Alphabet raises $85B for Google’s AI business: Record-breaking stock sale — another $40B planned next quarter. Alphabet posted $110B in Q1 revenue alone (22% YoY growth). Signals strong investor appetite for Anthropic’s expected ~$1T IPO, SpaceX IPO, and OpenAI waiting in the wings. Goldman Sachs CEO calls market sentiment “greed mode” for AI
출처: TechCrunch · CNBC
Google Gemma 4 12B: encoder-free multimodal that runs on laptops: A unified architecture that processes images and audio directly — no separate encoders needed. Runs on consumer laptops with 16GB RAM. Near 26B MoE performance at less than half the memory. A watershed for on-device multimodal AI
출처: Google Blog
Meta launches WhatsApp Business AI agent globally: After two years of testing in India and Mexico, the AI customer support bot is now available worldwide on WhatsApp and Instagram DMs. Handles inquiries, product recommendations, bookings, and lead qualification. Meta plans to charge via WhatsApp Business Premium subscription. An attempt to redefine WhatsApp as workflow software for SMBs
출처: TechCrunch
Google allows publishers to opt out of AI Search: Regulatory pressure forces opt-out options for AI Overviews, AI Mode, and Discover. Testing with UK publishers first, then global rollout. Traditional search rankings unaffected. Google notes AI Overviews has 2.5B MAU and AI Mode has surpassed 1B MAU — transparently signaling what publishers would be leaving
출처: TechCrunch
U of Toronto demonstrates AI worm using open-weight models: Researchers built autonomously propagating malware powered by freely downloadable AI models, capable of targeting any online device. Existing defenses are not ready. Findings were shared with national security bodies before publication — a deliberate warning about the risks of unregulated open-weight models
출처: Utoronto
Coralogix: Raised $200M to build the monitoring layer for AI agents. 5,000+ customers, $100M+ ARR, 60% YoY revenue growth. Over half of enterprise customers now use AI agents to query the platform — dashboards are fading, “agents watching agents” is becoming the norm
출처: TechCrunch
AethexAI: Ex-Goldman and Meta founders building voice AI specifically for Africa and the Middle East. Tiny models (300M–1.7B params, Kora series) to minimize latency, 17,000+ calls/day. Ships hard drives to radio stations for data collection, recruits university students for dialect annotation. Fundamentally different from ElevenLabs/Sierra which were built for Western markets
출처: TechCrunch
Google Dreambeans: New AI app that curates personalized “stories” (places, events, things to try) from your Google data — Gmail, Calendar, Photos, YouTube, Search history. Processes overnight, delivers in the morning. Privacy-forward design emphasized
출처: TechCrunch
Amazon shows AI-generated product images in search: Displays AI-made fake images to guide shopping searches instead of real product photos. Consumer confusion and trust concerns immediately raised
출처: TechCrunch
Stanford Law study: AI outperformed law professors on legal tasks (387 points on HN)
출처: Law
DDR5 RAM price surge: 32GB DDR5 now costs minimum $375 as AI demand squeezes the PC building market
출처: Tomshardware
YC P26: Hyper (“company brain” for agentic development), Rudus (AI for concrete contractors)
AI agent observability becomes mandatory infrastructure: Coralogix’s data shows over half of enterprise customers already query monitoring tools through AI agents. A recursive “agents watching agents” pattern is forming. Once this becomes standard, agents without observability will simply be undeployable — just as Datadog and New Relic became essential for cloud, AI-native monitoring will be a massive market
Laptop-grade multimodal is now real: Gemma 4 12B proves you don’t need cloud GPUs for capable multimodal AI. Encoder-free architecture that processes audio and vision directly, on 16GB consumer laptops. On-device AI is no longer limited to lightweight text models — expect an explosion of privacy-focused, offline-capable AI applications
AI search is becoming a regulated utility: Google’s publisher opt-out, combined with DuckDuckGo’s “no-AI” search surge, signals that AI search can no longer operate unilaterally. EU/UK regulation is likely to become a global template, reshaping the publisher-platform relationship
Voice AI’s real frontier is non-Western markets: While ElevenLabs and Sierra fight over English-language enterprise, markets where call volume is 3x the West — Africa, the Middle East — remain largely unserved. The pattern (small region-specific models + local data collection + on-the-ground partnerships) is replicable across Southeast Asia, Latin America, and Eastern Europe
AI agent observability for the mid-market gap: Coralogix is going enterprise with $200M. Between Coralogix and DIY, startups and mid-size companies deploying agents need affordable, lightweight monitoring — an open-source self-hosted option or usage-based low-cost SaaS
Voice AI for underserved language markets: AethexAI’s playbook is replicable. Korean dialects, Southeast Asian languages, Eastern European markets — wherever major voice AI platforms underperform due to training data bias, small models + local data collection is a viable entry strategy
AI search governance tooling: As Google’s opt-out goes global, publishers will need to manage their AI search presence across multiple platforms (Google, Bing, Perplexity, etc.). Analytics, opt-out management, and impact measurement bundled as a SaaS product
Is Google’s $85B raise a peak signal or just the beginning? Smart timing for Alphabet (raising when stock is high), but it could also be absorbing demand that would otherwise flow to Anthropic/OpenAI IPOs
Will publisher opt-outs actually matter? If major publishers exit en masse, does Google’s AI search product degrade, or is this a regulatory checkbox that changes nothing in practice?
AI worm ripple effects: U of Toronto’s demonstration makes the risks of open-weight models concrete. How fast do defenses evolve, and does this accelerate AI model regulation?
The recursive “agents watching agents” endpoint: When AI monitors AI that monitors AI, at what point does the chain become too brittle for human intervention — and who is accountable when it breaks?