AI Updates on 2025-09-04

Google releases EmbeddingGemma, a new open embedding model with 308M parameters that achieves state-of-the-art performance on the MTEB benchmark while being small enough to run completely on-device @sundarpichai
Perplexity announces Comet is now available for pre-orders on Android Play Store and available to Pro users in South Korea, Brazil, and Spain @AravSrinivas
Google announces Veo 3 integration into Google Photos' photo-to-video feature, upgrading the video generation capabilities @TechCrunch
Jina AI releases jina-code-embeddings, a new suite of code embedding models in 0.5B and 1.5B parameter sizes with SOTA retrieval performance supporting over 15 programming languages @JinaAI_

Andrew Ng identifies significant unmet demand for AI engineers who can use AI assistance to rapidly engineer software systems, while recent CS graduates face increased unemployment due to universities not adapting curricula to AI-native programming @AndrewYNg
Reid Hoffman discusses Stanford study showing 16% drop in entry-level jobs for 22-25 year-olds in AI-exposed fields, emphasizing need for new career pathways in the AI era @reidhoffman
Gergely Orosz criticizes Coinbase CEO's mandate to increase AI code generation percentages, arguing it focuses on tool usage metrics rather than business outcomes like customer satisfaction or product reliability @GergelyOrosz
Mustafa Suleyman highlights that frontier AI models are now 90% cheaper but 2.7x better than two years ago, emphasizing the leap forward in accessibility @mustafasuleyman
Deedy reports that 95% of Gen AI pilots don't fail according to MIT study, contradicting common narratives about AI project failure rates @deedydas
Lenny Rachitsky identifies evals as emerging must-have skill for product builders and AI companies, comparing it to SQL and Excel as fundamental competencies @lennysan
Sam Altman reports Codex usage up 10x in past two weeks, showing impressive momentum for AI coding tools @sama
Aravind Srinivas announces over one million people got Comet access in one morning, calling it the most widely deployed personal and agentic product in the world @AravSrinivas

Sam Altman observes increasing prevalence of LLM-run Twitter accounts, noting he's taking the dead internet theory more seriously @sama
Microsoft Research introduces the Sui Generis score to measure narrative diversity in LLM outputs, revealing how AI storytelling often creates repetitive, less unique narratives @MSFTResearch

Ribera, a Spanish healthcare company, uses AI to improve discharge systems for cataract surgery patients @Microsoft
OpenAI launches conversation branching feature in ChatGPT, allowing users to explore different directions without losing original thread @OpenAI
Google introduces Circle to Search translation feature and upgrades Gemini App image editing capabilities @TechCrunch
Notion databases now support AI-powered features for enhanced data processing and analysis @brian_lovin
TechCrunch reports OpenAI Jobs Platform set to launch in mid-2026, using AI to match candidates with businesses @TechCrunch
Supersonik AI launches as the first AI that can run live product demos, raising $5M led by a16z @danipolymath

Ethan Mollick shares research finding that LLMs' Theory of Mind abilities come from just 0.001% of their parameters, and breaking those specific weights causes loss of both belief tracking and language comprehension @emollick
Google DeepMind publishes Deep Loop Shaping method in Science Magazine that reduces noise in LIGO gravitational wave observatories by 10x or more, helping detect black hole mergers @GoogleDeepMind
Stanford researchers introduce Mixture of Contexts for generating minute-long videos in a single pass without drifting or forgetting historical context @GordonWetzstein
Research paper finds AI agents can be used for social science experiments when prompts are developed based on social science and game theory, making AI agent actions predictive of real human outcomes @emollick
New study evaluates AI agents' web browsing capabilities using Online Mind2Web benchmark, testing 9 models including GPT-5 and Sonnet 4 with different agent scaffolds @sayashk
Research paper challenges hallucination detection evaluation methods in LLMs, finding significant problems with current field practices @ziv_ravid
Hugging Face releases FineVision, a massive open-source dataset with 17.3M images and 24.3M samples for training Vision-Language Models @thibaudfrere