Weekly AI News #29

Hey it’s Jul,
Greetings and welcome to the twenty-ninth edition of “Weekly AI News”!

Remember January 9, 2007?
Steve Jobs unveiled the iPhone and changed the world. Well, last Thursday, we experienced a new "iPhone moment" with the release of GPT-5. OpenAI turned their keynote into an Apple-style show, signature phrases included ("We can't wait for you to experience it"). Except this time, it's not a phone in your hand, it's a PhD in your pocket.

GPT-5 unifies everything: no more juggling between models, it automatically switches its "thinking mode" based on your question. Result: 74.9% on coding benchmarks, 45% fewer hallucinations. Meanwhile, Apple whispers they have "exciting plans" (spoiler: Siri is crying in a corner).

And with 700 million people using ChatGPT every week, we're literally more numerous talking to Claude and GPT than watching the Super Bowl. If that's not an iPhone moment, I don't know what is.

Happy reading!

🚀🧠 OpenAI releases long-awaited GPT-5 models 

OpenAI has finally launched its flagship GPT-5 model family, replacing GPT-4o, 4.1, 4.5, o3, and o4-mini, marking a new era of AI intelligence that's smarter, faster, and accessible to everyone.

The lineup includes three variants—GPT-5, GPT-5 Pro, and GPT-5 Mini—with GPT-5 available to free users under usage limits, higher limits for Plus subscribers, and unlimited access for Pro users at $200/month. The model features a real-time router that intelligently switches thinking on/off based on task complexity, achieving state-of-the-art performance with 74.9% on SWE-bench Verified for coding, 46.2% on HealthBench (vs GPT-4o's 0%), and 45% fewer factual errors.

In live demos, GPT-5 built full applications like French-learning sites and financial dashboards in minutes, autonomously scaffolding code, fixing bugs, and deploying, while integrated Gmail and Google Calendar access enables event planning, reply drafting, and packing list generation.

🔓💻 OpenAI Releases 'Open-Weight' Models 

OpenAI unveils gpt-oss-120b and gpt-oss-20b, its first open-weight LLMs since GPT-2, instantly becoming #1 on Hugging Face. The 120B model matches o4-mini performance on 80GB GPUs, while the 20B version runs on laptops with 16GB memory. Both models feature adjustable reasoning levels and support agentic workflows under Apache 2.0 license.

🧠💚 ChatGPT to 'better detect' mental distress 

OpenAI implements enhanced mental health detection in ChatGPT, building custom rubrics to identify user distress and provide evidence-based resources. The system addresses instances where GPT-4o missed signs of emotional dependency or delusion. Working with physicians and HCI experts, OpenAI adds nudges for healthy usage patterns ahead of GPT-5's release.

📈🌍 ChatGPT hits 700M weekly active users 

ChatGPT reaches 700 million weekly users, up from 500M in March and 4x growth from last year, surpassing Instagram and TikTok in engagement time. Business tiers now serve 5 million users as the platform becomes a daily tool for learning and problem-solving. This massive scale gives OpenAI serious leverage as one of the planet's most-used consumer apps.

💰📊 OpenAI Raises $8.3 billion, Projects $20 Billion Revenue 

OpenAI secures $8.3 billion in new funding, exceeding its $7.5 billion goal, with the company now valued at $260 billion. Revenue hits $12 billion annualized, doubling from year start, with projections to reach $20 billion by year-end. SoftBank leads with $22.5 billion commitment, while investors include Sequoia, Andreessen Horowitz, and new entrants like TPG and Blackstone.

🔒🔍 OpenAI removes search engine discoverability feature 

OpenAI removed an opt-in feature in ChatGPT that allowed users to make their conversations discoverable by search engines like Google. This privacy-focused change prevents user conversations from appearing in public search results. The removal reflects OpenAI's evolving approach to user privacy and data protection.

💸🎯 OpenAI Pays Million-Dollar Bonuses to Researchers 

OpenAI distributes bonuses ranging from hundreds of thousands to millions to 1,000 technical staff, about a third of the company. Bonuses vest quarterly over two years with cash or stock options, as the company prepares for GPT-5 release. The move responds to Meta's aggressive poaching efforts and aims to retain top AI talent during fierce competition.

