June 18, 2026
English translation of the Korean original, prepared with AI assistance. Korean original
Signals 34
A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry
A case of autonomous AI research in which OpenAI and Molecule.one used GPT-5.4 to improve difficult reactions in medicinal chemistry.
Signal — The commercialization of AI-driven, end-to-end scientific discovery cycles (AI-native R&D) is accelerating.
OpenAI Blog
Introducing LifeSciBench
LifeSciBench is a new benchmark, designed and reviewed by experts, for evaluating real life-science research tasks and decision-making ability.
Signal — It is increasingly clear that the yardstick for AI performance is shifting from the amount of knowledge to reasoning ability in specific expert domains.
OpenAI Blog
Predicting model behavior before release by simulating deployment
OpenAI has presented 'deployment simulation', a method that uses real-world conversation data to run predictive safety evaluations before an AI model is deployed.
Signal — In the competition over AI performance, predictable safety, beyond capability, will become a key barrier to entry.
OpenAI Blog
Unlocking UK house-building with AI-accelerated planning
Google DeepMind is working with the UK government to develop an AI-based planning prototype that speeds up housing construction.
Signal — AI is entering an era in which it moves beyond the digital stack and acts as a real-world decision-making system (AI agent) that manages complex permitting and physical processes.
Google DeepMind
DiffusionGemma: 4x faster text generation
Research results show that applying diffusion techniques extensively made text generation in Gemma models four times faster.
Signal — Lighter LLMs and highly efficient inference will become the standard trend in AI services, and attention will focus on innovation in model architecture itself.
Google DeepMind
Investing in multi-agent AI safety research
Google DeepMind has confirmed $10 million in funding for research on the safety of multi-agent systems.
Signal — An era is starting in which AI at the commercial deployment stage treats safety, not performance, as its top KPI.
Google DeepMind
MolmoMotion: Language-guided 3D motion forecasting
A technology that predicts and generates 3D motion from a text description.
Signal — The era of large multimodal models is arriving, in which language is used as the intermediary to direct demanding physical and visual outputs (3D, complex simulations).
HuggingFace Blog
From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot
Strands Agents and LeRobot connect HuggingFace's open model ecosystem to robot hardware, enabling autonomous physical action.
Signal — The ultimate direction of AI is shifting from information processing to the market for robot agents that carry out physical tasks autonomously.
HuggingFace Blog
GLM-5.2: Built for Long-Horizon Tasks
GLM-5.2 is a large language model built for long-context understanding and complex multi-step reasoning (long-horizon tasks).
Signal — LLM development will concentrate on building autonomous agents that keep long-term memory and pursue goals, not just retrieve knowledge.
HuggingFace Blog
Hands Free, AIs Forward: NVIDIA XR AI Brings Agents to AR Glasses
NVIDIA XR AI offers a public beta of a framework for developing multimodal AI agents for AR glasses and XR devices.
Signal — The deployment of AI agents is evolving beyond screen interfaces to a stage where they combine directly with real space (spatial computing).
NVIDIA Blog
Coherent Breaks Ground on Expanded Texas Facility, Scaling AI’s Optical Backbone
Coherent is building an expansion plant in Texas, securing capacity to produce the optical components and 6-inch indium phosphide (InP) wafers that AI system interconnects need.
Signal — In the AI accelerator and server market, the standardization and mass production of optical interconnects will be the key competitive strength of the next generation.
NVIDIA Blog
HPE AI Factory With NVIDIA Expands for the Era of Agents
HPE AI Factory is being combined with NVIDIA to build agent-based AI operating environments for enterprises.
Signal — The focus of AI solutions is shifting from model training to autonomous, complex task execution (agentic workflow).
NVIDIA Blog
RL Systems Mind the Gap: Matching Trainer and Generator Throughput
To relieve the bottleneck in reinforcement learning (RL) training, the paper presents an asynchronous pipeline infrastructure and optimization techniques that resolve the throughput mismatch between trainers and generators.
Signal — End-to-end AI pipeline optimization, which minimizes latency across the whole training-to-deployment process, will emerge as a key trend.
SemiAnalysis
Is SMIC N+3’s Metal Pitch Smaller than Intel 18A’s?
A technical validation report analyzes whether SMIC's next-generation N+3 process has an advantage over, or a point of comparison with, Intel's 18A process in metal pitch.
Signal — Metal pitch and transistor design at ultra-fine process nodes will be the most important bottleneck for next-generation chip performance.
SemiAnalysis
Intel Should Raise Capital
The piece argues that it is time for Intel to take advantage of current market conditions to raise capital and to restore its technological standing through large-scale investment.
Signal — The coming AI infrastructure war will be fought not only on pure technical performance but also on companies' financial strategy and ability to raise funds in capital markets.
SemiAnalysis
Import AI 461: “Alignment is not on track”; FrontierCode; and synthetic research interns
To address AI safety problems, researchers are launching new safety-focused startups and laying out synthetic researchers and a structured portfolio of research bets.
Signal — The key bottleneck in AI investment and development will be securing safety (alignment) and reliability, not peak performance itself.
Import AI
Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing
AI research that applies the concept of finding and optimizing flaws in a reward system itself (reward hacking) to social systems.
Signal — As AI matures, the focus is shifting from maximizing performance to detecting structural vulnerabilities in systems and studying how they can be exploited.
Import AI
Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems
A research newsletter covering topics such as AI's macroeconomic growth rate, scaling laws for protein folding models, and analysis of existential risks from AI systems.
