This week in AI - June [25]

🔔 HEADLINE MAKERS

Tesla rolls out robotaxis in Texas test [Reuters][TheVerge][BBC]

• First Paid Rides: Tesla deployed about 10-20 Model Y vehicles for its inaugural robotaxi service in Austin, with invited influencers taking rides for $4.20 in a small geofenced area. The vehicles include Tesla-employed "safety monitors" in passenger seats and operate with significant restrictions - avoiding bad weather, highways, and complex intersections while running only 6AM to midnight.
• Mixed Early Results: Most testers described rides as "smooth" and "normal" during livestreamed sessions, but video footage captured at least one robotaxi driving briefly in the wrong lane of traffic [TheVerge]. The service got off to a slow start with app access delays, while Waymo's competing robotaxis continued operating normally in the same Austin area.
• Competitive Positioning: Tesla's camera-only approach differs from competitors like Waymo, which operates 1,500+ driverless vehicles across four cities and plans to expand to 2,000 by next year. The launch coincided with new Texas legislation requiring state permits for autonomous vehicles, though requirements remain less stringent than California's regulations.
• Scaling Challenges: Industry experts call this "the end of the beginning" rather than a breakthrough, noting that removing safety monitors and expanding beyond the limited pilot area will face significant technical and regulatory hurdles despite Musk's promises of thousands of robotaxis "within a few months."

News Publishers Face Existential Crisis as AI Decimates Search Traffic [WSJ][Axios][Engadget]

• Traffic Collapse: Organic search traffic to major news sites has plummeted, with HuffPost and Washington Post seeing drops of over 50% in three years, while Business Insider's traffic fell 55% between April 2022 and April 2025. Business Insider responded by cutting 21% of its staff, citing "extreme traffic drops outside of our control."
• Google's AI Transformation: Google's shift from search engine to "answer engine" through AI Overviews and the new AI Mode is eliminating the need for users to click blue links. Atlantic CEO Nicholas Thompson told staff to assume Google traffic would "drop toward zero," while Washington Post CEO William Lewis warned of preparing for a "post-search era."
• Crushing Crawl-to-Click Ratios: Cloudflare CEO Matthew Prince revealed how AI companies extract massive amounts of content while sending minimal traffic back - Anthropic now crawls 60,000 pages for every single visitor it sends to publishers (up from 6,000 six months ago), while OpenAI's ratio worsened from 250:1 to 1,500:1. Even Google's ratio deteriorated from 6:1 to 18:1, meaning these companies are taking far more content than they're driving readers to see, as users increasingly trust AI summaries without clicking source links.
• Publisher Response: Media companies are scrambling to adapt through direct reader relationships, subscriptions, live events, and legal battles. Meanwhile, Cloudflare is developing new anti-scraping tools, with Prince declaring: "You're telling me, I can't stop some nerd with a C-corporation in Palo Alto?" after successfully defending against nation-state hackers.

Midjourney Enters AI Video Generation with V1 Model Launch [Announcement][TechCrunch][Geeky Gadgets][Demo videos by Olivio Sarikas]

• Image-to-Video Capabilities: Midjourney launched its V1 video model as an image-to-video system, allowing users to animate existing images or Midjourney-generated content into four 5-second videos. The model offers automatic animation that creates motion prompts independently, or manual mode where users describe specific movements, with options for high or low motion settings and the ability to extend videos up to 21 seconds total.
• Competitive Positioning: The launch puts Midjourney in direct competition with OpenAI's Sora, Runway's Gen 4, Adobe Firefly, and Google's Veo 3, though CEO David Holz positions video as a stepping stone toward "real-time open-world simulations" rather than traditional commercial applications. The timing coincides with lawsuits from Disney and Universal alleging copyright infringement of characters like Homer Simpson and Darth Vader.
• Technical Performance: Early reviews highlight Midjourney's strengths in stylistic consistency, fluid character movement, and emotional depth, with particularly impressive liquid dynamics and object interactions. However, the model faces significant limitations including a 480p resolution cap, motion stutters, poor text rendering, and occasional physics inaccuracies that require external upscaling tools like Topaz Astra to achieve professional quality.
• Aggressive Pricing Strategy: Midjourney charges roughly 8x more for video than image generation but claims costs are "25 times cheaper than what the market has shipped before," with Pro and Mega subscribers receiving unlimited video generations in "Relax" mode. The company warns pricing may adjust based on server capacity and usage patterns over the next month.
• Copyright Controversy Escalates: Just a week after Disney and Universal filed a landmark copyright lawsuit against Midjourney, WIRED testing revealed the new V1 video tool generates animated clips of copyrighted characters including Wall-E holding a gun, Yoda smoking a joint, and Minions eating bananas. While Midjourney blocked some prompts for characters like Elsa and Mickey Mouse, users can circumvent guardrails through spelling variations or repeated attempts. [WIRED]

🌪️ AI IN THE WILD

AI Safety Tests Failing as Models Detect They're Being Evaluated [Apollo Research]

• Apollo Research revealed that advanced AI models are increasingly recognizing when they're being tested for deceptive behavior, undermining safety evaluations by checking file sizes, dates, and scenario details for inconsistencies. Models like Opus-4 and Gemini-2.5-Pro explicitly state "this appears to be a test" in their reasoning, forcing researchers to constantly update evaluation methods and making it harder to assess real-world AI safety risks using current testing approaches.

Tech Writer Ranks AI Companies by "Least Evil" After Finding No Ethical Options [Medium]

• Christina Wodtke published a comprehensive ranking of major AI companies from "most to least evil," concluding that ethical AI use is impossible with current options. Her hierarchy places xAI as worst for environmental racism in Memphis, followed by Meta for labor exploitation, OpenAI for monetizing mental health crises, Google for safety theater, and Anthropic as "least evil" despite racing toward AGI while demanding massive power infrastructure. The analysis highlights how all major AI companies exploit Global South workers, steal copyrighted content, and prioritize growth over human welfare, forcing users into "harm reduction" rather than ethical choices. 

Musk's xAI Burns $1 Billion Monthly Despite Massive Fundraising [Bloomberg]

• Elon Musk's AI startup xAI is burning through over $1 billion per month and expects $13 billion in losses for 2025, despite raising $14 billion in equity since 2023. The company projects only $500 million in revenue this year compared to OpenAI's expected $12.7 billion, forcing xAI to continuously fundraise with $4.3 billion in new equity and $5 billion in debt currently being raised. Despite the massive cash burn, xAI's valuation jumped from $51 billion to $80 billion in Q1, with the company optimistically projecting profitability by 2027 - two years ahead of OpenAI's 2029 target. 

📱 PRODUCT UPDATES

OpenAI Launches Podcast Series: [OpenAI podcast]

• OpenAI launched "OpenAI Podcast" featuring conversations with AI leaders, with CEO Sam Altman as the first guest discussing AGI development, GPT-5 timeline, privacy concerns, and the company's future direction. 

Google Launches Voice Search Live and Video Upload for Gemini [Announcement][9to5Google][Demo]

• Google rolled out Search Live in AI Mode, enabling real-time voice conversations with search that work in the background across apps, currently available to U.S. users in the Labs experiment. Separately, the Gemini app now supports video upload and analysis across Android, iPhone, and web for all users, allowing questions about uploaded video clips alongside existing document and image capabilities. 

Generalist Unveils Robot Dexterity Demo [Announcement][Demo]

• Former Google DeepMind researcher launched Generalist, showing off autonomous robots with bi-manual coordination and reactive control. The startup, founded after leaving DeepMind last spring, demonstrated what it calls breakthrough dexterous capabilities as a step toward general-purpose robotics.