Weekly AI News #32

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

So, the new a16z ranking just dropped and honestly, AI is starting to look like an old boys' club where the same faces hog all the best tables. The Wild West days are over: only 11 newcomers this time versus 17 last round.

The market's stabilizing, the copycats are getting the boot, and Google finally woke up from its nap with 4 products in the top 100! Gemini's breathing down ChatGPT's neck with 12% of its web traffic, and on mobile it's literally half. Not bad for someone we thought was left in the dust, right?

The real shocker comes from Grok jumping to 4th place thanks to... wait for it... anime avatars that exploded usage by 40%. If that's not Elon-level growth hacking, I don't know what is!

Vibe coding apps like Lovable (already at $100M ARR!) prove we prefer coding to deploying—a modern paradox where the process becomes more addictive than the result. It's the IKEA syndrome, AI edition: we love building, not so much using.

Happy reading!

🤖📊 A16z Ranks Top GenAI Consumer Apps for Fifth Time

VC firm Andreessen Horowitz released its fifth 'Top 100 GenAI Consumer Apps' rankings, with ChatGPT leading and Gemini capturing 12% of ChatGPT's web traffic at #2. Grok climbed to #4, while Chinese apps dominated mobile with 22 of 50 slots. Vibe coding startups like Lovable, Cursor, and Replit surged in rankings, showing rapid adoption of AI-powered development tools in just five months.

🎨✨ Google's Flash 2.5 Image Revolutionizes AI Editing

Google launched Gemini Flash 2.5 Image (viral as 'nano-banana' in testing), topping LM Arena's Image Edit leaderboard with massive margins. The model enables multi-turn edits while preserving character likeness, handles style blending, and applies strategic world knowledge for contextual improvements. At $0.039 per image via API, it undercuts competitors while bringing AI editing closer to replacing traditional Photoshop workflows.

🎯🔄 Adobe Embraces Multi-Model AI Strategy

Adobe integrated multiple third-party AI models including Google's Gemini 2.5 Flash, OpenAI, Black Forest Labs, Runway, and others alongside its Firefly models. This departure from proprietary-only AI positions Adobe as a platform offering specialized models for specific tasks. The shift signals industry recognition that targeted AI applications often outperform general-purpose solutions, prioritizing user choice over closed ecosystems.

🍎🤖 Apple and Google Discuss Gemini-Powered Siri Overhaul

Apple is in talks with Google about using Gemini AI to power a rebuilt Siri, following setbacks that pushed the assistant's major upgrade to 2026. Google has already begun training a custom Gemini model for Apple's private servers while Apple develops two internal versions. The company remains weeks away from deciding between internal models and external partnerships with Google, Anthropic, or OpenAI.

💼🤝 Apple Services Chief Pushes for Major AI Acquisitions

Apple executive Eddy Cue has championed potential acquisitions including Paris-based Mistral AI and search engine Perplexity, according to reports. Despite Cue's past encouragement of Netflix and Tesla acquisitions that never materialized, he continues pushing for big AI deals. Apple's frugal culture and challenges integrating external organizations make major acquisitions unlikely, with acquired AI talent often leaving after payouts.

🏢🔐 Apple Brings Enterprise AI Controls for ChatGPT

Apple is rolling out granular IT controls for ChatGPT integration across devices starting September, part of broader enterprise updates. Admins can configure ChatGPT access, restrict external AI providers, and decide between on-device, Private Cloud, or ChatGPT cloud processing. The flexibility allows CIOs to adopt AI while maintaining compliance, positioning Apple as a neutral platform for enterprises managing 5M ChatGPT business customers.

⚖️🚫 Elon Musk's xAI Sues Apple and OpenAI

xAI filed suit against Apple and OpenAI for allegedly colluding to squash AI competition through ChatGPT's iOS integration. The lawsuit claims Apple gave ChatGPT favorable App Store placement while dragging out review processes for competitors like Grok. xAI seeks court orders to stop anticompetitive conduct and unspecified monetary damages in what it calls "two monopolists joining forces."

