Runway just got caught in a major training data leak after 404 Media obtained an internal spreadsheet showing the company scraped thousands of YouTube channels and pirated films to train its flagship Gen-3 video model. Why it matters: Despite raising $141M from Google and Nvidia while dodging questions about data provenance, Runway now faces intense scrutiny as AI copyright tensions boil over.
Microsoft just started testing a radical Bing redesign that puts AI-generated answers front and center while banishing traditional blue links to a skinny right sidebar. The new layout turns search into multi-section deep dives with charts and videos. Microsoft claims it maintains publisher clicks, but it's currently limited to a handful of queries.
Reddit is now blocking web crawlers and search engines from indexing its recent posts unless they pay up. Google is currently the only major engine surfacing fresh Reddit results, following its $60 million AI training deal, while Bing and DuckDuckGo are left in the cold after Reddit updated its robots.txt file.
Mistral AI just dropped Mistral Large 2, a 123B-parameter flagship model built with a 128k context window designed to run efficiently on a single node. It delivers major gains in coding and math that rival GPT-4o and Claude 3 Opus, while actively curbing hallucinations. Non-commercial weights are free to download, offering serious frontier power without giant infrastructure.
The Verge's Decoder podcast just dug into the uncomfortable reality of AI's massive power appetite. While tech giants like Microsoft and Google blow past their emissions targets to pack data centers with GPUs, high energy use is already standard across video streaming and gaming. The real debate is how we decide if AI is actually worth the juice.
California lawmakers are advancing a controversial AI safety bill that would force tech giants to test models for catastrophic risks—like bioweapons or grid shutdowns—before public release. Tech heavyweights including Google and Meta strongly oppose the measure, arguing its mandates are technically unfeasible and could stifle innovation. Why it matters: With Republicans threatening to roll back federal AI rules, California is positioning itself as the battleground for tech regulation.
A deep dive by The Markup found that widely cited benchmarks like MMLU and HellaSwag—heavily touted by Google and Meta—are fundamentally broken. Many tests rely on outdated datasets or amateur trivia that fail to test actual reasoning or safety. Why it matters: Bragging rights on leaderboards don't equal real-world reliability, especially in sensitive areas like medicine.
Google just updated Gemini on Android to answer general questions—like checking the weather—straight from the lockscreen without requiring an unlock. Previously, asking anything beyond basic alarms or media controls prompted an annoying biometric check. A new toggle in preferences makes the hands-free voice experience actually usable.
Google DeepMind just hooked up Gemini 1.5 Pro to its discontinued Everyday Robots hardware, turning them into autonomous office navigators. The bot can process spoken commands and even follow hand-drawn whiteboard maps across a 9,000-square-foot office, notching a 90% success rate across 50 employee interactions.
Hugging Face just brought Google Cloud TPU v5e chips to its Inference Endpoints and Spaces, starting at $1.38 an hour. Powered by the open-source Optimum TPU library, developers can deploy and serve open models like Gemma, Llama, and Mistral in a few clicks. It gives builders a viable, budget-friendly alternative to scarce Nvidia GPUs.
Google researchers just published a paper warning that generative AI is flooding the web with fake images and videos to sway public opinion and run scams. The team found the barrage of synthetic slop is eroding trust in digital information while giving public figures an easy excuse to dismiss real evidence as AI-generated. Why it matters: It is an ironic admission from a company rushing to bake AI into everything.
Google just made it easier for political campaigns to disclose AI-generated content. Instead of requiring advertisers to insert their own notices, Google will now automatically slap an in-ad disclosure on YouTube Shorts, feeds, and in-stream video whenever an advertiser checks the "altered or synthetic" box. Other ad formats will still require manual disclaimers.
OpenAI co-founder Andrej Karpathy took the stage at UC Berkeley, framing modern LLMs as a brand-new operating system akin to 1980s personal computing. He revealed that GPT-4 began as an unglamorous internal Reddit chatbot, urging young developers to embrace the 10,000-hour rule and let tiny hackathon side projects snowball into real breakthroughs.
YouTube just updated its privacy policy to allow people to request the removal of AI-generated content that simulates their face or voice. Victims must file a first-party complaint under privacy violations, giving uploaders 48 hours to remove or blur the footage before platform review kicks in. Why it matters: It treats unauthorized likeness cloning directly as a personal privacy breach rather than broad misinformation.
Apple is gearing up to announce a deal bringing Google Gemini to its devices as soon as this fall, according to Bloomberg's Mark Gurman. The move adds another heavyweight alongside ChatGPT, with Anthropic also in the mix. Why it matters: bringing in third-party chatbots buys Apple time to build Apple Intelligence while taking a cut of partner subscription sign-ups.
Google just officially released Gemma 2, introducing ultra-efficient 9B and 27B open models for developers. The 27B variant punches above its weight by matching models twice its size while running full precision on a single Nvidia H100, and the 9B edges out Llama 3 8B. It is available now under a commercial-friendly license.
YouTube is reportedly offering Universal, Sony, and Warner lump sums of cash to legally train upcoming AI tools on songs from dozens of artists. Google is taking a strict permission-first approach to avoid the copyright battles currently hitting competitors like Suno and Udio.
Google just rolled out 110 new languages to Google Translate powered by its PaLM 2 model, marking its biggest expansion ever. The update opens up translation to more than 614 million speakers—around 8% of the global population—spanning from widely requested tongues like Cantonese to reviving Indigenous languages like Manx. PaLM 2 makes it far easier for models to learn closely related dialects without abundant standalone training data.
Google just opened up Gemini 1.5 Pro's 2-million-token context window to all developers, removing the waitlist. Both 1.5 Pro and Flash now also feature built-in Python code execution in an offline sandbox to tackle math and data reasoning, alongside Gemma 2 landing in Google AI Studio. Why it matters: developers can now feed entire codebases into Gemini while letting the model verify its own math by running code.
Google just opened up Gemini to teen students using Google Workspace for Education accounts across more than 100 countries, completely free of charge. The teen experience comes with built-in guardrails, automatic fact-checking, and zero chat data used for model training, while teachers get new AI lesson-planning tools in Classroom. Schools finally get a secure, supervised way to introduce GenAI to minors.
