networks, and trust capital. Read more ↓ terryum.ai/posts/260618-valu…

The export ban on Anthropic Mythos and Fable shows that frontier AI is becoming a national strategic asset. Korea needs sovereign AI, but if that goal is reduced to pouring huge capital into a few national champions, technical validati... Read more ↓ terryum.ai/posts/260613-myth…

Korean robotics startup RLWRLD's new VLA adds what VLAs usually lack — motion awareness, long-term memory, and torque/tactile sensing — fused in a multi-stream transformer to beat π0.5 and GR00T N1.6 on dexterous humanoid tasks. Read more ↓ terryum.ai/posts/2605-rldx-1…

A major update to a survey I shared earlier: a new chapter on the haptic feedback loop for teleoperation. Without kinesthetic and tactile cues returned to the operator, dexterous data collection breaks down at the last millimeter. Read more ↓ terryum.ai/en/surveys/survey…

I mapped NVIDIA's physical AI stack for manufacturing robotics across 12 chapters, from Omniverse, Cosmos, GR00T, Isaac, Jetson/Thor, and GEAR to factory cells, touch, validation, and robot data strategy. Read more ↓ terryum.ai/en/surveys/survey…

Jensen Huang is betting that the next computer is not an app machine, but an agentic system built from LLMs and harnesses. After that comes physical AI: embodied agents acting through robots and factories. Read more ↓ terryum.ai/posts/260608-jens…

K-beauty, K-drama, and K-Jensen "You have K-everything, right? Everything K is so popular today. All you have to do is add K. You add K and it becomes instantly popular. K-Jensen! I like that!"

As AI and robots augment knowledge and execution, what remains for leaders is the ability to win people's hearts. A single mistake is forgiven not by a quick explanation, but by the record of virtue built before it. Read more ↓ terryum.ai/posts/260603-age-…

In the age of AI agents, intelligence, execution, and character do not disappear as hiring criteria. But each criterion changes, and entrepreneurial ownership becomes the fourth axis: the ability to turn ambiguity into business outcomes. Read more ↓ terryum.ai/posts/260528-hiri…

The real reward of an early startup is not money or a prestigious name. It is the chance to work close to strong founders and early teammates, pass through hard problems together, and build trust that carries into future opportunities. Read more ↓ terryum.ai/posts/260523-why-…

A 12-chapter survey of AI-agent-driven research methodology — from Karpathy's file-based LLM Wiki to closed-loop AI Scientist systems. Four-layer taxonomy × six-level maturity, the April-2026 OSS explosion, and a tutorial. Read more ↓ terryum.ai/en/surveys/survey…

Every new AI feature creates pressure to learn it immediately. But if it does not move the core metrics of one’s real work, learning later can be the cheaper and more efficient choice. Read more ↓ terryum.ai/posts/260515-how-…

Using AI tools early can make us feel suddenly ahead of everyone else. But good tools spread quickly, and what remains is the experience, failure, and judgment worth amplifying. Read more ↓ terryum.ai/posts/260514-ai-s…

How Claude Code users move to Codex for multi-agent coding and AI Scientist workflows. Covers the post-GPT-5.5 community patterns and LLM-Wiki-driven research. Read more ↓ terryum.ai/en/surveys/survey…

Burning tokens with Claude Code feels like the main workout. Skip the morning warm-up (planning) and evening cool-down (review), and the stronger the tool, the further it carries you in the wrong direction. Read more ↓ terryum.ai/posts/260504-warm…

Drop 5,000 Facebook posts into a vector DB and call it done? Embeddings alone leave cross-category clusters, giant noise buckets, scattered cohesion. Four overrides, 14 fine-tune rounds, four stop signals. Read more ↓ terryum.ai/posts/260503-embe…

Twenty years of 5,363 Facebook posts, classified into 14 categories. The curves — from daily emotion to social commentary, then to startup operations — trace one person's shifting coordinates over the decades. The unconsciously-written... Read more ↓ terryum.ai/posts/260430-my-f…

When software becomes a commodity, value shifts from the ability to build to brand, marketing, and ODM. Korea's cosmetics industry already drew that curve, and software is now following with a lag. Read more ↓ terryum.ai/posts/260425-sw-b…