CES 2025 kicks off in Las Vegas on Jan 7, with major tech companies like AMD, Samsung, Toyota, and Nvidia showcasing innovations. Key events and press conferences will be live-streamed, with Media Day on Jan 6 featuring significant announcements.

Google CEO Sundar Pichai highlighted 2025 as a crucial year, prioritizing the expansion of the AI model Gemini to consumer applications, aiming to lead in AI technology.

DeepSeek-V3, the latest large language model, outperforms leading non-open-source models with a training cost of only $5.6 million. It uses innovative techniques like unsupervised loss load balancing and Mixture-of-Experts architecture, making it 11 times more efficient than Meta's LLaMA 3.

OpenAI will transition to a public benefit corporation in 2025, transferring control to its for-profit division while the non-profit focuses on charitable projects. This change aims to secure more funding for AI development.

Microsoft and OpenAI have agreed on a new AGI definition, crucial for future collaboration. OpenAI aims for $100 billion in profit to achieve AGI, while Microsoft focuses on commercial benefits. The agreement includes a clause limiting Microsoft's access to OpenAI's models upon AGI achievement, causing uncertainty. OpenAI seeks to remove this clause. Despite recent model successes, true AGI remains elusive. Microsoft plans to diversify AI models in its products and may invest in Anthropic, valui

ChatGPT experienced a service disruption on Thursday, with issues starting around 1:30 PM Eastern Time. OpenAI reported high error rates for ChatGPT, its API, and Sora. After emergency repairs, Sora resumed operations by 6:15 PM, but full restoration of ChatGPT and its API was still ongoing. Microsoft, OpenAI's cloud provider, also faced a power issue at a data center, affecting services in North America.

Tencent Research Institute launched DRT-01, an AI model for literary translation using Chain of Thought (CoT) technology. It enhances understanding of metaphors and similes, improving BLEU and CometScores. The model features a multi-agent framework and iterative optimization, ensuring high-quality translations. Trained on 400 public domain books, DRT-01 improves accuracy and explainability in translating figurative language.

The Guardian reported that ChatGPT Search can be misled by hidden text on websites, potentially generating inaccurate or even malicious content, highlighting a vulnerability in the AI search engine.

Microsoft and OpenAI's partnership includes a clause allowing OpenAI to end the deal if it achieves AGI, but this requires generating $10 billion in profit, a significant challenge for the currently unprofitable company.

Step Celestial's Step-1X-Medium model has been upgraded, increasing generation speed by over 30% and adding a base image enhancement feature for more detailed and versatile image creation. It now better captures traditional Chinese styles and supports English text in instructions, available on their open platform.

Zhipu AI launched CogAgent-9B-20241220, a specialized GUI interaction model based on GLM-4V-9B, open-sourced for community development. It excels in GUI tasks, supports bilingual inputs, and shows significant improvements over its predecessor.

DeepSeek V3, with 6850 billion parameters and a MoE architecture, excels in multilingual programming, achieving 60 TPS and outperforming other open-source models in coding, mathematics, and factual knowledge. It supports FP8 training and uses OCRv12 for multimodal capabilities.

Xiaomi is building a large GPU cluster to boost AI model development, led by founder Lei Jun. The company has been investing in AI for years, with a team now exceeding 3,000 people, and aims to integrate AI into various business segments.

Researchers from University at Albany, UC Santa Barbara, Amazon Alexa AI, and Meta introduced CoMERA, a framework that uses rank-adaptive tensor compression to reduce memory usage, computational costs, and training time while maintaining model accuracy. CoMERA achieved up to 361x compression in specific layers and 99x in full models, significantly lowering storage and memory requirements. It provides 2-3x faster training time per epoch for transformers and recommendation systems, making it more

Jack Clark of Anthropic refutes claims of AI progress slowing, citing OpenAI's o3 model as evidence. He predicts accelerated AI growth by 2025 through new training methods and increased computational power, despite rising costs.

Quantization in AI, aimed at enhancing model efficiency, is hitting performance limits. While it reduces computational complexity, a study shows that quantized models may perform worse if the original model is trained with large data over time. This trend negatively impacts companies relying on large models for cost reduction. Researchers suggest focusing on data quality and developing architectures for low-precision training.

The o3 model scored 75.7% on the ARC-AGI benchmark, reaching 87.5% with high computational resources, showing unprecedented task adaptability. However, it still fails at some simple tasks and requires human-annotated reasoning chains, indicating a gap from true AGI.

At CES 2025, AMD will face tough competition from Nvidia, which is expected to launch the RTX 5000 series. AMD plans to unveil its next-gen GPUs, possibly named RX 8000 or RX 9000, based on RDNA 4, aiming to strengthen its market position with advanced features like improved ray tracing and energy efficiency.

Tencent Research Institute launched DRT-o1 models, enhancing literary translation with long chain-of-thought technology. Trained on 63,000 sentences with metaphors and similes, the model uses a multi-agent framework for iterative refinement, improving BLEU and CometScores.

Researchers at OpenAI have developed "Deliberative Alignment," a new method to improve the safety of large language models by directly teaching them safety norms and using reinforcement learning, showing better performance in resisting harmful content and reducing false rejections.