Joo-Haeng Lee

Pebblous Inc CEO/Cofounder; Data Clinic; Data Magazine; Data Art Lab; Code Painting; Computer Graphics, CAD, Geometric Modeling, Robotics, HCI.

[Joo-Haeng Lee @joohaeng]

𝗜𝗦𝗢/𝗪𝗗 𝟮𝟲𝟮𝟲𝟰-𝟭 𝗛𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗥𝗼𝗯𝗼𝘁 𝗗𝗮𝘁𝗮𝘀𝗲𝘁𝘀 — About 13,000 humanoid robots shipped in 2025. By 2030 the figure is projected to pass 250,000. Yet the robots are multiplying far faster than their experience accumulates.

In June 2026, the first international draft standard for humanoid-robot datasets — ISO/WD 26264-1 — went public. It extends ISO 5259, the standard that made text and image data quality measurable, into the world of robot data.

Robot data is not a sensor log. It is a relationship between a robot's body, its action, the scene, the execution trace, and the outcome. When coordinate frames, calibration, and synchronization go unrecorded, the same motion becomes a different signal on a different robot. A 40-millisecond camera–IMU timing offset alone can push position estimates off by up to 10 meters.

So a million trajectories collected in mismatched formats are still a million isolated records. And because real-world data costs roughly 82 times more than simulation, the price of data you cannot reuse simply repeats.

The next bottleneck in Physical AI is not a bigger model or a better actuator. It is data that never accumulates because nothing standard holds it together.

▶ Read: https://t.co/ACGtAJUo8w

#Pebblous #DataClinic #DataQuality #DataJournalism #PhysicalAI #AIReadyData #HumanoidRobots #ISO26264 #ISO5259 #OpenXEmbodiment

[Joo-Haeng Lee @joohaeng]

On June 2, 2026, a U.S. executive order named "AI agents" for the first time. It did so inside a criminal-enforcement clause.

The order creates no new crime. It directs prosecutors to prioritize the existing Computer Fraud and Abuse Act against anyone who uses AI to break into a computer. The law treats AI as a tool and the person wielding it as liable.

Autonomous agents blur that line. When a human gives only a broad goal and the agent reaches systems it was never authorized to touch, who counts as the "individual using AI": the developer, the deployer, the operator? Liability does not vanish. It simply has to be reconstructed after the fact.

The raw material for that reconstruction is the log. Yet only 33% of organizations keep audit trails of courtroom-grade quality. A log does not make you innocent. But without one, there is no defense to begin with.

An action log is not a way to dodge regulation. It is the ticket that lets you stand in front of it.

▶ Read: https://t.co/40RspJdVEO

#Pebblous #DataClinic #DataQuality #AIGovernance #AIAgent #AgentEconomy #DataProvenance #CFAA #AIRegulation

[Joo-Haeng Lee @joohaeng]

AgiBotWorld는 로봇의 헛손질에 라벨을 붙인다. 실패한 시연을 버리지 않고 주석하는 데이터셋이 '클린 데이터=좋은 데이터'라는 통념을 뒤집는다 https://t.co/VUQGk7oBI4

[Joo-Haeng Lee @joohaeng]

𝗪𝗼𝗿𝗹𝗱 𝗠𝗼𝗱𝗲𝗹 — "How AI Understands the World and Predicts the Future"

A World Model is an attempt to represent the world not as fragmented pixels or tokens, but as an internal model of how objects move and interact. This is what allows a self-driving car to anticipate where a pedestrian will move next, a robot to mentally simulate the outcome before picking up an object, and a video generation model to produce physically plausible sequences. At the core of all these applications lies an internal model of the world. As a result, the concept of World Models connects seemingly distant fields like autonomous driving, robotics, and video generation.

Interestingly, this concept branches into two directions. One is about understanding the world. Approaches like Yann LeCun’s JEPA and DeepMind’s Dreamer focus on learning the principles of how the world works in abstract representation spaces, rather than reconstructing every pixel. The other is about predicting and generating the future. Models like Sora and Genie aim to simulate the world directly by generating the next plausible scene. They share the same name but have different goals and methods, and distinguishing the two makes the current AI research landscape much clearer.

