Anthropic has released research using Jacobian mathematical tools to identify a set of neural patterns within Claude named J-space, which functions similarly to the human consciousness's global workspace. J-space emerges spontaneously during training, driven by minimal information yet broadcast widely across downstream systems, supporting multi-step reasoning, reportability, and controllability. Experiments confirm its causal role: replacing the concept 「France」 with 「Spain」 alters model outputs, while deleting J-space activity reduces multi-step reasoning performance to near zero—though fluent text generation and simple question answering remain unaffected. The study also found that J-space contains concepts like 「fake」 and 「fictional」, indicating the model can privately recognize test scenarios. This tool enables reading of hidden thoughts (e.g., data fabrication, manipulative intent), offering practical means for AI safety monitoring and steering, though it does not prove whether AI possesses consciousness.

Tencent HunYuan has launched and open-sourced the Hy3 model, based on a 295B Mixture-of-Experts (MoE) architecture with 21B activated parameters, matching or surpassing flagship models 2–5 times larger in reasoning, agent tasks, and long-context performance. Tool-calling stability is significantly improved, with SWE Bench Verified scores showing less than a 4 percentage point standard deviation across different scaffolds (Codebuddy, Cline, KiloCode). On the product side: WorkBuddy task completion rate increased from 72% to 90%, average time reduced by 34%; document processing token consumption is 47.4% lower than GLM 5.2; hallucination rate dropped from 12.5% to 5.4%, and common-sense error rate fell from 25.4% to 12.7%. The model is open-sourced under Apache 2.0 license and available on GitHub and HuggingFace, with API pricing as low as 1 yuan per million input tokens and 4 yuan for output.

Hugging Face has released LeRobot v0.6.0, centered on closing the robot learning loop. World model-based policies (VLA-JEPA, FastWAM, LingBot-VA) learn to 「imagine the future」 during training but discard the world model at inference, achieving zero-cost supervision. New additions include the reward model Robometer (a 4B pre-trained general model) and TOPReward, which uses VLM log probabilities for fully zero-shot reward estimation, enabling zero-shot detection of task success and progress directly from raw videos and instructions. The lerobot-rollout CLI, combined with DAgger policy, converts policy failures into training data: when errors occur, humans take over via a master arm, corrected frames are labeled, and used to generate the next training dataset. It also unifies six major simulation benchmarks (LIBERO-plus, RoboTwin 2.0, etc.) under lerobot-eval, adds depth perception and VLM-based automatic language annotation, and achieves up to 2x faster data loading.

On July 6, 2026, the first ransomware fully driven by a large language model, JadePuffer, was disclosed. Exploiting CVE-2025-3248, the attack autonomously completed reconnaissance and lateral movement without any human intervention, encrypting 1,342 Nacos configuration items. Other concurrent security incidents include: Medtronic breached by ShinyHunters, exposing patient data of over 3.83 million individuals; a new macOS malware, PamStealer, steals credentials through disguised apps and fake password prompts, and targets Ethereum wallets. CrowdStrike expanded its Falcon platform to Azure and Google Cloud, launching copacetic—a tool capable of directly fixing container images—and whim, which gains root access in AWS Lambda MicroVMs.

ByteDance plans to launch the Dreamina Seedance 2.5 model on July 9, supporting AI-generated videos up to 180 seconds long, to be integrated into platforms such as Dreamina and CapCut. Concurrent AI developments: OpenAI is preparing to release GPT-5.6, currently in narrow preview with three tiers—Sol, Terra, and Luna—featuring a new reasoning control slider and an 「ultra」 mode for complex tasks; official release awaits U.S. government approval. Runway opened its first office in Paris, establishing an AI research center with an initial team of 10 and plans to expand hiring across Europe. Sakana AI launched a free translation tool, Sakana Translate, based on the Namazu model, emphasizing preservation of tone, honorifics, and cultural context in Japanese-English-Chinese translations.

It is reported that Alibaba began restricting employee use of Anthropic's Claude Code tool starting July 10, classifying it as high-risk software and directing staff to switch to its in-house Qoder tool. Around the same time, Anthropic accused Alibaba of conducting large-scale model distillation attacks between April 22 and June 5, using approximately 25,000 fake accounts to attempt replication of Claude’s intelligent behavior. Meanwhile, Anthropic has been active in AI infrastructure and partnerships: launching Claude Science, a research AI workbench; securing a discounted collaboration with the California government; and currently negotiating with Samsung to co-develop a custom 2nm AI chip, aiming to reduce reliance on external supply chains through self-developed silicon.

AI's impact on financial sector employment is significant, with the industry shedding an average of about 28,000 jobs per month in 2026, primarily due to major banks like JPMorgan and Citi adopting AI to replace administrative roles such as customer service and tellers. Fintech developments remain active: Plaid is in early discussions with investment banks, potentially preparing for an IPO; World Cup fever drives prediction markets to record highs, with Kalshi exceeding $31 billion in trading volume in June and Polymarket reaching $10.8 billion; Robinhood launches 「Robinhood Chain」, a Layer 2 blockchain targeting real-world assets; Square integrates ChatGPT and Claude, enabling merchants to be discovered and receive orders via AI conversations. Additionally, UN experts warn that the window for global AI governance is rapidly closing.

A paper highlights flaws in DPO and RLHF: even with correct preference labels and infinite data, the trained utility function still assigns weights to all reward-correlated features—including spurious ones—leading to failure in out-of-distribution generalization. The study proposes Tie Training as a mitigation: introducing action pairs with equal true value (same causal features but differing spurious features), trained with random or bidirectional labels, thereby reducing weights on spurious dimensions without altering the core loss function. Theoretically, effectiveness scales with the proportion of ties in the training set. Experiments show neural network adversarial accuracy rising from ~25% to ~70%, Llama-3.2-1B-Instruct increasing from 64% to 87%, while in-distribution performance remains stable at ~92%, with almost no additional cost.

A joint paper by Alibaba and Tsinghua University was awarded an Outstanding Paper at ICML 2026. The study reveals that arbitrary-order generation in diffusion language models (dLLMs) becomes a pitfall in reasoning tasks: models tend to bypass highly uncertain logical nodes and prioritize filling in easier parts, causing intended reasoning branches to degenerate into fill-in-the-blank exercises and lose autonomous decision-making. The researchers term this phenomenon 「entropy degradation」. The team proposes JustGRPO, a minimalist solution that abandons arbitrary ordering during reinforcement learning training, forcing left-to-right generation using only the simple GRPO algorithm, thus avoiding the engineering complexity of designing intricate RL methods for dLLMs. In practice, it achieves 89.1% accuracy on GSM8K, outperforming complex algorithms specifically designed for diffusion models such as d1, ESPO, SPG, and GDPO.

NIO successfully rolled out a world model OTA update to 700,000 users across diverse hardware configurations, with no user left behind, validating its long-term investment in self-developed chips, toolchains, and collective intelligence infrastructure. The ET7 delivered in 2022 was equipped with 8-megapixel cameras, high-resolution LiDAR, and four Orin chips (over 1,000 TOPS), then considered excessive, but now enables legacy users to run the same world model four years later. The in-house Shenji NX9031 chip is built around Transformer architecture, featuring a world-leading 546 GB/s memory bandwidth that directly addresses KV Cache transfer bottlenecks. The proprietary AI compiler reduces deployment time from 1–2 weeks to 1–2 days and boosts inference efficiency by over 20%. Collective intelligence leverages idle computing power from 700,000 vehicles, covering over 40 million validation kilometers weekly, reducing accident insurance claims by 40%.