Nobel laureate and AlphaFold pioneer John Jumper has announced his departure from Google DeepMind after nearly nine years, joining rival AI lab Anthropic. This marks the second top AI talent lost by Google within 48 hours, following Noam Shazeer—core author of the Transformer paper—joining OpenAI. According to Bloomberg, Jumper's recent work was deeply tied to AI programming, but DeepMind lacks a clear strategy on enterprise-grade AI coding tools. Anthropic recently launched Claude for Life Sciences and acquired AI biotech startup Coefficient Bio for approximately $400 million, with its growing focus on life sciences being a key draw for Jumper. Anthropic maintains an employee retention rate of 80%, the highest among frontier labs, positioning itself as a magnet for elite AI talent.

Zhipu’s GLM-5.2 has topped the Design Arena single-turn HTML web design benchmark, defeating closed-source models including Claude Fable 5, becoming the first open source champion model under the MIT license. Despite having only 744B parameters and no vision capability, it outperformed proprietary rivals up to 6.7 times larger in scale, at just 1/7 to 1/11 of their cost. Its success lies in highly usable code: 91% of sessions used TailwindCSS, 51% used font-awesome, and it correctly invoked libraries such as chart.js and three.js. The model generated 25% more code than competitors, with an average generation time of 304.7 seconds (about twice that of Fable 5), trading speed for higher page fidelity. However, it still ranks second in game development, data visualization, and 3D design leaderboards, trailing behind Fable 5.

NVIDIA, CMU, and UC Berkeley have jointly introduced the ENPIRE framework, enabling multiple Coding Agents to autonomously complete an entire robotics research cycle—from reading papers and modifying algorithms to real-world testing—each controlling a dual-arm robot. In the Pin Insertion task, agents first attempted behavioral cloning, then introduced online reinforcement learning while tuning regularization and batch size, boosting success rate from near zero to 99% within three hours. The framework features auto-reset, auto-scoring, and safety control interfaces, making the physical world as iteratively accessible as a software development environment. With 8 agents operating across 8 robots in parallel, the time to achieve target success rates was reduced from 1.5 hours to 40 minutes, exhibiting human-lab-like 「experience inheritance」 phenomena.

Norway has announced an almost total ban on the use of artificial intelligence in primary education, aiming to protect children's cognitive development and reflecting concerns that over-reliance on AI may undermine critical thinking skills. Meanwhile, despite the U.S. government prohibiting Anthropic from releasing its Fable 5 model, related data appears to be spreading unchecked, revealing a disconnect between regulatory measures and actual technological diffusion. These events collectively highlight the intensifying tension between rapid AI advancement and societal norms, educational policies, and regulation. Regulators are attempting to set boundaries, but technology often spreads faster than policy can anticipate.

Open-source SDK startup Waniwani has secured an $8 million seed round. Its product enables financial service providers like insurers to quote and sell policies directly through AI platforms such as ChatGPT, positioning itself as an early player in AI distribution infrastructure. As consumers increasingly make decisions via AI, traditional distribution channels like SEO and digital advertising are becoming obsolete. Waniwani adopts a freemium model—offering an open-source SDK alongside paid infrastructure—similar to Stripe and Twilio. The founding team combines deep expertise in finance and AI: the CEO previously led BCG’s generative AI practice, and the co-founder founded France’s largest online insurance provider. A Bank of America report suggests AI-driven channels could threaten $15 billion in revenue for the insurance brokerage industry; on February 9, relevant stock prices dropped sharply, reflecting market repricing.

Anthropic has launched Claude Code Artifacts, productizing AI agent workflows into dynamic, team-shared, updatable, and reviewable workspaces. It captures terminal progress into real-time, updatable visual pages rather than static reports, turning agent outputs into direct objects of team collaboration. Pages are generated based on codebases, connectors, and conversational context, including reasoning processes, evidence sources, and current status—like a live worksite. The core value is reducing communication overhead in understanding agent outputs, applicable in incident investigation, delivery reviews, and security audits. The feature is currently available in private beta for Claude Team and Enterprise organizations, with default privacy settings and permission designs geared toward enterprise governance.

Redis creator antirez has published a rebuttal to the widely circulated narrative in the U.S. AI community that Chinese models rely on API distillation, sparking technical debate. He presents an 「impossible triangle」: APIs do not provide full logits or chain-of-thought traces, mathematically reconstructing complex models from sparse outputs is infeasible, and there is no clear information pathway. Critics argue he defines 「distillation」 too narrowly, pointing to cases like Alpaca and Vicuna where collecting large volumes of instruction outputs via APIs for fine-tuning constitutes viable black-box distillation. Researcher Nathan Lambert notes the term 「distillation attack」 has been weaponized as a moral label. Reports indicate Chinese labs are young, lean, and pragmatic, possessing world-class capabilities in pretraining, RL, and post-training, with progress driven by intensive engineering optimization rather than distillation.

Peking University has proposed LIFE-HARNESS, a method that enhances deterministic LLM Agent performance through runtime intervention without updating model weights, achieving an average relative improvement of 88.5% across 116 out of 126 settings. The study identifies that agent failures often stem from interface mismatches between model and environment—not insufficient model capability—such as incorrect tool call formatting, unexecutable actions, or looping trajectories. The method employs a four-layer design to handle environment rule definition, task workflow reuse, action format correction, and trajectory recovery, mining reusable failure patterns from training trajectories. Harnesses derived from Qwen3-4B training data can transfer to 17 other models, indicating they learn stable environmental structures rather than model-specific behaviors, demonstrating cross-model generalization.

OpenAI has introduced the Handoff feature for Codex, allowing users to seamlessly migrate ongoing programming tasks between local machines and remote servers via natural language commands, including full Git state. This enables developers to continue coding across devices without manually syncing code states. It continues a series of recent Codex enhancements, further bridging local development and cloud execution workflows, reducing friction in multi-environment switching, and making AI-assisted programming more aligned with real-world cross-device development scenarios.

A team from Westlake University and The Chinese University of Hong Kong, Shenzhen, has introduced DrPO, applying He Kaiming’s drift model to preference post-training in single-step text-to-image models. The method constructs drift directions in feature space using only reward rankings, attracting high-score samples and repelling low-score ones, bypassing reliance on denoising trajectories. Since rewards are used only for ranking and not backpropagation, gradient computation occurs in the form of drift regression within the feature space, achieving a 3.51x speedup over DRaFT—which requires gradient backpropagation—under HPSv3 rewards. The method also supports non-differentiable, rule-based, or procedural scoring signals such as GenEval, broadening its applicability. Ablation studies show latent-MAE features outperform those from pretrained models.