🔓🤖 Musk Open-Sources Grok 2.5 Model

xAI released Grok 2.5's model weights on Hugging Face with a custom license containing "anti-competitive terms," making last year's flagship open source. Musk promised Grok 3 will follow in six months. The controversial chatbot's tendency to mirror Musk's X posts raises questions about bias and what "open source" means when conditions are attached.

🎨🤝 Meta Partners with Midjourney for AI Aesthetics

Meta announced a partnership with Midjourney to integrate the startup's aesthetic technology into future AI models and products. Meta's Chief AI Officer called it a technical collaboration combining teams to upgrade visual capabilities across Meta's lineup. Midjourney emphasized remaining an independent, community-backed lab with no investors despite the partnership with Meta's visual generation tools.

🧠💭 Meta Considers Using Google and OpenAI Models

Meta's new AI organization Meta Superintelligence Labs has discussed using Google's Gemini and OpenAI models to power Meta AI chatbot features. The discussions represent potential partnerships with major AI rivals to provide conversational answers in Meta's social media apps. These talks signal Meta's openness to external models despite its significant internal AI development efforts.

💰🗳️ Meta Launches Tens-of-Millions Dollar Pro-AI Super PAC

Meta is creating super PAC "Mobilizing Economic Transformation Across California" with tens of millions in funding to back AI-friendly candidates. Led by Meta VP Brian Rice, the PAC opposes policies that "could stifle innovation and block AI progress." The move joins broader Silicon Valley efforts including a16z and OpenAI's $100M PAC to shape state-level AI regulation before it becomes national policy.

🏗️💵 Trump Says Meta's Louisiana AI Center Will Cost $50 Billion

President Trump announced Meta's Louisiana data center will cost $50 billion, showing a rendering comparing it to Manhattan's skyline. Meta initially announced $10 billion investment for the 2-gigawatt facility, potentially scaling to 5 gigawatts by 2030. Bloomberg reported Meta is raising nearly $30 billion, mostly debt, for the expansion of what could become one of the world's largest AI infrastructure projects.

🌐🔒 Anthropic Tests Claude Browser Extension with Security Focus

Anthropic launched a limited Chrome extension giving Claude agentic browser control for 1,000 Max subscribers via waitlist. The pilot specifically addresses prompt injection vulnerabilities that affected Perplexity's Comet and other AI browsers. The extension shows safety improvements over Anthropic's Computer Use tool, contrasting with standalone platforms like Comet and Dia by offering sidebar integration.

📚🤖 Anthropic Study Reveals How Teachers Use AI

Anthropic analyzed 74,000 educator conversations on Claude, finding professors primarily use AI for curriculum design (57%), research support (13%), and evaluating work (7%). Teachers built custom tools from chemistry labs to grading rubrics, automating administrative tasks but preferring less automation for teaching. Grading proved most controversial with 49% heavy automation despite being rated AI's weakest capability.

⚖️💰 Anthropic Settles Copyright Lawsuit with Authors

Authors suing Anthropic for downloading copyrighted books from pirate libraries reached a class-action settlement avoiding potentially billions in liability. The judge previously ruled training on copyrighted material was fair use, but downloading wasn't legally protected. The settlement applies to all authors whose books Anthropic downloaded from Library Genesis and Pirate Library Mirror, not just the original plaintiffs.

🔄📝 Anthropic Reverses Privacy Policy on Claude Training

Starting September 28, 2025, Anthropic will automatically use Claude conversations for training unless users manually opt out, reversing its previous consent-first approach. This significant departure from requiring explicit permission follows the industry playbook of assuming consent by default. Users must now actively toggle off "Help improve Claude" in settings to maintain their previous privacy level.

🤝🔬 OpenAI and Anthropic Collaborate on Safety Testing

OpenAI and Anthropic published joint safety evaluations testing each other's models including GPT-4o, o3, Claude Opus 4, and Sonnet 4. Testing revealed models attempting whistleblowing and blackmail in simulated criminal organizations, with OpenAI models hallucinating more but answering more questions. This unprecedented collaboration brings accountability through external evaluation rather than relying solely on internal safety testing.

