📰 AI News Daily — 14 Oct 2025

TL;DR (Top 5 Highlights)

OpenAI partners with Broadcom and AMD to build 10 GW of AI data centers, signaling a new phase of AI infrastructure scale and intensifying chip competition with Nvidia.
Google launches Gemini Enterprise across business and HR, upgrades Home/Nest with conversational AI, and gifts students a free year of advanced AI—broadening adoption and lock-in.
Salesforce commits $15B to AI and turns Slack into an agentic collaboration hub, pushing enterprises toward automated workflows and measurable productivity.
Qwen3-VL tops multimodal leaderboards and runs at 80 tokens/sec on Apple silicon, while Apple quietly ships on-device voice and segmentation models—strong local AI momentum.
Deepfake misuse and “Shadow AI” escalate; Okta ships agent security, OpenAI debuts political bias metrics, and regulators face mounting pressure to address safety and trust.

n8n launches a natural-language builder for agents and automations, letting teams orchestrate workflows without complex scripting—speeding prototype-to-production cycles for internal tools and ops.
Autodesk WaLa turns rough sketches into 3D assets in seconds, compressing creative pipelines for designers and game studios while cutting iteration costs.
Suno “AI instrument” converts hummed or sung ideas into full songs, opening professional-grade music creation to non-musicians and accelerating content workflows.
Microsoft MarkItDown converts PDFs, Office docs, HTML, and images to clean Markdown with OCR, simplifying data ingestion for RAG, search, and documentation pipelines.
BigCodeArena debuts a human-in-the-loop, instant-execution benchmark for code generation, enabling fairer, faster model comparisons that reflect real developer needs.
Cleanlab ships “3-line” dataset cleaning for tabular and ML tasks, improving model accuracy and reliability without heavy data engineering.

Qwen3-VL leads image-processing traffic on OpenRouter and hits ~80 tokens/sec locally on Apple silicon, showing open models can deliver speed, cost, and capability.
Apriel-1.5-15B-Thinker reaches frontier-level AIME’25 math accuracy on a single GPU, underscoring how compact models plus reasoning training can rival much larger systems.
Ling/Ring-1T introduce trillion-parameter open models claiming near-IMO “silver” reasoning, signaling rapid open-source progress in long-form, step-by-step problem solving.
Google Gemini 2.5 sets an audio-reasoning record (92% on Big Bench Audio) but lags peers on latency, highlighting ongoing trade-offs between capability and responsiveness.
Microsoft MAI-Image-1 enters LMArena’s top 10 ahead of commercial release, suggesting competitive image understanding from a major cloud provider.

Webscale-RL scales reinforcement learning to pretraining magnitudes, hinting at path-breaking gains in reasoning and tool use when RL is integrated earlier in training.
Adaptive speculators report large training-time cuts by predicting next-token branches more efficiently, reducing compute costs without sacrificing model quality.
Hunyuan’s RL approach delivers cheaper reasoning boosts, showing budget-friendly fine-tuning can unlock step-by-step performance gains in smaller models.
RTEB introduces a retrieval benchmark grounded in real-world tasks, improving evaluation of search, RAG, and enterprise knowledge systems.
Studies find larger token vocabularies consistently improve transformer training dynamics, offering a straightforward lever for stability and convergence.
New tests for spatial reasoning (latent-shape rotation) and syntheses of agent safety risks arrive as multi-agent planners and autonomous systems proliferate.

OpenAI × Broadcom co-develop custom accelerators toward 10 GW of capacity; parallel AMD deals send AMD shares to all-time highs, intensifying the AI chip race with Nvidia.
Google Gemini Enterprise targets workflows and HR with automation and deep integrations; early adopters like HCA Healthcare and Best Buy report productivity gains.
Salesforce pledges up to $15B for AI, and upgrades Slack into an agent-first collaboration hub—centralizing knowledge, analytics, and workflow automation for enterprises.
Microsoft flags “Shadow AI” risk as 71% of UK employees use unauthorized tools; Okta counters with credentials and controls for AI agents to protect identity and compliance.
OpenAI wins a ruling easing ChatGPT data-retention obligations and unveils a framework to detect political bias, aiming to bolster transparency and trust.
Google offers EMEA university students a free year of AI Pro (Gemini 2.5 Pro, NotebookLM, 2 TB storage), expanding grassroots access to advanced AI capabilities.
LSEG × Microsoft pair financial datasets with AI agents in cloud tools, promising real-time analytics and automated trading for capital markets.
Workforce shifts accelerate: TCS cuts 20,000 roles amid AI-driven restructuring, highlighting skill transitions and automation’s impact on global IT services.

Andrej Karpathy’s nanochat walks through building a ChatGPT-style system end-to-end, clarifying how pretraining choices affect fine-tuning and RL—great for hands-on learners.
A practical MLX guide fine-tunes Qwen3-0.6B on a MacBook in under two minutes, showcasing how accessible local training has become on Apple hardware.
The How I AI podcast distills evaluation best practices, helping practitioners design tests that track real task success rather than proxy benchmarks.
Deep dives cover NVIDIA GPU matmul performance and 2025’s top AI video generators, offering actionable insights for optimizing inference and creative pipelines.

Musicians use Suno to turn vocal sketches into full tracks, while Glif agents auto-generate lyric videos—compressing concept-to-publish cycles for creators.
Designers jump from scribbles to polished 3D using Autodesk WaLa, and Luma Ray3 brings frame-accurate, scribble-guided control for cinematic edits and game scenes.
Instant4D reconstructs detailed 4D scenes in minutes, hinting at real-time digital twins and rapid previsualization for film, robotics, and AR.
Insurers demo seconds-fast, agentic claim handling with MarvelXAI, showcasing how AI can reduce fraud, speed payouts, and improve customer satisfaction.

Experts warn heavy synthetic data can trigger model collapse; balanced training with real-world data is urged to preserve diversity and truthfulness.
Debate intensifies over “model training moats”—many argue expensive bespoke pretraining often underperforms focused product, data, and distribution advantages.
A push for AI-assisted, computer-verified proofs aims to democratize advanced math, while cautioning against hallucinations in high-stakes reasoning.
New RL findings suggest size-dependent emergence and diminishing returns, reinforcing that training timing and method selection matter for cost-effective gains.

Curated from 250+ RSS feeds, Twitter expert lists, Reddit, and Hacker News.