Microsoft Investigates DeepSeek
Microsoft continues to show strong growth across its business sectors, especially in cloud and AI. CEO Satya Nadella predicts a 175% growth in AI-related business for the year. DeepSeek, a Chinese AI company, now has its AI model R1 available on Microsoft’s Azure AI Foundry and Github, allowing users to test and integrate it. Microsoft praises R1 for its cost-effectiveness, making advanced AI accessible with minimal infrastructure investment. This seems like a jab at the more expensive OpenAI.
However, Microsoft and OpenAI are investigating whether DeepSeek has violated usage terms by accessing OpenAI data improperly via API. DeepSeek’s models reportedly claimed to be ChatGPT, suggesting unauthorized use of OpenAI’s output, which is against OpenAI’s usage policies.
Anthropic CEO Debunks AI Myths
Dario Amodei, CEO of Anthropic, clarifies misconceptions in the AI debate. The Claude 3.5 Sonnet model was developed for a few tens of millions of dollars, not billions as rumored. Amodei criticizes DeepSeek R1, saying it only reproduces known methods. He highlights the “shifting the curve” principle, where costly production today becomes cheaper over time, emphasizing optimization over innovation. The real innovation is in Deepseek-V3, not R1.
Amodei predicts future AI systems will need millions of chips and could surpass human intelligence by 2026/27. Reinforcement Learning is becoming a key focus in AI development. He views US chip export controls to China as essential for maintaining global balance.
Alibaba’s Record AI Model
Alibaba introduces Qwen2.5-Max, trained with over 20 trillion tokens, surpassing competitors like DeepSeek-V3 and Llama 3.1, which have 14-15 trillion tokens. In benchmarks, it competes with models like GPT-4 and Claude 3.5. Alibaba uses methods like Supervised Fine-Tuning and Reinforcement Learning, though the data source is unclear, possibly involving synthetic data. Qwen2.5-Max is available via Alibaba’s Cloud API and Qwen Chat, with a focus on attracting developers with competitive pricing, despite Chinese government censorship.
Meta’s Personalized AI Bubble
Mark Zuckerberg is optimistic about Meta’s future, with 3.3 billion daily users by end of 2024. Meta AI reached 700 million users in December, and aims to personalize responses based on user data from Facebook and Instagram. The upcoming Llama 4 model will be multimodal, processing text, sound, images, and video, working independently to understand user preferences. Meta continues to rely on advertising, now optimized with AI.
OpenAI’s ChatGPT Pro Success
OpenAI’s ChatGPT Pro subscription generates at least $25 million monthly, but the company still incurs losses due to high server costs. CEO Sam Altman aims for $12 billion in revenue by 2025, with $8 billion from ChatGPT. Price increases are likely.
Figure AI’s Robot Safety Center
Figure AI, testing its humanoid robot Figure 02, is establishing a safety center for robot-human interaction standards. Led by Rob Gruendel, the focus is on stability, human and animal recognition, AI behavior, and accident prevention. The findings will inform guidelines and certifications for various applications, with quarterly progress updates.
LKA Chief Warns of Identity Theft
Mario Germano, head of the Rhineland-Palatinate State Criminal Police Office, warns against misuse of public audio and video data by AI. He notes that AI can create convincing fake content, posing risks for personalized phishing. Germano advises using family code words to prevent scams like impersonation tricks.