Google's Gemma 4 12B runs multimodal AI locally on 16GB laptops—but only with quantized weights and modest context windows. Full precision needs more memory. The tradeoff: privacy over speed. https://t.co/2jM4IdTb1p
Napkin AI's bet: workers need one sharp diagram before they need a full AI deck. 2M users in six months suggests the narrower play—text to editable visuals—has real traction. https://t.co/vgWkX6Cm8u
OpenAI wants Congress to federalize AI rules—then override state laws with them. The catch: its proposed safety institute would recommend, not block, deployment. https://t.co/diSea3B4WB
New Lancet study: endoscopists who relied on AI to spot precancerous growths saw their unaided detection drop from 28.4% to 22.4%. The risk isn't AI use—it's substitution over tutoring. https://t.co/hlMdsPUjXF
Claude-Mem lets you search across AI coding sessions instead of re-explaining context each time. Trade-off: you're running a local database with hooks into your editor. https://t.co/2TmIyaEdr5
Microsoft deployed its own coding model to Copilot users today, cutting the cord on OpenAI and Anthropic weights. The catch: nobody knows if it's 5 billion or 137 billion parameters. https://t.co/h39UhgRioT
Microsoft rolled out MAI-Code-1-Flash to VS Code users on June 2, a 137B-parameter code model with a 256K token window. CLI and API support delayed. https://t.co/6Nuxik6Ybc
Microsoft embedded AI agent containment into Windows itself. Agents run in isolated containers with declared file/network access, enforced by the kernel. https://t.co/BUYHcUDeHx
Trump signs voluntary AI review order: agencies get 30 days to assess frontier models before release. No mandatory oversight—that's Congress's call. #AI#Policy#Cybersecurity https://t.co/JIPji4sKgg
Odysseus hit 29K GitHub stars in 3 days as a self-hosted AI workspace. Safe first test: localhost only, dummy data, authentication on, no live mailbox. #AI#SelfHosted#DevTools https://t.co/Wyp0X0JCnR
Anthropic expands Project Glasswing to ~150 new organizations across 15+ countries. The cyber-capable Claude model now reaches ~200 vetted partners in critical… https://t.co/SF79uD6NDS
Mathematicians respond to OpenAI's private AI proof: disclose your tools, stay accountable, let peers review the work. The Leiden Declaration puts transparency rules in front of the field. https://t.co/wsfPHXMOsF
NVIDIA's RTX Spark puts 128GB unified memory and CUDA into mainstream Windows PCs this fall. The local-AI buying decision just got more complex—and cheaper isn't always the answer. https://t.co/dz7ilK9vFl
Nvidia's Nemotron 3 Ultra ranks first among US open models but scores 6 points behind China's Kimi K2.6. The real play: bundling it with tools optimized for Nvidia hardware. https://t.co/ZJBojfPb8d
MiniMax launched M3 with 1M-token context and native multimodal input. Weights arrive in 10 days; API access is live now. $0.60 per million input tokens. #AI#LLMs#Coding https://t.co/GWrvkHPj89
OpenClaw's creator joined OpenAI as model labs tighten API access. A $85-125 Raspberry Pi running the gateway is now the most durable way to own an AI assistant you… https://t.co/TzufFEpbgu
Anthropic filed for IPO ahead of OpenAI at $965B valuation, but the lead rests on unaudited revenue and a two-month-old rival figure. https://t.co/lV3T3Mm3PA
Nvidia is moving beyond graphics cards into the Windows PC processor slot with RTX Spark—a chip combining Grace CPU, Blackwell GPU, and up to 128GB unified memory. https://t.co/hvJS2KIFsS