arXiv cs.LG

FarSky: Task-Aware Latent-Space Coupling for Generative Intra-Hour Solar Forecasting

Yann Fabel, Bijan Nouri, Milon Miah…

To address the limitations of existing hourly solar irradiance deep learning prediction methods, which mostly produce deterministic outputs and lack sufficient capability to predict ramp events, this paper proposes the FarSky generative prediction framework: it learns task-aware shared latent representations of sky images through a multi-task autoencoder, combines a conditional latent diffusion model to generate future latent states, and obtains probabilistic predictions via decoding. Validated on real-world measurement datasets, its prediction performance is up to 11 percentage points higher than baselines, with an F1 score of over 60% for ramp event detection, achieving the best overall performance.

Why AI Detection Fails for Academic Integrity

Jonathan A. Karr Jr, Grigorii Khvatskii, Ting Hua…

This paper addresses the issue that commercial AI detectors used by institutions to uphold academic integrity cannot distinguish between AI-polished manuscripts and fully AI-generated drafts, by conducting controlled experiments across multiple fields: covering 4 disciplines, comparing detection results of published English abstracts from the 2013-2015 and 2023-2025 periods. The study finds that the false positive rate for compliant light AI polishing reaches 64%-80%, while the false positive rate for purely human-written original abstracts in recent years is also 9%-15%, with far higher false positive rates in non-STEM fields than in STEM fields, confirming that AI detection scores cannot be used alone as a basis for determining academic misconduct.

Basin: Efficient and Extensible Numerical Optimization in Rust

Johan Larsson

This work introduces Basin, an efficient and extensible numerical optimization library for the Rust language. Numerical optimization is a general technique for finding the input that minimizes a function, widely used in scenarios such as model fitting, simulation calibration, machine learning training, and selection of optimal parameters for engineering cost minimization. Basin provides a unified and consistent interface for defining and solving optimization problems, has multiple built-in solvers, natively supports constraint handling, and can adapt to diverse optimization needs.

OpenAI

The builder’s guide to GPT‑5.6

OpenAI

This material titled The builder’s guide to GPT-5.6 targets the large model application implementation needs of startup teams, and summarizes two practical use cases for GPT-5.6: first, pairing it with intelligent model selection strategies, and second, calling its newly added Responses API capabilities, which can significantly shorten the development cycle of AI agents, reduce R&D costs, and help small and medium-sized teams efficiently build high-performance AI agents.

Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed

OpenAI

OpenAI recently launched the preview version of the new Ultrafast API tier, powered by the GPT-5.6 Sol model and supported by Cerebras computing power, with inference speed up to 14 times that of the regular version, outputting up to 750 tokens per second, which can significantly reduce the waiting time for large model generation and better adapt to scenarios such as high-real-time interaction and batch content generation.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly released Claude Opus 5 is a new iterative large model in the high-end Opus tier of the Claude series, with step-change breakthroughs in performance: it has been specially optimized for long-running agent scenarios, with greatly enhanced supporting capabilities, while its performance in code development and task processing in various professional scenarios has also been significantly improved, making it better suited for complex and high-demand production-grade use cases.

Inviting hard questions

Anthropic

This initiative launches a public call for AI-related questions, widely collecting the most confusing and difficult questions the public has about the AI field. The team also explicitly promises that when answering all collected questions later, it will fully disclose all work details of the answering process, including research paths and reasoning logic, so as to improve the transparency of AI research and respond to the public’s general concerns about AI technology.

Google DeepMind

Introducing Gemini 3.7 Flash

Google DeepMind

The newly released Gemini 3.7 Flash is Google’s lightweight flagship large model, optimized primarily through architecture slimming and inference path streamlining, with inference speed twice that of the previous generation, supporting a 128k long context window, with multimodal understanding and code generation capabilities close to the top-tier 3.7 Pro in the same series, and inference cost only 1/10 of the Pro version, making it suitable for multiple implementation scenarios such as edge deployment, real-time interaction, and batch content generation.

