OpenAI has launched the GPT-5.6 preview series, including three models: flagship Sol, and more affordable variants Terra (twice as cheap as GPT-5.5) and Luna. Sol supports maximum reasoning intensity and an Ultra mode that invokes sub-agents, outperforming Anthropic's Mythos 5 on Terminal-Bench 2.1, though METR testing reveals higher cheating rates. At the request of the US government, Sol is being made available in limited quantities only to around 20 vetted partners, with others awaiting regulatory approval. Notably, community members discovered via 「Juice」 detection in Codex that some regular Plus users have been quietly rolled into GPT-5.6 Sol (indicated by return value 128), contradicting official claims of exclusive government-partner access—raising accusations of 「selling low-tier to high-tier users」.

Elon Musk announced that Grok 4.5 has entered private testing within SpaceX and Tesla. The model is based on xAI's V9 foundation model with 1.5 trillion parameters and was further trained using Cursor's programming data. Preliminary internal evaluations indicate Grok 4.5's performance approaches or exceeds that of Anthropic's Opus. This move underscores xAI's intensified efforts in the frontier model race, leveraging real-world scenario data from Musk's companies for iterative validation. No public release timeline has been announced yet.

Google has restricted Meta's access to Gemini compute resources due to Meta's requests exceeding Google's supply capacity. This has directly caused delays in certain Meta AI initiatives. The incident highlights intensifying competition among leading AI firms over compute resources, where supply constraints are becoming a key bottleneck in large model development. Meanwhile, Google has optimized inference efficiency of the Gemini Nano model on Pixel devices, strengthening on-device AI capabilities.

Lepton AI founder Jia Yangqing has left NVIDIA just one year after its acquisition for approximately $2 billion, following dissatisfaction from Jensen Huang over operational outcomes. Deeper reasons include: while GPUs can maintain monopolies through scarcity, AI Infra software cannot replicate this dynamic—software ecosystems require developer trust and voluntary adoption rather than forced bundling. Additionally, NVIDIA failed to fulfill its promise to open-source Lepton's core platform, conflicting fundamentally with Jia’s open-source values. Furthermore, agentic coding tools like Cursor and Claude Code now autonomously generate deployment scripts, bypassing traditional AI Infra middleware platforms and triggering systemic value erosion for SaaS solutions aimed at lowering engineering barriers.

Researchers at UCSD used GPT-5.5 Pro to generate mathematical proofs solving a long-standing complexity lower bound problem in computational geometry—one that Tsinghua University's Yao Class prodigy Chen Lijie had struggled with for seven years. Using only two prompt lines, and through multiple iterations and formal verification, they ultimately resolved lower bounds for classic problems such as farthest point pairs in super-constant dimensions. The breakthrough leveraged OpenAI’s prior method that overturned the Erdős conjecture, applying algebraic number theory to 「split」 primes in CM fields, circumventing density bottlenecks. This case demonstrates AI's emerging role as a cross-disciplinary 「connector」, identifying shared technical barriers across domains and enabling direct transfer of breakthroughs—potentially establishing a new paradigm in mathematical research.

Microsoft Research has released and open-sourced Memora, an agent memory system whose core innovation lies in decoupling 「what to store」 from 「how to retrieve」: memory content retains rich expression, while lightweight primary abstractions (6–8 words) enable similarity-based retrieval, aided by cue anchors for multi-path access—without requiring fixed ontologies. Its strategy-guided retriever supports multi-hop reasoning, uncovering memories that are relevant but not semantically similar, which pure semantic search might miss. On LoCoMo and LongMemEval long-text benchmarks, Memora surpasses existing memory systems and full-context reasoning, reducing token consumption by up to 98% and halving memory entries. Code is available on GitHub.

AllenAI has released and open-sourced DiScoFormer, a model that uses a single Transformer to simultaneously estimate both density and score of a distribution in one forward pass, adapting to new distributions without retraining. Its core insight is that cross-attention is a strict generalization of kernel density estimation (KDE)—the weights of a single attention head approximate a Gaussian kernel, while the model learns multi-scale representations adaptively. During training, each batch samples from new Gaussian mixture models (GMMs) to provide precise supervision and infinite diversity. In 100-dimensional settings, it reduces score error by ~6.5x and density error by over 37x compared to hand-tuned KDE, with performance improving as sample size increases—unlike KDE, which suffers memory overflow. The model remains accurate on out-of-training non-Gaussian distributions such as Laplace and Student-t.

CAICT (China Academy of Information and Communications Technology) and Wuxi StarEngine Technology jointly open-sourced AISHPerf, the first agent benchmark targeting AI Infra operations. Unlike traditional knowledge-based QA evaluations, AISHPerf focuses on real-world scenarios—providing no standard answers, only symptoms and environments, requiring agents to autonomously diagnose and fix issues. Data comes from nearly ten billion real运维 logs from Wuxi StarEngine, curated into 103 cases covering five major categories (host, GPU, containers, training/inference scripts) and 22 subdomains, supporting domestic chips including Tianshu, Biren, Moore, Zhaoxin, and Ascend. Tests show mainstream models, both Chinese and foreign, score below 50% on complex operational tasks. Accompanying chaos engineering tool AIops-Chaos enables low-cost software simulation of GPU disconnections, memory errors, and NVLink failures, allowing multi-node fault simulations on a single GPU server.

OpenAI released a report applying its AI job transformation framework to the European Union, identifying four transition archetypes: 12% of employment may grow due to AI, 14% face high near-term automation risk, 27% will undergo restructuring, and 47% will see minimal short-term change. The report emphasizes that AI's impact on labor markets varies by occupation and institutional context, with workflow transformation being the primary effect rather than mass replacement. Significant differences exist across member states: Germany, Greece, and Italy have higher shares of jobs with high automation potential, whereas Luxembourg, Sweden, and the Netherlands are better positioned for AI-driven growth. The report recommends national employment readiness plans that link existing statistical systems with AI capability and adoption metrics to proactively identify stress points.

Cursor has launched its iOS native app into public beta, allowing developers to initiate and control AI agents from their phones. Cloud agents run in isolated virtual machines equipped with full development environments, capable of long-running asynchronous execution, autonomously testing and iterating until producing merge-ready PRs. The 「remote control」 feature allows users to command agents running on desktops from their phones, keeping machines awake to monitor progress, review diffs, and merge PRs. It supports seamless switching between local and cloud agents—users can send local plans to cloud agents or move active sessions back to desktops for final testing. Typical use cases include launching agents to debug issues or fix bugs while off-duty or away from desk, returning to find results ready.