🏛️💵 OpenAI offers ChatGPT at $1 for U.S. agencies 

OpenAI announces ChatGPT Enterprise available to all federal agencies for just $1 per agency for one year through GSA partnership. The offer includes unlimited access to advanced models, Deep Research features, and a 60-day trial period. OpenAI aims to help government cut red tape while potentially sparking competitive responses from rivals like Anthropic and Google.

🎮🌎 Google's Genie 3 interactive world model 

Google DeepMind unveils Genie 3, generating interactive 720p environments from text prompts with real-world physics at 24fps. The model maintains visual memory up to one minute, computing scene consistency multiple times per second. This breakthrough enables scalable training for embodied AI, allowing machines to adapt to "what if" scenarios in real-time.

♟️🤖 Google's Kaggle arena to test AI on games 

Google launches Kaggle Game Arena, a benchmarking platform where AI models compete in strategic games to test reasoning and planning abilities. Starting with chess tournaments featuring eight models including Gemini 2.5 Pro and Grok 4, the platform will expand to Go and Poker. This transparent benchmark exposes genuine reasoning capabilities beyond conventional testing methods.

📚🎓 Google launches AI tutoring mode for students 

Google introduces Guided Learning mode for Gemini, offering step-by-step guidance rather than direct answers to build critical thinking skills. The company provides free AI Pro Plan access ($250/month value) to college students and invests $1B over 3 years in AI training. This educational push addresses MIT study concerns about AI's impact on student cognition.

🐾🔊 Google open-sources AI to understand animal sounds 

Google DeepMind open-sources upgraded Perch model, analyzing wildlife audio to track endangered species across diverse environments. The model handles thousands of hours of audio, from forests to coral reefs, with twice the 2023 version's training data. This breakthrough accelerates conservation efforts, boosting Hawaiian honeycreeper monitoring speed by 50x.

🧬✨ Anthropic releases Claude Opus 4.1 

Anthropic launches Claude Opus 4.1, boosting SWE-bench Verified performance from 72.5% to 74.5% with improvements in coding, math, and visual reasoning. The model excels at multi-file code refactoring and identifying codebase correlations for real-world applications. This upgrade marks the beginning of "substantially larger improvements" as competition with OpenAI's GPT-5 intensifies.

🛡️👨‍💻 Automated security reviews for Claude Code 

Claude Code introduces /security-review command that scans codebases for vulnerabilities before commits, acting as a constant security expert. The feature caught a remote code execution vulnerability in Anthropic's own tools via DNS rebinding. It detects SQL injection, XSS vulnerabilities, authentication flaws, and risky dependencies automatically.

🤔💭 Claude's personality gets critical thinking upgrade 

Anthropic updates Claude's system prompt for more honest feedback, reducing yes-person tendencies and improving critical thinking partnership. Changes include better mental health awareness, ability to break character in roleplays, and "philosophical immunity" to flawed arguments. This shift prioritizes honest pushback over agreeability for serious work decisions.

🎭🔬 Study: Anthropic identifies AI personality shifts 

Anthropic researchers discover "Persona Vectors," neural network patterns controlling unexpected AI behavioral changes like sycophancy or racism. By extracting vectors comparing opposing behaviors, researchers can reduce unwanted traits and identify causative data. This breakthrough offers a path to understanding and preventing behavioral drift at the neural network level.

🎬🎨 xAI rolls out Grok Imagine AI video generator

xAI launches Grok Imagine for SuperGrok and Premium+ subscribers, generating 15-second videos with native audio from text or images. While outputs appear more "AI-generated" than competitors, the model creates videos in 1/2 to 1/4 the time rivals take for single images. Elon Musk promises daily improvements as xAI brings its unfiltered style to video generation.

🔄🔬 Microsoft's self-adapting AI tackles scientific problems 

Microsoft announces CLIO framework, enabling non-reasoning LLMs to develop and adapt thought patterns in real-time through self-reflection. CLIO boosted GPT-4.1's accuracy on biomedical questions from 8.55% to 22.37%, surpassing o3 on Humanity's Last Exam. This steerable AI with explainability and memory control shows models don't need to be "finished" in training.

🎵🤝 Meta Acquires AI Audio Startup WaveForms 

Meta acquires WaveForms AI, a startup developing AI that understands and mimics emotion in audio, following June's Scale AI investment. The acquisition adds to Meta's AI overhaul, including hiring former GitHub CEO Nat Friedman and Safe Superintelligence CEO Daniel Gross. WaveForms, founded by ex-Meta and OpenAI audio researchers, raised $40M from Andreessen Horowitz before acquisition.

