GSV's AI News & Updates (04/18/25)
OpenAI Social Network, Claude Voice Debut, GPT 4.1, Agent2Agent Protocol, China's AI Education Plan, ID Verification for Models, Shopify Memo, AI Student Success Predictions, The Future of Search
General 🚀
OpenAI is building a social network: OpenAI is exploring a ChatGPT-linked social network, with early prototypes featuring image generation and a shared feed — potentially setting the stage for a new rivalry with X and Meta. A social layer would give OpenAI access to real-time user data, a critical edge in model training.
A lot of OpenAI model releases + a new developer tool:
GPT 4.1 (in standard, mini, and nano): OpenAI’s new flagship expands context windows to a million tokens, improves reliability across modalities, and undercuts prior models on cost (landing just as GPT-5 gets officially delayed).
o3 and o4-mini: A leap forward in tool-using, image-understanding AI. With o3 positioned as the most powerful reasoning model to date and o4-mini optimized for efficiency at scale, both models natively blend visual and textual inputs, execute multi-step tasks, and self-select tools like code, search, and image analysis.
Codex CLI: A lightweight, open-source agent that connects its models directly to your terminal, letting them write, edit, and act on code locally.
OpenAI in Talks to Acquire Windsurf for $3B: OpenAI is negotiating its largest acquisition yet. The deal would deepen OpenAI’s push into the developer market and signal a wave of consolidation as the AI tooling race heats up.
From Google’s Cloud Next conference:
Google’s Gemini 2.5 Flash: a new AI model optimized for speed, cost-efficiency, and real-time use cases like customer service and document parsing. Unlike its more expensive siblings, Flash is built for scale.
Agent2Agent Protocol for AI Agent Interoperability: This open interoperability framework enables AI agents from diverse platforms and vendors to communicate and coordinate tasks seamlessly.
Google unveiled Firebase Studio: a browser-based, AI-powered platform designed to enable users to build, deploy, and manage custom applications rapidly. “It’s like lovable+cursor+replit+bolt+windsurf all in one testing catalog.”
Google to embrace Anthropic’s standard for connecting AI models to data: MCP is an open standard that enables AI models to connect with various data sources and applications, facilitating tasks such as accessing business tools, content repositories, and app development environments.
Claude Comes to Google Workspace + Launches Research Mode: Anthropic is integrating Claude into Gmail, Calendar, Docs, and even Drive—turning the AI assistant into a powerful workplace companion that understands your workflow in context. This move brings Claude closer to being an enterprise-grade agent, alongside a new “Research Mode” that mirrors OpenAI’s Deep Research.
Anthropic Is Readying a Voice Assistant Feature to Rival OpenAI: Its first step into AI voice assistants—with early versions including three personalities: Airy, Mellow, and Buttery. As voice becomes the next battleground for AI assistants, Claude is moving closer to being a full-stack productivity agent.
Perplexity in Talks to Power AI Assistants on Samsung and Motorola Phones: The partnerships could position Perplexity as a viable alternative to Google’s Gemini across Android devices—just as the startup eyes a raise of up to $1B at a potential $18B valuation.
Canva Expands with AI Assistant, Coding, and Sheets: The platform now also offers Canva Code, allowing users to create interactive applications like maps and calculators through prompts. Additionally, Canva Sheets has been introduced, providing spreadsheet functionalities with AI-powered insights and chart generation.
Microsoft Enables Copilot Studio to Use Your Computer Like a Human: Microsoft’s new “computer use” feature lets Copilot Studio agents interact with websites and desktop apps directly—clicking buttons, filling forms, and navigating screens without APIs.
Education and the Future of Work 📚
The Real Reason Why Students Are Using AI to Avoid Learning: The core issue isn't AI itself but the broader digital environment, particularly social media platforms like TikTok and Instagram, which have conditioned users for instant gratification. This environment fosters a preference for quick solutions, making students more inclined to use AI as a shortcut rather than engaging deeply with challenging tasks.
Teachers Worry About Students Using A.I. But They Love It for Themselves.: Educators are increasingly using generative A.I. in their own work, even as they express profound hesitation about the ethics of student use.
Utah Shares Plan for Statewide AI Education: More than half of public districts in the state now have an AI policy in place. Last August, Utah launched one of the country’s first statewide RFP processes for school AI tools.
China’s AI-Powered Education Overhaul: In a sweeping reform bid, the Ministry of Education announced this week that AI will be central to cultivating what it calls “core competitiveness” in young talent, focusing on skills like problem-solving, independent thinking, and collaboration.
China's Yuanfudao unveils next-gen AI learning device: The Xiaoyuan AI learning machine is a hybrid tablet and robotic study companion designed to personalize education. Building on the success of its previous model, which sold over a million units, the new device features a smart dock that adjusts for scanning, prints exercises, and helps detect conceptual misunderstandings through advanced diagnostics.
Community Colleges Battle Wave of ‘Bot’ Students Exploiting Financial Aid: Fake student enrollments, often powered by AI-generated applications and coursework, are flooding community college classes to fraudulently access aid. At Southwestern College, professors say they’re spending more time spotting bots than teaching, with some classes dropping from full to nearly empty once fakes are removed.