This hub collects five Pebblous articles on World Models, ranging from an introductory five-step guide to a comprehensive survey, deep dives into JEPA, comparisons of the three main approaches, and insights into the limits faced by VLMs and VLAs. The articles are organized to naturally guide readers from beginner concepts to advanced understanding.

Explore Pebblous’s World Model series:

https://t.co/CQ30CnbNRj

#pebblous #WorldModel #DataGreenhouse #PebbloSim #PebbloScope #Blog

[Joo-Haeng Lee @joohaeng]

𝗪𝗼𝗿𝗹𝗱 𝗠𝗼𝗱𝗲𝗹 — "AI가 세계를 이해하고 미래를 예측하는 법"

월드 모델(World Model)은 AI가 세상을 단편적인 픽셀이나 토큰의 나열이 아니라, 사물이 어떻게 움직이고 서로 영향을 주는지에 대한 내부 모형으로 표상하려는 시도입니다.

자율주행차가 다음 순간 보행자가 어디로 움직일지 가늠하고, 로봇이 물건을 집기 전 결과를 머릿속으로 시뮬레이션하며, 영상 생성 모델이 물리적으로 그럴듯한 장면을 이어 그리는 일 — 그 바탕에는 모두 세계에 대한 내부 모형이 깔려 있습니다. 그래서 월드 모델은 자율주행·로보틱스·영상생성이라는 서로 멀어 보이는 분야를 하나로 꿰는 개념이 되었습니다.

흥미롭게도 이 개념은 두 갈래로 갈립니다. 하나는 세계를 이해하는 길입니다. Yann LeCun의 JEPA, DeepMind의 Dreamer처럼 픽셀을 일일이 복원하기보다 추상적인 표현 공간에서 세계의 작동 원리를 학습하려는 접근입니다. 다른 하나는 미래를 예측하고 생성하는 길입니다. Sora나 Genie처럼 다음에 펼쳐질 장면 자체를 직접 만들어 내며 세계를 시뮬레이션하는 접근입니다. 같은 이름을 쓰지만 목표와 방법이 다른 두 흐름을 구분해서 읽으면, 지금 AI 연구의 지형이 한결 또렷하게 보입니다.

이 허브는 페블러스가 월드 모델을 다룬 글 다섯 편을 한곳에 모았습니다. 개념을 처음 접하는 독자를 위한 다섯 단계 입문부터, 전체 지형을 조망하는 총정리 서베이, JEPA 기술 심화, 세 갈래 접근의 비교, 그리고 VLM·VLA가 부딪힌 한계까지 — 입문에서 심화로 자연스럽게 이어지도록 배치했습니다.

페블러스의 월드 모델 글 모음:
https://t.co/NL2WDdbDwm

#pebblous #WorldModel #DataGreenhouse #PebbloSim #PebbloScope #Blog

[Joo-Haeng Lee @joohaeng]

The technology to generate synthetic data improves every year. The technology to evaluate it remains primitive.

67% of enterprises already use synthetic data, yet no objective standard exists for proving its quality. Without a way to tell who produced good data, rewards can't be fair, and the best producers quietly leave the market. The same information asymmetry shows up in the 170x gap between the $319B global data broker market and the $1.8B pure-play data marketplace: buyers simply cannot verify quality before they buy.

Pebblous registered patent No. 10-2969403 computes quality scores along three axes — Fidelity, Utility, Privacy — and derives each producer's contribution directly from those scores, all the way to reward distribution. The structural twist is routing around Shapley value, whose coalitions explode past a million once participants pass twenty.

The limits are real. Standards define what to measure; how to measure it automatically and tie it to compensation is still a blank. When the EU AI Act mandates training-data quality proof for high-risk AI in August 2026, that blank stops being a cost and becomes a barrier to entry.

Whoever holds the proof technology takes the market first.