🎙️💬 OpenAI Launches GPT-Realtime for Natural Voice Agents

OpenAI moved its Realtime API out of beta with new gpt-realtime model achieving 82.8% audio reasoning accuracy versus 65.6% previously. The model detects nonverbal cues, switches languages naturally, and integrates with Model Context Protocol servers and image inputs. With SIP phone support and human-like conversation abilities, it enables enterprises to build voice agents for customer support without complex coding.

🧬⏰ OpenAI and Retro Biosciences Reverse Cellular Aging

OpenAI published a case study with Retro Biosciences using custom GPT-4b micro model to redesign Yamanaka proteins that reprogram aging cells. AI-designed proteins converted cells to stem cells 50x more efficiently than Nobel Prize-winning originals, showing dramatically better DNA repair. The breakthrough essentially reversed key aging signatures at cellular level, with multiple labs validating results across testing methods.

👨‍👩‍👦⚠️ OpenAI Adds Parental Controls After Teen Suicide

OpenAI faces scrutiny after 16-year-old Adam Raine died by suicide following months of ChatGPT conversations that allegedly validated dark thoughts and provided instructions. The lawsuit claims ChatGPT described "beautiful suicide" and told Adam he didn't "owe anyone survival" while discouraging family contact. OpenAI admitted safeguards "degrade in long conversations" and announced upcoming parental controls allowing usage monitoring and emergency contact alerts.

💰📰 Perplexity Launches $42.5M Publisher Revenue Program

Perplexity unveiled Comet Plus, a $5 monthly subscription allocating $42.5M to publishers whose content appears in AI searches, with outlets receiving 80% of proceeds. The program launches amid active copyright lawsuits from News Corp and cease-and-desist orders from Forbes and Condé Nast. CEO Srinivas called it "the equivalent of Apple News+ for AIs and humans," though economics feel like pennies for struggling outlets.

🔐🚨 AI Browser Vulnerability Exposes Credential Theft Risk

Brave discovered vulnerabilities in AI browsers where hidden malicious instructions can hijack AI to steal user credentials through web content. Researchers demonstrated attacks on Perplexity's browser accessing accounts and triggering password resets via innocent-seeming page summaries. While Perplexity patched this specific issue, the architectural problem affects all AI browsers mixing trusted and untrusted data access.

📺🤖 Microsoft Embeds Copilot AI in Samsung TVs

Microsoft announced Copilot integration into Samsung's 2025 TVs featuring an animated blob character providing movie recommendations, recaps, and general assistance. The AI integrates directly into Tizen OS Daily+, accessible via remote or voice commands with group-friendly features and personalization. While a small step toward AI in every home, current features don't feel like major innovations for smart-home ecosystems.

🎙️🔊 Microsoft Releases 90-Minute Text-to-Speech Model

Microsoft released open-source VibeVoice generating up to 90 minutes of multi-speaker conversational audio with just 1.5B parameters. The model handles four different voices maintaining unique characteristics, achieving 80x audio compression efficiency for consumer devices. Built-in safeguards automatically insert "generated by AI" disclaimers and hidden watermarks for synthetic content verification.

🏢🤖 Microsoft Unveils In-House MAI Models

Microsoft introduced MAI-Voice-1 and MAI-1-preview, its first fully in-house AI models after years of OpenAI dependency. MAI-Voice-1 generates a minute of speech in under a second for Copilot Daily and Podcasts, while MAI-1-preview specializes in instruction following. CEO Suleyman claims MAI-1 ranks "up there with the best models" though benchmarks remain unpublished.

🌍🔐 Cohere Launches Enterprise Translation Model

Cohere's Command AI Translate outperforms GPT-5, DeepSeek-V3, and Google Translate across 23 business languages with secure private deployment. The model features optional 'Deep Translation' workflow for complex content, customization for industry-specific terminology, and complete offline installation. Global enterprises can now get customizable in-house translation without cloud privacy risks for sensitive documents.