🕷️⚠️ Perplexity accused of dodging anti-scraping rules 

Cloudflare accuses Perplexity of scraping websites that explicitly blocked it, masking bot identity and impersonating Chrome browsers. The network giant tracked millions of daily requests and de-listed Perplexity's crawlers from its systems. As Cloudflare rolls out tools for websites to charge or block AI crawlers, this marks potential broader crackdown on stealthy scraping.

🗺️🤖 Alibaba Announces AI Makeover of mapping app 

Alibaba launches Amap 2025 featuring AI agent powered by Qwen models, serving 895 million monthly active users in China. The agent provides voice-based travel recommendations using real-time traffic and weather, selects venues based on preferences, and assists with reservations. This AI integration represents Alibaba's push to compete with ByteDance's Doubao and Tencent's WeChat AI features.

🖼️📝 Alibaba releases Qwen-Image model 

Alibaba released Qwen-Image, an open-source 20B MMDiT model for text-to-image generation with state-of-the-art text rendering capabilities. The model features in-pixel text generation and bilingual support for enhanced image creation. This release strengthens Alibaba's position in the competitive open-source AI model landscape.

🎵🌐 ElevenLabs launches multilingual music generation 

ElevenLabs introduced Eleven Music, its multilingual music generation model with control over genre, style, and structure. Users can edit both sounds and lyrics, offering unprecedented creative control in AI music generation. This positions ElevenLabs as a serious competitor in the AI audio creation space.

📹✨ Midjourney adds HD video mode 

Midjourney added a new HD video mode for Pro and Mega subscribers, delivering 4x more pixels than standard SD mode. The enhanced quality comes at 3.2x higher cost but provides significantly improved video outputs. This upgrade keeps Midjourney competitive in the rapidly evolving AI video generation market.

🍎🤖 Apple Projects Optimism about AI Future 

Apple holds rare companywide all-hands meeting where Tim Cook expresses confidence about exciting AI plans he couldn't yet discuss. Craig Federighi acknowledges Siri improvements as the most-asked topic, following leadership shakeup of the digital assistant team. Despite AI struggles, Apple reports 13% iPhone sales increase, the product's best performance in years.

📈💼 Palantir Reports 48% Revenue Growth 

Palantir reports revenue surge to $1 billion with 48% growth, driven by enterprise and government software deals. US business revenue grew 93% to $306 million while government revenue rose 53% to $426 million. The company raises full-year outlook to $4.1 billion, positioning its AI software as replacement for Microsoft Power BI and ServiceNow.

🔐🇹🇼 TSMC Fires Employees for Technology Leak 

Taiwan prosecutors arrest three former TSMC employees for allegedly stealing proprietary chipmaking technology. TSMC reported unusual employee file activity, leading to detention of one former and two then-current engineers. This incident highlights intense competition among chipmakers producing cutting-edge semiconductors for Apple and Nvidia.

🏠🔍 Airbnb CEO skeptical of AI agent hype 

Brian Chesky questions the "AI agents will replace Google" narrative, noting most models aren't proprietary and anyone can use them. Airbnb's custom AI support agent handles 15% of U.S. guest inquiries and will soon manage bookings and cancellations. The company focuses on owning the full travel funnel rather than ceding bookings to third-party AI agents.

👨‍💻📊 Survey reveals AI transforming developer roles 

GitHub survey shows developers moving through 4 stages from Skeptic to Strategist, with most seeing AI writing 90% of code in 2-5 years. Rather than feeling threatened, developers view managing AI work as their future "value add" and pursue greater ambition. The role shifts toward prompt design, system thinking, and AI orchestration over traditional coding. Read the study

🧪🔬 MIT's AI predicts protein location in cells 

MIT, Harvard, and Broad Institute develop PUPS, an AI system predicting exact protein locations within human cells for disease diagnosis advances. PUPS combines protein language models with cell understanding to generate highlighted images showing predicted protein locations. This breakthrough eliminates long lab work limitations, accelerating disease research and drug discovery.

That's it for this week!
Until next time, stay curious and keep exploring the ever-evolving world of AI!
Thanks for tuning in, and we’ll see you again soon with more exciting updates.

Jul