14th October - AI News Daily - Qwen3-VL Tops Multimodal Charts, Runs 80 Tokens Per Second on Apple Silicon

14/10/2025 12 min Temporada 1 Episodio 117
14th October - AI News Daily - Qwen3-VL Tops Multimodal Charts, Runs 80 Tokens Per Second on Apple Silicon

Listen "14th October - AI News Daily - Qwen3-VL Tops Multimodal Charts, Runs 80 Tokens Per Second on Apple Silicon"

Episode Synopsis

Send us a text🌍 INAI • The Open AI HubThe Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge. 2 Million+ tools, models, agents, tutorials & daily news—free for all, updated every day.https://github.com/inai-sandy/inAI-wikiKey Infrastructure & Partnerships: OpenAI partners with Broadcom and AMD to build 10 GW AI data centers, intensifying competition with Nvidia. AMD shares hit all-time highs following the deal.Enterprise AI Adoption: Google launches Gemini Enterprise across business and HR workflows, with early adopters like HCA Healthcare and Best Buy reporting productivity gains. Salesforce commits up to $15B to AI and transforms Slack into an agent-first collaboration hub for enterprise automation. LSEG and Microsoft combine financial datasets with AI agents for real-time analytics.Model Performance: Qwen3-VL leads multimodal leaderboards and achieves ~80 tokens/sec on Apple silicon. Apriel-1.5-15B-Thinker reaches frontier-level AIME'25 math accuracy on a single GPU. Ling/Ring-1T introduce trillion-parameter open models with near-IMO "silver" reasoning. Google Gemini 2.5 sets audio-reasoning record at 92% on Big Bench Audio. Microsoft MAI-Image-1 enters LMArena's top 10.New Tools: n8n launches natural-language builder for agents and automations. Autodesk WaLa converts sketches to 3D assets instantly. Suno "AI instrument" transforms hummed ideas into full songs. Microsoft MarkItDown converts documents to Markdown with OCR. BigCodeArena debuts human-in-the-loop code generation benchmark. Cleanlab ships 3-line dataset cleaning.Research Advances: Webscale-RL scales reinforcement learning to pretraining magnitudes. Adaptive speculators cut training time by predicting next-token branches efficiently. Hunyuan's RL approach delivers cheaper reasoning boosts. RTEB introduces real-world retrieval benchmark. Larger token vocabularies improve transformer training dynamics.Security & Policy: Deepfake misuse and "Shadow AI" escalate; Microsoft reports 71% of UK employees use unauthorized AI tools. Okta ships agent security credentials and controls. OpenAI debuts political bias detection framework and wins ruling easing data-retention obligations. Google offers EMEA students free year of AI Pro. TCS cuts 20,000 roles amid AI-driven restructuring.Learning & Community: Andrej Karpathy's nanochat tutorial covers building ChatGPT-style systems end-to-end. MLX guide fine-tunes Qwen3-0.6B on MacBook in under two minutes. Discussions warn against synthetic data overuse triggering model collapseand debate model training moats versus product advantages.Support the show

More episodes of the podcast AI News Daily