Signal — The debate over AI progress is shifting its weight from a race on performance to building social safeguards and systematically proving scientific utility.
Import AI
After Orthogonality: Virtue-Ethical Agency and AI Alignment
The paper presents a philosophical theory that human rationality comes from alignment with 'practices', networks of practical activity, and not from orientation toward final goals.
Signal — The next stage of AI safety research will go beyond defining what to do, to modeling the complex human procedures and value judgments of how to act.
The Gradient
AGI Is Not Multimodal
A theoretical warning that the current success of multimodal AI does not amount to general human intelligence (AGI).
Signal — The next breakthrough in LLMs will be physical action and grounding in the real world, not knowledge reasoning.
The Gradient
Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
The piece discusses a paradigm shift in machine learning research and analyzes the tension between approaches based on mathematical principles and approaches based on scaling compute.
Signal — As returns on compute fall, an academic approach that designs the fundamental structure of the model itself (shape, symmetry) will matter again.
The Gradient
When Rules Learn: A Self-Evolving Agent for Legal Case Retrieval
A framework that uses an LLM-based agent to optimize existing sparse retrieval methods such as BM25 for legal case search, on its own and without parameter training.
Signal — Meta-optimization, in which LLMs autonomously improve proven classical AI and information retrieval systems rather than only generating content, will become a main research trend.
arXiv cs.AI
SkillChain-Gym: A Benchmark for Reskilling-Aware Production-Inventory Control under Disruptions
A simulation benchmark for production inventory management that includes maintaining workforce skills and retraining as essential variables.
Signal — The scope of AI is expanding beyond optimization inside the digital stack to designing intelligent systems that handle complex physical and social constraints in the real world, such as labor and resources.
arXiv cs.AI
Quantifying Consistency in LLM Logical Reasoning via Structural Uncertainty
The paper presents a method to quantify the reliability of a large language model's (LLM) logical reasoning by measuring the stability of the model's own pairwise preference rankings over multiple sampled candidate solutions (structural uncertainty).
Signal — The LLM evaluation paradigm will move away from simple answer accuracy toward consistency of the reasoning process and reliable measurement of uncertainty.
arXiv cs.AI
The Critical Role of Model Selection in Causal Inference: A Comparative Analysis of Classification Models within the InferBERT Framework for Pharmacovigilance
A comparative study showing that the choice among classification models is critical for causal inference in pharmacovigilance (drug safety analysis).
Signal — In AI-based medical research, the shift from a scale-centered approach to one centered on methodological rigor will accelerate.
arXiv cs.LG
Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis
A research paper that systematically evaluates how well foundation models represent cancer, using multiple modalities (whole-slide images and transcriptomic profiles).
Signal — For general-purpose AI models to enter clinical decision-support systems, a high degree of explainability and reliability checks based on multi-signal fusion will be the key gateway.
arXiv cs.LG
MODE: Modality-Decomposed Expert-Level Mixed-Precision Quantization for MoE Multimodal LLMs
To cut the memory cost of MoE-MLLMs, the paper presents an expert-level mixed-precision quantization method that accounts for modality bias (cross/intra-vision bias).
Signal — Efficient data processing and compression of complex multimodal inputs is the main bottleneck for the next wave of AI commercialization.
arXiv cs.LG
MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision
MemSlides is an agent framework built on hierarchical memory that separates long-term memory, working memory, user-profile memory and tool memory.
Signal — Going forward, every specialized AI workflow will require memory and revision as core functions, beyond simple generation.
arXiv cs.CL
RepSelect: Robust LLM Unlearning via Representation Selectivity
RepSelect is an unlearning method that selectively collapses the principal components of weight gradients, so that an LLM deeply forgets specific knowledge or values while keeping its general abilities.
Signal — Implementing the 'right to be forgotten' for AI models in technical terms will become a key research topic in next-generation LLM development.
arXiv cs.CL
From Parasocial Scripts to Dyadic Persistence in Autonomous AI-Agent Communities
An empirical analysis finds human-like parasocial relationships (PSI) and interaction patterns in online communities run by autonomous AI agents.
Signal — A new approach that models the social and psychological rules governing AI agents' behavior will be the next trend.
arXiv cs.CL
Roelof Botha joins SpaceX’s board of directors
Roelof Botha, a Silicon Valley VC investor, has joined the board of SpaceX.
Signal — Watch for changes in the governance and capital-raising structures of large infrastructure companies.
TechCrunch AI
After unveiling ridiculously expensive AR glasses, Snap’s stock takes a dive
Snap's shares fell after the smart glasses it launched failed to meet market expectations.
Signal — This shows that success depends not just on launching hardware but on combining user experience (UX) with a platform that blends deeply into daily life.
TechCrunch AI
NEA’s Tiffany Luck says enterprises are still figuring out their AI ROI
After an initial phase of excessive AI use (token consumption), companies are starting to doubt the actual return on investment (ROI) and to manage costs.
Signal — AI investment is entering a stage where it must prove real business results (profits) and not just hype.
TechCrunch AI
World leaders want American AI. They just don’t want America to be able to turn it off.
Leaders of major countries see the risk of the US abruptly cutting off access to AI technology as a serious security problem.
Signal — The regionalization of AI in pursuit of technological sovereignty, and the building of open-source-based ecosystems that reduce reliance on the US, will accelerate.
TechCrunch AI