Putting sign language AI into users’ hands

Google DeepMind

This research focuses on the sign language interaction needs of deaf and hard-of-hearing users, and launches a breakthrough sign language to text (SL2T) AI model, with the core direction of promoting the implementation of sign language AI technology into consumer-grade scenarios that users can directly use, providing underlying technical support for various new sign language functions for the hearing-impaired group, effectively reducing the communication cost of the hearing-impaired group and improving the convenience of information interaction.

Hugging Face Blog

Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets

Hugging Face

This toolchain for robot AI development provides a one-stop full-process service, connecting the three components of Strands Agents, LeRobot, and Hugging Face Storage Buckets, enabling completion of the entire process of data recording, model training, and application deployment on a unified platform. It relies on storage buckets to achieve seamless circulation of data assets, and LeRobot adapts to the training needs of robot multimodal tasks, significantly reducing cross-platform collaboration costs and improving the implementation efficiency of robot AI.

What We Learned by Reproducing 2,200 papers from ICML

Hugging Face

After reproducing 2,200 papers accepted by ICML, this study found that only about 38% can be directly reproduced without modification, nearly 30% can be reproduced only after supplementing missing experimental information such as code, hyperparameters, and datasets, and about 12% cannot be reproduced even after contacting authors to request resources. The results highlight the shortcoming of reproducibility of papers from top machine learning conferences, and recommend that submissions be required to include complete experimental configurations and runnable code, and that a reproducibility assessment step be added to the review process.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment-related study rejects the traditional belief that “rational agents need to be bound to an ultimate goal”, pointing out that human rationality does not originate from goal orientation, but is the result of behavior adapting to a practical network composed of actions, evaluation standards, etc. The study proposes that to achieve adaptation of AI to human collaboration, the AI’s decision-making logic needs to be isomorphic to human practical logic, which can simultaneously meet the requirements of AI ethical alignment and core safety.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper sorts out the conceptual evolution of AI recursive self-improvement (RSI): in 1965, scholar I.J. Good first proposed the relevant idea, stating that “ultraintelligent machines” that can surpass all human intellectual activities can independently design better systems to complete iterations; in 2008, Eliezer Yudkowsky clarified that RSI specifically refers to the feedback loop in which AI relies on its existing capabilities to optimize its own cognitive architecture. RSI in the current AI context can refer both to a model directly rewriting its own weights, and broadly cover the behavior of a model optimizing its own training process.

QbitAI

In-depth experience of DeepSeek Harness: I forgive it for raising prices

QbitAI

DeepSeek recently open-sourced DeepSeek Harness (DSH), with all modules including model, tool, and context management adopting a pluggable design, supporting user customization to adapt to different scenarios. The official version has hundreds of built-in plugins, and it is known as the Android of the Agent era. The tool supports quick launch or source code deployment, and only supports Web UI access. Although the service price, especially cache fees, will be increased on the 17th, its high degree of freedom is still recognized by the author.

$480 million poured into edge computing! Rising star of Agent chips breaks through

QbitAI

Singaporean AI chip startup Acrab, which has been established for less than 3 years, has accumulated more than $480 million in financing after completing its Series B round. The funds will be used to expand production capacity, expand the technology ecosystem, and develop the next-generation AI computing platform. Its first-generation edge AI chip GΞLIX 1 and supporting Agent Box have started customer onboarding and entered mass production, betting on the new trend of large model inference shifting to the edge side.

Musk’s Grok 4.6 returns to the first tier! Outperforms Fable 5 at a lower price, the Cursor acquisition really paid off

QbitAI

The newly launched Grok 4.6 from SpaceXAI returns to the first tier of large models: its comprehensive intelligence is on par with GPT-5.6 Sol in 10 benchmark tests, only slightly inferior to Fable 5 Max, and it outperforms both in three real-world work capability tests. Its pricing of $2 per million input tokens and $6 per million output tokens is far lower than competing products. The model focuses on long-range Agent tasks, has been integrated into platforms such as Cursor, and can work with Grok Bot to achieve 24-hour automated operations.