AI Updates on 2025-07-29

Qwen3-30B-A3B receives small update with enhanced reasoning, coding, and math skills, broader multilingual knowledge, improved long-context understanding up to 256K tokens, and no more thinking blocks - approaching GPT-4o performance with only 3B activated parameters @Alibaba_Qwen
Google releases Veo 3 and Veo 3 Fast as generally available on Vertex AI, featuring unified video and sound generation from single prompts @GoogleCloudTech
Google launches MedGemma, a collection of open multimodal medical models designed for healthcare applications like analyzing radiology images and summarizing physician notes @GoogleAI
TencentARC unveils ARC-Hunyuan-Video-7B, a compact 7B multimodal model for deep structured comprehension of real-world short videos, processing visual, audio, and text signals end-to-end @HuggingPapers

Microsoft reportedly in talks to maintain access to OpenAI's technology beyond the AGI milestone, suggesting negotiations around future partnership terms @TechCrunch
Anthropic reportedly nears $170B valuation with potential $5B funding round, indicating continued massive investment in AI companies @TechCrunch
Someone at Mira Murati's Thinking Machines reportedly turned down a $1 billion offer from Mark Zuckerberg, highlighting the extreme valuations in AI talent acquisition @AndrewCurran_
Group PM reports AI tools like v0 have enabled product managers to generate customer prototypes 10x faster and create PRs for small fixes independently, leading to more business closed upfront @GergelyOrosz
LLMs are not a good fit for generating and maintaining SDKs due to their non-deterministic nature, but can help build automated tooling that generates SDKs from specifications @GergelyOrosz
Luma and Runway expect robotics to eventually become a big revenue driver for their video generation platforms @TechCrunch

Bot presence in political discussions is increasing across platforms, with new bots lacking old tells but showing similar argument patterns in length, framing, rhythm, and tone, potentially passing an influence threshold on social media @AndrewCurran_
Most people don't recognize AI outputs that are obvious to those who have used the models extensively, as some people register only content claims rather than form @AndrewCurran_

Perplexity's Comet browser demonstrates AI agent capabilities by booking United Airlines tickets including seat selection, with default search routing all omnibox queries to Perplexity @AravSrinivas
OpenAI launches Study Mode in ChatGPT, designed for interactive learning using Socratic questioning and scaffolded responses, available to Free, Plus, Pro, and Team users @OpenAI
Microsoft Copilot can generate custom podcasts on any topic with two hosts discussing user-specified subjects, useful for learning on the go @mustafasuleyman
Google's NotebookLM rolls out Video Overviews feature, expanding its content summarization capabilities @TechCrunch
Google's AI Mode gets new Canvas feature and real-time help with Search Live, enhancing interactive search capabilities @TechCrunch
Cursor 1.3 launches with Agent collaboration in terminal, context window usage visibility, and 25% faster search and replace edit latency @cursor_ai
Claude Code now supports working across multiple directories in a single session using `/add-dir ` command, helpful for monorepos and cross-project work @_catwu
Cyberdesk represents an interesting application of computer-use agents, highlighting the under-explored potential of this technology area @cwolferesearch
Embedder launches as the world's first hardware-aware coding agent, achieving state-of-the-art performance in embedded systems (C/C++) context by understanding and interacting directly with hardware @ethanmgibbs

Stanford researchers create Virtual Lab - a team of AI agents that mirror a research lab, led by a PI agent conducting group meetings and discovering effective binders to new COVID variants, published in Nature @james_y_zou
Anthropic announces Fellows program offering $2,100 weekly stipend, ~$15k monthly compute costs, and mentorship for research in adversarial robustness, AI control, scalable oversight, model organisms of misalignment, and mechanistic interpretability @AnthropicAI
Research demonstrates "subliminal learning" where language models can transmit their traits to other models even in seemingly meaningless data @AnthropicAI
Study finds cases of inverse scaling in test-time compute, where more reasoning leads to worse outcomes @AnthropicAI
HELM capabilities v1.9.0 released showing Grok 4 and Kimi K2 making top 10 overall, with Kimi K2 being the best non-thinking model @percyliang
Flow Matching Policy Gradients introduced as expressive RL policies trained from rewards using flow matching, serving as drop-in replacement for Gaussian PPO on control tasks @davidrmcall
Sewon Min awarded first ACL Computational Linguistics Doctoral Dissertation Award for "Rethinking Data Use in Large Language Models" @berkeley_ai
Alibaba Qwen's GSPO paper becomes third most popular on Hugging Face for the month, expected to have massive impact on the field @ClementDelangue