▶ Read: https://t.co/VOcn4KFvpp

#Pebblous #DataClinic #DataQuality #DataJournalism #SyntheticData #AIReadyData #DataGreenhouse #EUAIAct

[Joo-Haeng Lee @joohaeng]

Self-driving cars, robots, and Sora. Three fields that look unrelated turn out to lean on a single idea. A world model — AI's attempt to represent the world not as a flat stream of pixels but as an internal model of how things move and influence one another.

A car anticipates where a pedestrian will step next. A robot simulates the outcome before it grasps an object. A video model paints physically plausible scenes one after another. Each rests on an internal model of the world underneath.

What is striking is that the same name splits into two paths. One is understanding the world: approaches like Yann LeCun's JEPA and DeepMind's Dreamer learn how the world works in an abstract representation space instead of reconstructing every pixel. The other is generating the future: approaches like Sora and Genie simulate the world by producing the scenes that come next.

Reading these two currents as distinct brings the AI research landscape into sharper focus. Pebblous has opened a hub gathering five pieces — a five-level primer, a full survey, a JEPA deep dive, a comparison of three approaches, and the limits VLM and VLA ran into.

Seeing is not understanding. The next challenge is not a model that sees more, but one that understands the world.

▶ Read: https://t.co/CQ30CnbNRj

#Pebblous #DataClinic #DataQuality #DataJournalism #PhysicalAI #JEPA #Sora #WorldModel

[Joo-Haeng Lee @joohaeng]

Roughly 30% of generative AI projects stall on data quality, not the model. The thing worth fixing first sits one step earlier than the architecture.

The problem runs deeper into the data. About 55% of the causes behind LLM hallucinations trace back to data, and the largest share is bias in training data — a failure of representativeness. The reverse holds too: prepare data into a retrievable form and bolt it on, and the hallucination rate falls from 50% to 13.9%.

AI-Ready Data is not perfect data but data prepared and tracked for a purpose. Its conditions are seven quality dimensions — accuracy, completeness, consistency, timeliness, plus representativeness, validity, uniqueness — lineage that follows source and transformation, and governance that fixes access and accountability.

These conditions are moving from advice to regulation. From 2 August 2026, the EU AI Act requires documentation of where training data came from and how it was processed, and the data lineage market is growing 23.1% a year.

Once data's value moves from holding to tracking, data quality has to move with it: from a one-time inspection to an operation that keeps measuring the state.

▶ Read: https://t.co/DQ6kV0fBNj

#Pebblous #DataClinic #DataQuality #DataJournalism #AIReadyData #DataGovernance #RAG #EUAIAct

[Joo-Haeng Lee @joohaeng]

The world's largest ERP vendor just committed more than €1B (about $1.16B) over four years to an 18-month-old German startup. What Prior Labs, acquired by SAP on May 4, builds is neither a chatbot nor an image model. It is a foundation model for tables.

TabPFN is pretrained once on synthetic data, then reads a real table whole at inference time and predicts immediately, with no retraining — reading the table like a prompt. In 2025 Nature reported that a single 2.8-second inference surpassed the accuracy of an ensemble tuned for four hours. On the same day, SAP also bought the data lakehouse Dremio, securing the model layer and the data layer at once.

That performance carries a qualifier: small-to-mid tables, roughly under 10,000 rows. On large data, a TFM may accept up to a 40,000× latency penalty to gain 0.8% in accuracy. It does not replace XGBoost everywhere; it is fast and strong on small tables.

But the moment a model skips retraining, flaws in the input table are no longer diluted — they flow straight into the prediction. Missing values, schema drift, non-standard code values, and label noise all become degraded performance. Only 44% of manufacturing ERP data is AI-ready, and bad data costs the average company an estimated $12.9M a year.

The stronger the model gets, the more data quality matters, not less. That is the real question this deal leaves behind.

▶ Read: https://t.co/4Am8KStTyG

#Pebblous #DataClinic #DataQuality #DataJournalism #AIReadyData #DataGovernance #SAP #TabPFN

[Joo-Haeng Lee @joohaeng]

SAP가 창업 18개월 된 Prior Labs에 4년간 €10억 투자를 약속했다. 챗봇이 아니라 '표를 읽는' 파운데이션 모델 TabPFN이다.

