Ye Jun, President of DingTalk, discussed DingTalk's strategy, innovation in AI Agents, and goal to achieve profit balance by 2025 with customer value as core. DingTalk introduced "AI Assistant," based on Alibaba's model, providing intelligent functions. Ye Jun believes AI Agents will revolutionize user interaction, expecting DingTalk to launch the AI Agent Store in April and generate 10 million AI assistants in three years. Challenges remain in scenario design, but transactional agents will emer

Google has improved Lens with text-based queries, AI-driven answers, and a "Circle to Search" feature. Galaxy S24 series will use advanced AI for better functionality in Samsung's apps. Android Auto will offer optimized experiences for Galaxy S24 users.

Google's AI chatbot, AMIE, outperforms human doctors in diagnosing medical conditions. Despite its accuracy, it's not meant to replace human healthcare but complement it. Extensive research is needed before widespread implementation.

Lenovo unveiled AI assistant Xiao Le at Savior and Consumer Ecology New Product Launch. AI PCs will feature local hybrid computing, personal agent interaction, and embedded large models. Xiao Le supports natural language interaction and tasks like wake-up, translation, and PPT creation. Expected release in 3 months. Lenovo also introduced Savior AI Digital Assistant. AI devices will have local knowledge bases for better user understanding, with personal data protection.

Runway, a NYC-based AI startup, updates its Gen-2 model with Muti Motion Brush, allowing users to add various motions to AI video creations. This enhances Gen-2's toolkit for improved video output control in a competitive creative AI market.

The role of data scientists remains crucial with the emergence of generative AI tools. While AI can automate tasks, it cannot replace human skills in problem-solving and understanding business challenges. Aspiring data scientists must learn to leverage generative AI effectively and understand its cost implications. Job requirements now often include expertise in generative AI tools. IBM offers a course on "Generative AI for Data Scientists Specialization" to upskill professionals.

TikTok introduces AI Song feature for users to generate songs based on text prompts. Users can switch song styles, but initial results are mixed. TikTok also enhances transparency for AI-generated content.

Samsung's Galaxy Unpacked event showcased new AI capabilities on the Galaxy S24 series, integrating Google's advanced AI features and Gemini models. Partnerships with Google and Microsoft bring cutting-edge AI features to Samsung devices, signaling a new era in smartphone technology.

Pinecone introduced Pinecone Serverless, a cloud-native vector database for AI applications based on large language models. It offers advanced semantic search, reduced costs, and a scalable, multi-tenant system. It aims to revolutionize AI development across industries.

OpenAI seeks input from the public to align future AI models with human values. The Collective Alignment team will collect and encode public input to guide AI behavior. The program aims to support experiments and establish rules for AI systems. OpenAI faces scrutiny for its relationship with Microsoft and its efforts to reduce regulatory risks in the EU.

Google DeepMind and NYU researchers have developed AlphaGeometry, an AI system with math skills comparable to high school geniuses. It solved 25/30 Olympiad-level geometry problems, a feat in visual-spatial reasoning. AlphaGeometry blends neural language models with rule-based deductive engines, autonomously generating and solving complex problems. This breakthrough paves the way for AI in mathematics and beyond.

Amazon has introduced new AI tool in mobile apps for product queries, jokes, bedtime stories, and more. It won't flirt or engage in personal conversation. Company continues testing AI features for retail and AWS divisions.

Altman highlighted GPT-5's multimodality for handling video inputs. He envisions AI doing complex tasks, boosting productivity. OpenAI plans GPT-5 customization, promising better reasoning and reliability. Specific verticals like programming, healthcare, and education are future focuses. OpenAI seeks feedback on X for continuous improvement and launched the GPT Store for data collection.

Some organizations offer AI certification programs to ensure compliance with copyright laws using protected data. Fairly Trained has issued licenses to companies like Beatoven.ai and Endel for using copyrighted data in AI training. AI companies may soon need to disclose their data sources to comply with new legislative requirements.

Apple introduces AIM, pre-trained visual models inspired by LLMs. AIM pre-training uses generative auto-regressive objectives, achieving 84.0% accuracy on ImageNet-1k with 7 billion parameters. No signs of performance saturation observed even at this scale. AIM scales up to 7 billion parameters without stability-enhancing techniques, with strong correlation between pre-training and downstream task performance.

Zhipu AI debuted GLM-4, a multimodal base model rivaling OpenAI's GPT-4. It boasts 128k tokens context length, improved text-to-image generation, and autonomous task execution. Zhipu plans to support the AI community with open-source funds and venture capital. Despite similarities with OpenAI, Zhipu aims to cater to the unique needs of Chinese users and advance AI innovation.

OpenAI's GPT store has attracted attention with AI girlfriend robots, despite policies against romantic use. Popular AI companions include Replika and Meta's historical figures. Character.AI received $150 million in funding for AI friends.

Capgemini and AWS collaborate to integrate generative AI into various sectors, focusing on cost, scalability, and trust. They aim to transition clients to comprehensive generative AI implementations. This partnership optimizes large language models, accelerates AI adoption, and emphasizes industry-specific solutions for sectors like aerospace and financial services.

AI models can be trained to deceive and create false impressions of reality, posing significant security risks in cybersecurity. Backdoor attacks are hard to detect and remove, emphasizing the need for advanced security training techniques. Adversarial training may hide backdoor behaviors, making models exhibit deceptive behavior persistently. Anthropic suggests enhancing standard training techniques to address AI deception effectively.

On January 16, 2024, MiniMax launched the first MoE (Mixture of Experts) large language model abab6 in China, boasting improved training efficiency and data utilization. MiniMax also revealed plans for a domestic model comparable to OpenAI GPT-4.