Google kills Earth AI feature after one day over misinformation fears
Plus: DeepSeek cuts costs by 60%, Snapchat blocks AI videos, and Sam Altman backs calls to slow the industry down.
TL;DR
- Google yanked its Earth AI image generator one day after launch, citing rapid misuse to create geospatial misinformation.
- DeepMind's Gemini Robotics 2 can control robots of any shape and coordinate multiple units at once.
- DeepSeek's new Flash model matches OpenAI's GPT-5.6 Luna while costing 60 percent less per task.
- Sam Altman joined voices calling for the AI industry to slow down and prioritize safety over speed.
Models and research
DeepMind robotics surge; efficiency models emerge
New vision-language-action models enable robots at any scale, while startups prioritize efficiency and cost over raw parameters.
Google DeepMind unveiled Gemini Robotics 2, a vision-language-action model capable of controlling robots from tabletop arms to full humanoids. The system integrates image recognition, language processing, and motor control to coordinate multiple units simultaneously. Developers can join the waitlist for early access (via Google DeepMind Blog).
DeepSeek cuts costs by 60 percent. DeepSeek's new V4 Flash model scores 50 on the Artificial Analysis Intelligence Index—just one point behind OpenAI's GPT-5.6 Luna—while priced 60 percent lower per task. The model supports a one-million-token context window and open-weights under MIT license, with costs driven down partly by DeepSeek's 98 percent cache discount (via The Decoder).
Thinking Machines bets on efficiency over scale. The startup led by Mira Murati released Inkling Small, achieving near-parity with its larger predecessor on reasoning tasks while using less than a third of the parameters. The model averages 24K output tokens per task compared to 45K for DeepSeek and 78K for GPT-5.4 mini, emphasizing token efficiency and fine-tuning with user data (via The Decoder).
LLM security flaw leaves models vulnerable to attack. Researchers at MIT Technology Review argue a fundamental architectural flaw makes large language models impossible to fully secure against hacking. The finding, presented at the International Conference on Machine Learning, suggests defenders face an inherent asymmetry in protecting systems designed for openness (via MIT Technology Review).
Industry and business
Sam Altman backs calls to slow down AI development
OpenAI's CEO joins the chorus for caution while the company rolls out safety initiatives across Europe and combats scam operations.
Sam Altman said it is time for the AI industry to "pace" itself after years of pushing full speed ahead. Both OpenAI and Anthropic have publicly backed a petition calling for measured progress, signaling a shift toward restraint from sector leaders (via TechCrunch).
OpenAI launches responsible AI effort across Europe. OpenAI published updates on advancing responsible AI practices in Europe, laying groundwork for regulatory compliance and trust-building with policymakers (via OpenAI News).
OpenAI disrupts criminal scam operation. The company deployed AI to identify and disrupt a large-scale criminal scam operation, demonstrating a practical application of AI systems for threat detection and law enforcement support (via OpenAI News).
Products and tools
Google kills Earth AI feature; Snapchat bans AI videos
Platforms move swiftly to block AI slop and misinformation, signaling industry-wide pushback against low-quality machine-generated content.
Google yanked a new AI image-generation feature from Google Earth one day after launch after users deployed it to create fabricated imagery overlaid on satellite maps. Critics flagged the risk of geospatial misinformation reaching journalists and researchers who depend on the platform's visual integrity. Google stated it had "seen people sharing screenshots of generated imagery that appear to violate our policies" (via TechCrunch).
Snapchat cuts AI-generated videos from recommendation engine. Snapchat adjusted its algorithm to exclude fully AI-generated content from Spotlight eligibility, aiming to keep the platform "a place where people can discover authentic creativity from real people." The move mirrors similar steps by YouTube, LinkedIn, Meta, and Substack (via TechCrunch).
Policy and safety
Researchers flag structural security flaw in large language models
Fundamental design principles that make LLMs useful appear to make them inherently vulnerable to adversarial attack.
MIT Technology Review reported on research from the International Conference on Machine Learning arguing that large language models face an unsolvable security paradox. A fundamental architectural flaw—the very openness that makes LLMs capable—creates persistent vulnerability to hacking. Researchers contend that defenders face an inherent asymmetry: any fix that adds security reduces the model's ability to generate novel responses (via MIT Technology Review).
Hardware and compute
EU pools €30 billion for AI computing; private sector trails
Europe launches infrastructure bidding to catch up to U.S. spending, though total funding remains dwarfed by individual tech giants' data center budgets.
The European Commission opened bidding for up to seven AI computing gigafactories, pooling €10 billion in public funding with at least €20 billion in expected private investment. The 18-nation initiative includes hardware partners AMD, Nvidia, and Qualcomm. Applications close November 12, with construction starting in 2027. The €30 billion package aims to serve startups, companies, research institutions, and governments—but remains roughly one-twentieth of current annual U.S. tech spending on data centers (via The Decoder).