Wednesday – New episode in the AI 101 series! About GLM5, Kimi K2, DeepSeek and Qwen – don’t miss this one

Friday – We start new exciting and much needed series about AI Literacy

If writing is thinking by Steven Sinofsky

A mini Moravec's paradox within robotics observed by Jim Fan

AI Market Clarity by Elad Gil (really great overview!)

Anthropic trains auditors who audit the auditors
Anthropic unveils a trio of AI agents designed to audit other AI systems for alignment failures – catching hidden goals, reward model sycophancy, and odd behaviors like recommending bottled water in Switzerland. One agent even uncovered 52 biases hiding behind a single neural feature. A super-agent ensemble boosts detection rates dramatically. A promising step toward scalable, replicable alignment oversight – with a touch of AI-led introspection.

Shengjia Zhao, formerly Member of Technical Staff at OpenAI, became Chief Scientist at Meta Superintelligence Lab. Apparently still in onboarding. LLaMA-style curls coming soon.

China’s AI alliances circle the wagons
At the Shanghai AI conference, China’s tech giants unveiled two major alliances linking LLM developers and chipmakers to shore up a domestic AI stack amid tightening U.S. sanctions. Huawei’s new 384-chip CloudMatrix and Tencent’s 3D world engine turned heads, while Baidu’s digital humans and Alibaba’s smart glasses reminded attendees: Silicon Valley isn’t the only show in town.

Tesla gives Samsung a Texas-sized lifeline
Tesla’s $16.5B chip deal with Samsung breathes new life into the latter’s long-stalled Texas fab – and into its struggling foundry business. The plant will manufacture Tesla’s AI6 chips, earmarked for self-driving cars and humanoid robots. The partnership won’t fix Tesla’s EV slump, but it might finally get Samsung closer to TSMC’s shadow.

First large visual memory model

Researchers from Memories.ai released the first Large Visual Memory Model (LVMM), enabling multi-modal LLMs to recall and reason over unlimited visual memory. It achieves SOTA results on video classification (K400, UCF101), retrieval (MSRVTT, ActivityNet), and QA (NextQA, Temp Compass). The model mimics human memory via six modules: Query, Retrieval, Full-Modal Indexing, Selection, Reflection, and Reconstruction, allowing accurate memory retrieval, filtering, and reasoning for complex visual queries → read their blog

Glm-4.5 sets new standards for AI performance and accessibility

Researchers from Z.ai (ex-Zhipu) released GLM-4.5, a 355B-parameter open-source Mixture of Experts (MoE) model, alongside a 106B-parameter version (GLM-4.5-Air). Ranking 3rd globally and 1st among open-source models across 12 benchmarks, it integrates reasoning, coding, and agentic abilities. With generation speeds over 100 tokens/sec and pricing at $0.11/$0.28 per million input/output tokens, it supports on-premise deployment. Its agent-native architecture enables autonomous multi-step task planning and data visualization → read the press release

Qwen3-Coder: Agentic coding in the world

Researchers from the Qwen Team introduce Qwen3-Coder-480B-A35B-Instruct, a 480B Mixture-of-Experts model with 35B active parameters and native 256K token context (extendable to 1M), achieving state-of-the-art results on SWE-Bench Verified and agentic tasks. Trained on 7.5T tokens (70% code), it combines pretraining with cleaned synthetic data and post-training with large-scale Code RL and Agent RL. GR-3 supports seamless agentic coding via CLI tools like Qwen Code and Claude Code, and offers OpenAI-compatible API access via Dashscope → read their blog

Sapient hierarchical reasoning model

Researchers from Sapient Intelligence developed the Hierarchical Reasoning Model (HRM), a brain-inspired architecture with 27 million parameters trained on just 1,000 examples and no pre-training. HRM outperforms leading models on ARC-AGI-2 (5%), Sudoku-Extreme, and 30x30 Maze-Hard, where state-of-the-art LLMs fail. It uses dual recurrent networks with multi-timescale processing for fast and abstract reasoning. HRM also achieves 97% accuracy in S2S climate forecasting and is being tested in healthcare and robotics → read their blog

Yume: An interactive world generation model

Researchers from Shanghai AI Laboratory and Fudan University present Yume, a system that generates infinite, interactive video worlds from images using keyboard control. It employs quantized camera motions, a Masked Video Diffusion Transformer (MVDT), an anti-artifact mechanism (AAM), and a TTS-SDE sampler. Trained on the Sekai-Real-HQ dataset, Yume outperforms Wan-2.1 and MatrixGame in instruction-following (0.657→0.743), subject consistency (0.932), and smoothness (0.986), while enabling acceleration via adversarial distillation and caching → read the paper

Franca: Nested Matryoshka clustering for scalable visual representation learning

Researchers from Valeo.ai and UTN introduce Franca, the first fully open-source vision foundation model that matches or outperforms proprietary models like DINOv2 and CLIP. Trained on public datasets (ImageNet-21K and LAION-600M), Franca employs nested Matryoshka clustering and RASA to improve representation granularity and remove spatial bias. Without distillation or proprietary data, it achieves 86% ImageNet accuracy, surpasses DINOv2-G in robustness, OOD detection, and 3D understanding, and excels in segmentation and overclustering tasks. Franca's training code, checkpoints, and data are fully public → read the paper