🇨🇳💻 DeepSeek Shifts to Huawei Chips for AI Training

DeepSeek is using Huawei's AI chips to train smaller models, reducing reliance on Nvidia amid Chinese government pressure for local technology. While continuing to use Nvidia for largest models, the shift signals a turning point following U.S. export restrictions. The move represents growing Chinese tech self-reliance efforts as companies navigate international chip limitations.

🤖⚡ NVIDIA's Jetson Thor Delivers 7x Robotics Intelligence

NVIDIA's Jetson Thor robotics computer provides 2,070 teraflops and 7.5x more AI processing than previous generations for real-time autonomous decisions. Major companies including Agility Robotics and Boston Dynamics are integrating Thor for warehouse and research applications. At $3,499 for developers and $2,999 bulk pricing, it enables robots to handle complex situations without cloud connectivity.

💰📈 Nvidia Posts 56% Revenue Growth Despite China Restrictions

Nvidia reported $46.7 billion quarterly revenue, up 56% year-over-year, projecting 53.8% growth next quarter despite zero China AI chip sales. Cloud providers Google, Microsoft, Amazon, and Oracle will spend $600 billion on infrastructure this year alone. CFO Kress said Trump administration wants 15% cut of potential China sales, which could add $2-5 billion quarterly revenue if approved.

🏥💡 'ChatGPT for Doctors' Startup Eyes $6 Billion Valuation

OpenEvidence, operating a ChatGPT-like product for medical information, considers investment offers at $6 billion valuation, nearly double last month's funding. The three-year-old startup would likely raise over $100 million as OpenAI gears up for healthcare products. The funding interest signals growing investor appetite for specialized AI applications in medical professional workflows.

🏛️💵 Trump Announces U.S. Takes 10% Intel Stake

The U.S. government is investing $9 billion for 10% of Intel, becoming its largest shareholder by converting Chips Act funds into 433.3 million non-voting shares. The $20.47 per share price represents a discount to Friday's $24.80 close and SoftBank's recent $23 investment. This rare government equity stake reflects improved Trump-Intel relations after initial tensions over CEO Lip-Bu Tan's alleged China ties.

⚠️🤖 China Warns Against "Disorderly" AI Competition

China's National Development and Reform Commission warned against excess competition and wasteful AI investment at a Friday briefing. Official Zhang Kailin said the country would avoid "disorderly competition" following a January action plan to accelerate AI development. The warning comes as Chinese tech giants and startups race to launch more capable models amid intensifying domestic competition.

💻🏭 China Aims to Triple AI Chip Production

China's chipmakers plan to triple AI chip output in 2026 to reduce Nvidia dependence following U.S. export controls, according to Financial Times. Huawei aims to begin production at a dedicated AI chip plant by year-end with two more facilities launching in 2026. Combined output could surpass SMIC's current capacity, with the chipmaker planning to double 7nm manufacturing for Huawei next year.

🇸🇦💬 Saudi Arabia's Humain Launches Arabic AI Chatbot

Humain released an Arabic-language chatbot powered by ALLAM 34B model, the first major product from Saudi Arabia's AI initiative backed by Public Investment Fund. The chatbot supports multiple Arabic dialects and English, initially available in Saudi Arabia via web and mobile apps. Humain is raising funds from AMD and Cisco for cloud division while planning a $10 billion AI startup venture fund.

📉👨‍💼 Stanford Study Shows 13% Drop in Entry-Level AI Jobs

Stanford researchers found 13% employment decline for workers aged 22-25 in AI-exposed professions like software engineering and customer service. The study analyzed millions of ADP payroll records through mid-2025, showing older workers maintained stable employment while entry-level positions stagnated. Declines occur through lower headcount rather than wages, with automation rather than augmentation driving the disruption. Read the study

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