AI Updates on 2025-07-28

Zhipu AI releases GLM-4.5 and GLM-4.5-Air models with MIT license, featuring 355B total parameters (32B active) and 106B total parameters (12B active) respectively, both with 128K context length and native function calling @reach_vb
xAI's video generation model Imagine is preparing to launch integrated with Grok, featuring sound capabilities similar to Veo 3 @AndrewCurran_

Jefferies raised China AI capital expenditure forecast for 2025 by 40% to $108 billion, citing that NVIDIA's entire stock of H20 chips only meets about half of China's potential demand @AndrewCurran_
Tesla signs $16.5 billion chip contract with Samsung running until 2033, with Tesla assisting in maximizing manufacturing efficiency for AI chip production @AndrewCurran_
Perplexity usage in India growing rapidly, with CEO noting this as proof that search has changed forever @AravSrinivas
LLMs now direct the majority of discretionary purchases but generate no ad revenue, raising questions about the sustainability of this model @snowmaker
Anthropic introduces new weekly rate limits for Claude Pro and Max plans due to unprecedented demand for Claude Code, affecting less than 5% of subscribers @AnthropicAI
Software engineering roles may need to evolve dramatically with widespread coding assistant use, potentially creating distinct categories: infrastructure/backend/security engineers, research engineers, and app/frontend developers @sayashk

Chinese universities are encouraging students to use more AI rather than restricting it, representing a different approach to AI adoption in education @techreview
UNICEF is investigating how advancing neurotechnology could affect children's rights, with MIT researchers serving as advisors to the project @medialab

Microsoft launches Copilot Mode in Edge browser featuring multi-tab context analysis, voice navigation, and smart task handoff capabilities @mustafasuleyman
Claude can now read and update Notion pages and Linear tickets directly through MCP, enabling project management and documentation updates from conversations @AnthropicAI
Google Chrome adds AI-powered store summaries to help US shoppers make purchasing decisions @TechCrunch
Tesla's FSD Supervised demonstrates understanding of toll booth interactions, automatically proceeding after transaction completion using pillar and side repeater cameras @Tesla_AI
Salient raises $60M Series A for AI agents handling consumer loan servicing, processing over $1B in transactions and cutting handling times by 60% @a16z
Hugging Face launches Jobs CLI powered by uv, enabling one-command VLM-based OCR processing of documents @vanstriendaniel

Language models can create intricate ASCII art without being specifically trained for visual art creation, representing an emergent capability @AITechnoPagan
Direct Preference Optimization (DPO) works by training an implicit reward model and recovering the RLHF-optimal policy in closed form, making it more stable and resource-efficient than PPO-based RLHF @cwolferesearch
DSPy few-shot example selection improved Qwen classification performance from 50% to 88%, demonstrating the importance of proper example curation @MaximeRivest
New GLM-4.5 models show impressive benchmark performance with AIME24 score of 91.0 versus Claude 4 Opus's 75.7, and MATH 500 score of 98.2 versus GPT-4.1's 96.7 @reach_vb
Research on real-time AI companions identifies challenges in achieving 10hz human conversation frequency versus current 1-2hz LLM reaction times, requiring advances in multi-modal processing and long-context understanding @ericjang11