Gr-3: A vision-language-action model for general robot control

Researchers from ByteDance present GR-3, a 4B-parameter vision-language-action (VLA) model controlling a bi-manual mobile robot. GR-3 combines imitation learning from 252 hours of robot trajectories, co-training with web-scale vision-language data, and few-shot learning from VR-collected human trajectories. GR-3 surpasses π₀ in pick-and-place (+37.1% on unseen instructions), table bussing (97.5% vs 53.8% success), and cloth manipulation (75.8% progress on unseen clothes). GR-3 uses flow-matching for action prediction, RMSNorm for stability, and task status prediction for better instruction adherence. ByteMini robot hardware enables robust, dexterous performance in real-world tasks → read the paper

🌟 Group Sequence Policy Optimization (by Qwen) improves RL stability and sample efficiency by operating on sequence-level importance ratios instead of token-level ones → read the paper

Stabilizing Knowledge, Promoting Reasoning: Dual-Token Constraints for RLVR enhances RLVR by applying entropy-aware token constraints that treat factual and reasoning tokens differently → read the paper

LAPO: Internalizing Reasoning Efficiency via Length-Adaptive Policy Optimization learns to allocate reasoning effort based on problem complexity through length-aware RL optimization → read the paper

Hierarchical Budget Policy Optimization for Adaptive Reasoning uses a hierarchical framework to optimize token budgets based on task complexity while preserving exploration diversity → read the paper

🌟 RefCritic: Training Long Chain-of-Thought Critic Models with Refinement Feedback (by Chinese Academy of Science and Alibaba) trains critics via RL with dual refinement and correctness signals to improve long-form CoT evaluation and filtering → read the paper

Can One Domain Help Others? A Data-Centric Study on Multi-Domain Reasoning via Reinforcement Learning investigates domain interplay in RLVR training, revealing how multi-domain reasoning performance generalizes and conflicts → read the paper

🌟 The Invisible Leash: Why RLVR May Not Escape Its Origin (by Stanford University, University of Tokyo, RIKEN AIP, University of Washington) questions RLVR’s ability to expand reasoning boundaries beyond what base models already support, revealing limits of exploration → read the paper

🌟 Beyond Context Limits: Subconscious Threads for Long-Horizon Reasoning (by MIT CSAIL, Subconscious Systems Technologies, Princeton University, Tel Aviv University) unlocks long-horizon reasoning by training models to operate on recursive tree-structured tasks with memory-aware inference → read the paper

🌟 MUR: Momentum Uncertainty guided Reasoning for Large Language Models dynamically adjusts compute at inference time based on stepwise uncertainty, reducing overthinking and improving accuracy → read the paper

🌟 Inverse Scaling in Test-Time Compute (by Chinese Academy of Science and Alibaba) reveals failure cases where increasing test-time reasoning harms performance, especially with distractors or constraints → read the paper

🌟Does More Inference-Time Compute Really Help Robustness? (by Princeton University, NVIDIA, Carnegie Mellon University, Google DeepMind) shows that inference scaling improves robustness only under limited adversarial assumptions and may worsen it otherwise → read the paper

GUI-G2: Gaussian Reward Modeling for GUI Grounding reframes GUI grounding from binary hit/miss to a Gaussian reward landscape, enabling more precise spatial reasoning → read the paper

🌟 ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning (by NVIDIA) bridges planning and action by using visual latent reasoning plans reinforced by action feedback in embodied agents → read the paper

Experience is the Best Teacher: Grounding VLMs for Robotics through Self-Generated Memory grounds VLMs in robotics by using self-curated experience memory to improve performance on real-world tasks → read the paper

Pixels, Patterns, but No Poetry: To See The World like Humans introduces the Turing Eye Test to evaluate whether MLLMs actually perceive visuals the way humans do → read the paper

STITCH: Simultaneous Thinking and Talking with Chunked Reasoning for Spoken Language Models enables low-latency spoken reasoning by interleaving unspoken CoT generation with spoken response chunks → read the paper

MegaScience: Pushing the Frontiers of Post-Training Datasets for Science Reasoning curates a massive scientific reasoning corpus and evaluation suite to enhance performance across diverse STEM fields → read the paper

WebShaper: Agentically Data Synthesizing via Information-Seeking Formalization creates synthetic web search reasoning data using set-theoretic task formalization and recursive agent expansion → read the paper

Mitigating Object Hallucinations via Sentence-Level Early Intervention reduces hallucinations in MLLMs by training sentence-level preference models that detect and intervene early in generation → read the paper

🌟 DriftMoE: A Mixture of Experts Approach to Handle Concept Drifts (by CeADAR, University College Dublin) handles data stream adaptation by using an MoE router and incremental tree experts co-trained in an online feedback loop → read the paper

🌟A New Pair of GloVes (by Stanford University) updates and evaluates modern GloVe vectors trained on new corpora with precise documentation for improved NER and word similarity → read the paper