Artificial intelligence is transforming retail, helping to optimize store layouts, manage assortment, and forecast demand. Major chains are implementing AI solutions to improve efficiency and personalize shopping.
Магнит
Anthropic's Claude Opus 5 earned $75,264 in the FoodTruck Bench simulator, outperforming the previous record held by GPT-5.5 by 13.6%. The key success factor was not sales volume but a higher average portion price ($12.69 vs $11.86) and a well-thought-out strategy for managing locations, prices, and inventory.
Anthropic
Chats created using Claude AI's publish feature have been indexed by Google's search engine, as discovered by Reddit users. Anthropic responded quickly—the request stopped working, but many links remain active and contain personal data. Private (unpublished) chats were not affected.
Anthropic
Researchers from BAIR (Berkeley AI) introduced two new methods for protecting large language models against prompt injection attacks: StruQ and SecAlign. Both approaches require no additional computational overhead or manual effort and effectively reduce attack success rates while preserving model utility. StruQ is a structured instruction setting, and SecAlign is a special preference optimization that achieves an even higher level of protection.
Meta
OpenAI
Researchers from Adobe Research, Stanford University, and Princeton University have developed the LSSVWM architecture for video world models that addresses the problem of long-term memory. The model uses State-Space Models (SSMs) and a block-wise scanning scheme, as well as dense local attention, to combine efficiency with context retention. Experiments on the Memory Maze and Minecraft datasets showed significant improvement in information retention over long time intervals.
Princeton University (Sengupta Lab)
A team from MIT published the paper "Self-Adapting Language Models" with the SEAL framework, enabling large language models to update their own weights through self-editing and reinforcement learning. Experiments on Llama-3.2-1B-Instruct and Qwen2.5-7B showed significant improvements in few-shot learning and knowledge integration.
MIT
Meta
Alibaba/Qwen
OpenAI
DeepMind
ByteDance has developed Astra, an innovative architecture featuring two sub-models (Astra-Global and Astra-Local), designed to overcome the limitations of traditional navigation systems for mobile robots in complex indoor environments. Astra employs a System 1/System 2 paradigm, where the global model handles localization and the local model manages path planning and odometry.
ByteDance
Researchers from BAIR (Berkeley AI) introduced the PEVA (Predicting Ego-centric Video from human Actions) model, which generates next first-person video frames based on past frames and a sequence of human actions specified via 3D pose changes. The model uses an autoregressive conditional diffusion transformer trained on the Nymeria dataset and can predict atomic actions, simulate counterfactual scenarios, and support long video generation up to 16 seconds. PEVA can also be used for planning by evaluating different action sequences based on similarity to a target image.
Berkeley Artificial Intelligence Research (BAIR)
Alibaba Qwen introduced the Group Sequence Policy Optimization (GSPO) algorithm to address instability and model collapse during long RL training. GSPO uses sequence-level optimization, offering higher efficiency and robustness than GRPO, especially for MoE models. The algorithm underpins the Qwen3 models.
Alibaba/Qwen
Alibaba released Qwen-Image, a 20-billion parameter foundational image generation model. It outperforms competitors in complex text rendering, including multi-line captions and Chinese characters, and supports precise image editing.
Alibaba/Qwen
Alibaba Qwen has introduced Qwen-Image-Edit, an image editing model based on Qwen-Image (20B). It supports semantic and external editing, precise text editing in two languages, and achieves state-of-the-art results on open benchmarks.
Alibaba/Qwen
NVIDIA has revealed details about the Blackwell Ultra accelerator: the chip has 208 billion transistors, 160 SMs, 640 fifth-generation tensor cores, supports the new NVFP4 format, and is equipped with 288 GB of HBM3e. The throughput of NVFP4 computations has reached 15 petaflops, and the company has abandoned FP64 development.
NVIDIA
Taiwan Semiconductor Manufacturing Company
Researchers from BAIR (Berkeley Artificial Intelligence Research) have presented the first complete quantitative theory of word2vec training. They proved that in practically relevant regimes, training reduces to principal component analysis (PCA) matrix factorization, with each learned feature corresponding to an interpretable concept. The theory allows computing all features in advance based on corpus statistics and algorithm hyperparameters.
Berkeley Artificial Intelligence Research (BAIR)
Alibaba Qwen has introduced Qwen3Guard, the first safety model in the Qwen family, available in two variants: Qwen3Guard-Gen for offline analysis and Qwen3Guard-Stream for real-time streaming moderation. The model supports 119 languages and a three-tier risk classification (Safe, Unsafe, Controversial).
Alibaba/Qwen
Researchers from BAIR (Berkeley AI) have introduced the Transitive RL (TRL) algorithm, based on a 'divide and conquer' paradigm for off-policy reinforcement learning (RL). TRL reduces the number of Bellman recursions logarithmically, avoiding the error accumulation problems typical of TD learning. The algorithm showed superior results on complex long-horizon tasks without requiring tuning of the hyperparameter n as in n-step TD.
