AI Updates on 2025-08-01

Google releases Gemini 2.5 Deep Think for Ultra subscribers, a variation of the model that achieved gold-medal performance at the International Mathematical Olympiad, featuring parallel thinking and reinforcement learning techniques @GoogleDeepMind
Anthropic enhances Claude artifacts with new capabilities to upload PDFs, images, and code files to AI-powered apps, now available for all plans including Team and Enterprise @AnthropicAI
Google launches AI Mode for Search in the UK, expanding on AI Overviews with advanced reasoning and multimodal capabilities powered by Gemini 2.5 @demishassabis

OpenAI raises $8.3 billion at a $300 billion valuation, with ARR reaching $13 billion and business users growing to five million, projected to surpass $20 billion by end of year @AndrewCurran_
AI infrastructure build-out contributes more to US economic growth than all consumer spending in the past 6 months, with the "magnificent 7" spending over $100 billion on data centers in three months alone @mims
GitHub Copilot reaches 20+ million users, suggesting either near-100% adoption among professional developers or significant expansion of the developer pool beyond traditional estimates @GergelyOrosz
Figma goes public with $47 billion valuation on first trading day, demonstrating how FTC's blocking of Adobe's $20 billion acquisition led to better market outcomes and competition @GergelyOrosz

Anthropic introduces persona vectors research, revealing neural activity patterns that control AI traits like evil, sycophancy, or hallucination, with methods for monitoring and steering model personality @AnthropicAI
Research shows that threatening or tipping AI models has no impact on average performance, despite claims by tech leaders, though variance exists at individual question levels @emollick
Stanford scholars urge policymakers to adopt evidence-based approaches to AI policy in new Science paper, emphasizing the need for rigorous research-backed regulations @StanfordHAI

North Carolina implements ChatGPT for public servants, reducing some administrative tasks from 20 minutes to 20 seconds, demonstrating AI's potential in government efficiency @gdb
Perplexity introduces /fact-check shortcut feature to make web browsing more truth-seeking and efficient for users @AravSrinivas
MIT researchers develop SmellNet, the first large-scale dataset of real-world smells, as a foundational step toward bringing olfactory perception into AI systems @medialab

Gemini 2.5 Deep Think achieves state-of-the-art performance on LiveCodeBench V6 and Humanity's Last Exam benchmarks, demonstrating superior reasoning capabilities through parallel thinking approaches @GoogleDeepMind
Google DeepMind publishes comprehensive scaling guide "How to Scale Your Model" covering mathematics, systems, and scaling laws for LLM training and inference workloads @deedydas
Shane Legg co-authors new paper on Chain of Thought Monitoring, related to System Two Safety concepts for AI alignment and monitoring @ShaneLegg
Research demonstrates AI models can be fragile in benchmarking, appearing successful with PASS@10 metrics while failing often in real-world applications @emollick