모델은 재학습을 건너뛴다. 그래서 입력 표의 결함이 예측에 그대로 흐른다. 모델이 강해질수록 데이터 품질은 더 중요해진다.

https://t.co/dYiHHe10Tf

#페블러스 #데이터품질 #SAP #TabPFN

[Joo-Haeng Lee @joohaeng]

Coralogix just raised a $200M Series F. Not for collecting logs, but for watching what AI agents actually do.

Agents fail without errors: response times normal, error rate zero, decision wrong. Someone has to watch them.

Autonomy without observability isn't delegation. It's neglect.

https://t.co/yaPj9vmrph

#Pebblous #AIAgent #Observability #Coralogix

[Joo-Haeng Lee @joohaeng]

The Washington Post struck "training" from its OpenAI deal. Summaries, citations, and links are allowed; training on the content is not.

How AI buys data has shifted from buying once to renting continuously. Data isn't a held asset. It pays only while it flows.

https://t.co/BzBV7TcsuY

#Pebblous #DataQuality #OpenAI #DataLicensing

[Joo-Haeng Lee @joohaeng]

Most "open-source robot models" release the weights and keep the training data closed.

Ai2's MolmoAct 2 opened all of it: 720 hours of bimanual data, the code, the evals. Then it beat closed models at 87.1% real-world success.

The moat in robot AI is moving from the model to the data.

https://t.co/wvHtHI7mAU

#Pebblous #MolmoAct2 #PhysicalAI #DataSovereignty

[Joo-Haeng Lee @joohaeng]

Bezos's Prometheus raised $12B at a $41B valuation. The moat it pitched isn't the model. It's physical experiment data that OpenAI and Google can't scrape.

As models commoditize, value comes down to the data nobody else can reach.

https://t.co/fzx18j4gFX

#PhysicalAI #DataMoat #Prometheus

[Joo-Haeng Lee @joohaeng]

Malaysia plans to bring an AI governance bill to cabinet in June 2026. It's ASEAN's first bid to protect both training-data inputs and AI outputs as intellectual property.

While the West asks what to ban, Malaysia asks who owns the data first. Regulation read not as prohibition, but as turning data into an asset.

https://t.co/RR5XXEvsEP

#Pebblous #AIGovernance #Malaysia #ASEAN #DataSovereignty

[Joo-Haeng Lee @joohaeng]

A 269-page federal AI bill would freeze state regulation of model "development" for three years. Its named first target: California's AB 2013, the law requiring disclosure of training data.

The regulatory front is moving from a model's behavior to its training data, and from disclosure to audit.

https://t.co/a9JKghDzCE

#Pebblous #DataClinic #AIGovernance #TrainingData #GreatAmericanAIAct

[Joo-Haeng Lee @joohaeng]

미국 연방 AI 법안(Great American AI Act)이 주의 AI '개발' 규제를 3년간 막으면서, 첫 표적으로 캘리포니아 AB 2013을 지목했습니다. 학습 데이터를 공개하라던 법이죠.

규제의 전선이 모델의 행동에서 학습 데이터로, 공개에서 감사로 옮겨가고 있습니다.

https://t.co/AJTZd3hPSc

#페블러스 #데이터클리닉 #AI거버넌스 #학습데이터 #GreatAmericanAIAct

[Joo-Haeng Lee @joohaeng]

AI가 과학자 개인은 강하게, 과학 전체는 좁게 만들었다
Nature 게재 4,130만 편 분석… 23년간 AI 도구 도입 후 주제 다양성 4.6%·탐구행동 22% 감소 https://t.co/Oi7GopkslL

[Joo-Haeng Lee @joohaeng]

스킬 문서가 학습하기 시작했다 — 행동 운영체제의 시대

Microsoft SkillOpt 논문(arXiv:2605.23904)이 보여준 자기진화 알고리즘과 페블러스가 본 다음 카테고리 https://t.co/Ax7XKCFmqT

[Joo-Haeng Lee @joohaeng]

I just published "Should We Score Spatial AI? Five Criteria from PebbloSim’s Perspective"
https://t.co/knMipwXFPU