Berkeley Artificial Intelligence Research (BAIR)
The article explores alternatives to standard autoregressive transformer-based LLMs: linear attention hybrids (MiniMax-M1, Qwen3-Next, DeepSeek V3.2, Kimi Linear), text diffusion models, code world models, and small recurrent transformers. The author notes a return to classical attention in MiniMax-M2 and the complexity of linear attention in production.
Moonshot AI
MiniMax
Alibaba/Qwen
DeepSeek
Google/DeepMind
Mistral
Meta
Hugging Face
Allen Institute for AI
xAI
OpenAI
IBM
NVIDIA
Mistral AI announced strategic partnerships with SAP and Helsing to build a sovereign AI stack in Germany and Europe, as well as the opening of an office in Germany.
Mistral
Mistral AI has announced Mistral 3, a new generation of models including dense models Ministral 3 (3B, 8B, 14B) and a sparse mixture-of-experts model Mistral Large 3 (41B active, 675B total parameters). All models are released under the Apache 2.0 license and are available on multiple platforms including Hugging Face, Amazon Bedrock, and Mistral AI Studio.
Mistral
Mistral AI
NVIDIA
Red Hat
vLLM
DeepSeek released its flagship model V3.2 with open weights, which is comparable in performance to GPT-5 and Gemini 3.0 Pro. The article details architectural changes, including the new DeepSeek Sparse Attention (DSA), and the evolution from V3 through R1 and V3.1 to the hybrid model combining reasoning and standard chat.
DeepSeek
Mistral AI has released the next generation of programming models, Devstral 2, in two sizes (123B and 24B parameters), available under open licenses. The models demonstrate high performance on the SWE-bench Verified benchmark despite smaller size and cost. Also introduced is the CLI agent Mistral Vibe for autonomous development task solving.
Mistral
Mistral AI
2025 was a pivotal year for large language models: RLVR and GRPO-based reasoning models dominated. DeepSeek R1 showed that reasoning can be developed through reinforcement learning, and also reduced the estimated cost of training state-of-the-art models. Further expansion of RLVR to other domains and increased attention to inference-time scaling and continual learning are expected.
DeepSeek
OpenAI
Alibaba/Qwen
Moonshot AI
NVIDIA
Google/DeepMind
Nous Research has released the NousCoder-14B coding model, which outperforms many proprietary systems. The model was trained in 4 days on 48 Nvidia B200 GPUs and achieved a 67.87% accuracy on the LiveCodeBench v6 test, 7.08 percentage points higher than Alibaba’s baseline Qwen3-14B. The release coincides with the buzz around Anthropic’s Claude Code tool.
Nous Research
Alibaba/Qwen
Anthropic
Anthropic has introduced Cowork, a new AI agent capability that expands the functionality of the Claude Code tool to non-technical users. The feature is available to Claude Max subscribers through the macOS app. The entire product was built in about a week and a half using Claude Code itself.
Anthropic
OpenAI
Google/DeepMind
Microsoft
The Mistral AI team discovered a slow memory leak in vLLM when serving the Mistral Medium 3.1 model with graph compilation enabled. Using pmap and BPFtrace, they found that the leak occurs not in the heap but through direct mmap calls that bypass the glibc library.
Mistral
In October 2025, Hurricane Melissa became the strongest hurricane ever to hit Jamaica. Thanks to Google DeepMind's AI model WeatherNext, the National Hurricane Center (NHC) forecasted for the first time that the storm would intensify from Category 1 to Category 5 five days before landfall. This enabled early evacuations and saved lives.
Google/DeepMind
Google DeepMind
Google Research
Beijing has officially launched the 'Digital Space-1' project, an experimental satellite for testing 'space brain' technologies capable of autonomous perception, cognition, and decision-making. Chinese scientists and companies aim to shift from 'ground control' to 'onboard intelligence' to address the management challenge of tens of thousands of low-orbit satellites.
Moscow integrated artificial intelligence into radiologists' workflows. AI assists doctors in analyzing images, improving diagnostic accuracy and speed.
A study by K2 NeuroTech found that 65% of companies face a gap between management expectations for AI adoption and actual capabilities. 56% rated their data landscapes as partially ready or not ready for new requirements, and 52% consider IT infrastructure as needing modernization. Meanwhile, 79% of employees already use public AI services, but only 29% of companies have real security control.
Cognizant and Anthropic have announced an expanded partnership: Claude will be embedded into Cognizant's industry platforms. The goal is to help clients bridge the gap between AI promises and real business outcomes.
Anthropic
Cognizant
Despite loud announcements, fully autonomous AI agents in sales often fail to meet expectations: according to UserGems, half of such agents are deactivated within a year. Human salespeople generate 2.6 times more revenue than autonomous systems. However, agent-based lead qualification (AQL) shows good results, increasing conversion of inbound inquiries to validated SQL from 13% to 40%.