Using AI to predict student success in higher education: A new Brookings report finds that Black and Hispanic students are significantly more likely to be wrongly predicted to fail, even when they go on to graduate. These false negatives stem from entrenched biases in the data and model design.
Google Offers Free AI Premium Plan to U.S. College Students Until Mid-2026: Unlocking tools like Gemini Advanced, NotebookLM Plus, and Veo 2—through June 2026. The move positions Google alongside OpenAI and Anthropic in the battle to capture the student market early.
Shopify CEO says staffers need to prove jobs can’t be done by AI before asking for more headcount: In a recent internal memo titled "AI usage is now a baseline expectation," Lütke stated that before requesting additional headcount or resources, teams must demonstrate that AI cannot accomplish the tasks at hand.
Venezuelan migrants relied on clickwork to survive. Now AI is replacing them: As annotation jobs dry up and freelance rates crash, workers find themselves stuck between outdated infrastructure and an increasingly inaccessible future of AI labor.
AI Is Beating Doctors at Empathy: As healthcare systems strain under time and staffing pressures, patients are turning to AI not just for information, but for compassion. Tools like ChatGPT are offering emotionally intelligent, accessible responses—sometimes more humane than those from overburdened physicians.
From Golf Swings to Surgery: New AI Can Watch, Reason, and Coach: Video-STaR teaches itself to understand and reason about video content—without human-annotated labels. In tests, it could analyze a diving routine, score difficulty levels like a human judge, and offer nuanced feedback.
Startups
Ilya Sutskever’s New AI Startup SSI Hits $32B Valuation With No Product: SSI is betting on a new path to superintelligent AI—one Sutskever calls a “different mountain to climb.” The deal reflects continued investor fervor for elite AI founding teams, even amid broader market uncertainty.
Mira Murati’s AI startup is reportedly aiming for a massive $2B seed round: With no product in market, the valuation rests on its talent pool: former OpenAI stars and early visionaries in AI.
Tech 💻
The rise of AI ‘reasoning’ models is making benchmarking more expensive: Independent researchers are now spending thousands just to verify what companies claim in blog posts. And that’s assuming they can even get access. With high token counts, pricey model usage fees, and opaque access policies, benchmarking is fast becoming a pay-to-play game.
Cohere Releases Embed 4 for Multimodal, 200-Page RAG Search: Designed for messy, real-world data like scans and handwriting, Embed 4 targets finance, healthcare, and manufacturing—and aims to become the retrieval backbone for enterprise AI agents.
The Business Case for (and Against) Giant Context Windows: While massive prompts enable full-document analysis, studies show diminishing returns beyond 32K tokens, raising questions about latency, cost, and long-range recall. For most companies, hybrid approaches blending retrieval and large prompts may offer the best ROI as the context window arms race matures.
An excellent deep dive into what tariffs mean for AI infrastructure
Safety and Regulation ⚖️
Google DeepMind Is Hiring a 'Post-AGI' Research Scientist: Google is preparing for a future with AGI, ASI, and machine consciousness. The listing covers everything from economics and education to sentient machines.
Ireland probes Musk’s X for feeding Europeans’ data to its AI model Grok: The investigation revives GDPR concerns and highlights growing EU-U.S. friction over AI and data governance.
Apple Will Analyze On-Device User Data to Improve AI—Without Sending It to the Cloud: Apple will begin comparing synthetic training data against users’ real emails—entirely on-device. The new method aims to preserve privacy while helping Apple close the gap with OpenAI and Google, whose models are trained on broader datasets.
Wikipedia Offers AI-Ready Dataset to Curb Scraper Bots: The move aims to provide developers with a cleaner, sanctioned alternative to raw text scraping—while keeping Wikipedia’s open-access ethos intact.
OpenAI May Require ID Verification for Access to Future Models: OpenAI is implementing "Verified Organization" status that may become mandatory for accessing advanced APIs. The move aims to curb misuse, enhance security, and guard against data exfiltration after reports of API abuse linked to foreign actors.
Police Are Using AI to Go Undercover Online: In Arizona and beyond, law enforcement agencies are quietly testing AI-powered bots that pose as activists, children, and sex workers online—collecting data and engaging targets via text, social media, and messaging apps. The company behind it, Massive Blue, markets the tool as a crime-fighting breakthrough.
Other
Google’s AI Overviews Chief on the Future of Search: As search shifts from keyword-matching to conversational, multimodal, and personalized discovery, Google is betting on factuality, transparency, and human curiosity as its long-term differentiators.
India Backs AI Startups to Close the Global Gap: Modi’s government is pushing to scale homegrown AI capabilities—offering GPUs and $1.2B in funding to build Indian LLMs trained on local languages. But while public efforts ramp up, private sector investment still lags behind global peers.
Google AI is helping decode dolphin communication: Built on decades of underwater fieldwork and optimized for mobile deployment, it’s the closest we’ve come to translating dolphin society.