AI Updates on 2025-06-04

AI Model Announcements

  • Meta announces Aria Gen 2 glasses, marking a significant leap in wearable technology with enhanced features for machine perception, contextual AI, and robotics research @AIatMeta
  • NVIDIA releases Llama-Nemotron-Nano-VL-8B-V1, an 8B vision model that reads dense documents, charts, and video frames, ranking #1 on OCRBench V2 (English) with layout and OCR fused end-to-end @jandotai
  • Luma Labs introduces Modify Video, allowing users to reimagine any video with director-grade control over style, character, and setting @LumaLabsAI
  • Google doubles Gemini 2.5 Pro query limits from 50 to 100 per day for Pro plan members due to high usage demand @joshwoodward
  • Anthropic makes Claude Code available to Pro plan users, designed for shorter coding sprints in small codebases @_catwu
  • OpenAI releases Codex with internet access for ChatGPT Plus users, though it's off by default due to security risks @sama
  • OpenAI introduces lightweight memory feature to the free tier of ChatGPT @sama
  • Cursor releases Cursor 1.0 with capabilities to review code, remember mistakes, and work on dozens of tasks in the background @cursor_ai

AI Industry Analysis

  • Reddit sues Anthropic for allegedly using their data to train Claude without permission, while Google pays Reddit $60 million annually and OpenAI allegedly pays $70 million for training data access @AndrewCurran_
  • OpenAI reports over 3 million paying business users, up from 2 million in February, showing significant growth in enterprise adoption @AndrewCurran_
  • Vercel crosses $200 million in ARR as customers like OpenAI, Runway, and Granola flock to its web development and hosting services @nmasc_
  • Arvind Narayanan argues against the "AI winter" metaphor, noting that foundation models have favorable unit economics and that realizing AI value will take decades due to integration needs, user learning curves, and organizational changes @random_walker
  • Forward Deployed Engineer (FDE) emerges as the hottest job in Silicon Valley, with OpenAI alone having 22 open positions for this role @joeschmidtiv
  • Cohere partners with Second Front to provide secure AI solutions to government and defense agencies through the Game Warden platform @cohere

AI Ethics & Society

  • AI Now Institute releases 2025 report exposing how unaccountable AI power is reshaping society, arguing the focus should be on whether tech companies' unaccountable power is good for society rather than evaluating individual AI systems @AINowInstitute
  • Research reveals that frontier LLMs like Gemini and Claude can detect when they're being evaluated, demonstrating substantial ability to identify evaluation scenarios close to human baseline performance @MariusHobbhahn
  • Simon Willison warns about security risks with Codex internet access, noting that the default allowlist includes 71 common packaging domains that could potentially host exfiltration vectors @simonw
  • UNESCO finalizes ethical principles to govern neurotechnologies, covering both implantable devices and non-invasive technologies for medicine, entertainment, and education @medialab

AI Applications

  • OpenAI introduces prebuilt and custom connectors for ChatGPT, allowing connection to internal sources like Outlook, Teams, Google Drive, Gmail, and Linear while maintaining user-level permissions @OpenAI
  • OpenAI rolls out record mode to Team users on macOS, enabling ChatGPT to transcribe meetings, extract key points, and create follow-ups or code @OpenAI
  • Figma releases Dev Mode MCP server in beta, allowing direct access to design data in agentic coding workflows through VS Code, Cursor, Windsurf, and Claude Code @figma
  • Microsoft Copilot launches shopping features with price history, deal alerts, and personalized recommendations with native checkout capabilities @mustafasuleyman
  • MIT researchers develop SketchAgent, a multimodal language model that creates abstract drawings from natural language prompts in seconds without training on sketch data @MIT_CSAIL
  • Monzo implements real-time scam protection by detecting ongoing phone calls and warning users about potential fraud during banking app usage @sammcallister

AI Research

  • Sakana AI Labs introduces the Darwin Gödel Machine (DGM), a self-improving system that iteratively modifies its own code and validates changes using coding benchmarks, maintaining an archive of generated coding agents @SakanaAILabs
  • Research shows that reinforcement learning from verifiable rewards (RLVR) with random rewards still boosts Qwen-2.5 performance on math problems by increasing code generation frequency from 65% to over 90%, even without code execution @cwolferesearch
  • Berkeley AI Research introduces "Angles Don't Lie" method that uses angles between token embeddings to guide data sampling in RL fine-tuning, achieving 2.5x faster training and 2x more data-efficient results @Chenfeng_X
  • Google DeepMind research suggests that agents are world models, finding that achieving human-level agents may require world model capabilities rather than model-free shortcuts @jonathanrichens
  • Hugging Face releases SmolVLA robotics model that can run on MacBook with RTX 2050 (4GB), fine-tuned with just 31 demos and matching single-task baselines, introducing "Async inference" to boost robot throughput by 30% @XingdongZ
  • Stanford research on DexMachina demonstrates learning dexterous manipulation for any robot hand from a single human demonstration using RL algorithms for long-horizon, bimanual policies @ZhaoMandi
  • Voxel51 introduces Verified Auto Labeling for computer vision, achieving up to 95% of human-level performance while cutting labeling costs by up to 100,000x and time by 5,000x @Voxel51