📰 AI News Daily — 04 Oct 2025

TL;DR (Top 5 Highlights)

OpenAI hits $500B valuation, eclipsing SpaceX; Sora 2’s viral rise sparks deepfake and copyright concerns.
NVIDIA crosses $4T, underscoring GPU dominance as enterprise AI spending surges.
Google ships Gemini 2.5 Flash Image and “nano banana,” pushing fast, flexible creative AI to the mainstream.
Microsoft retools Sentinel with agentic AI and a unified data lake, tightening enterprise threat detection.
Perplexity launches Comet, a free AI browser targeting research-focused browsing and Chrome’s dominance.

Perplexity Comet launched a free AI browser for research and productivity, challenging Chrome by integrating answers, citations, and personalization directly into the browsing experience.
Microsoft 365 Premium and an open-source Agent Framework for .NET and Python enable multi-agent workflows that automate documents, data tasks, and approvals across enterprise systems.
Salesforce MuleSoft Agent Fabric centralizes orchestration, governance, and integration of AI agents, helping enterprises prevent agent sprawl while securing data access and audit trails across applications.
Chrome DevTools added the Model Context Protocol, letting AI assistants inspect pages, debug issues, and simulate user behavior—speeding web troubleshooting without tedious reproduction steps.
Google Jules gained a CLI and API, letting teams embed the coding agent into workflows, automate repetitive tasks, and customize behaviors—closing gaps between IDE assistance and continuous integration.
Snowflake Cortex AI for Financial Services unifies proprietary and third-party data for secure insights, automated analysis, and fraud detection—simplifying AI deployment for banks and investment firms.

Google Gemini 2.5 Flash Image delivers fast generation, flexible aspect ratios, and natural-language editing, plus the free “nano banana” tool—broadening creative AI access for developers, creators, and businesses.
IBM Granite 4 combines Mamba and Transformer architectures for faster, more efficient enterprise reasoning—positioning IBM for customer service, analytics, and regulated-industry workloads.
Anthropic Claude Sonnet 4.5 tied Opus 4.1 atop Text Arena; strong coding, tool-use, and vision gains signal a razor-thin frontier among leading models.
Zhipu GLM-4.6 published full test sets and trajectories, improving evaluation transparency while closing gaps on code-edit success—an encouraging move for reproducible benchmarking.
Alibaba Qwen3‑VL‑235B (FP8) halves compute needs with minimal quality loss, suggesting precision tweaks can unlock cheaper, scalable multimodal deployments.
Huawei SINQ quantization achieves high accuracy with sharply reduced memory, underscoring rapid progress in efficient model serving for long-context and on-device scenarios.

MIT and IBM published scaling laws predicting LLM performance at lower cost, giving smaller teams a playbook to budget training and avoid wasteful over-parameterization.
MIT and McMaster used AI to discover enterololin, a new antibiotic targeting Crohn’s-related bacteria—accelerating drug discovery and offering hope against antimicrobial resistance.
UCLA demonstrated an optical generative model that produces images using light instead of GPUs, hinting at ultra‑low‑power AI hardware and new directions for photonic computing.
Scientists combined satellite imagery and AI to refine Serengeti migration data, enabling more accurate conservation decisions and resource planning for an iconic ecosystem.
New AI-driven protein and DNA design tools promise faster therapeutics and agriculture, but raise biosecurity risks—prompting calls for stronger safeguards and oversight frameworks.
Apple researchers explored detecting hallucination spans within model outputs, improving reliability by flagging uncertain passages for review rather than rejecting entire responses.

OpenAI reached a $500B valuation, surpassing SpaceX; ahead of DevDay on Oct 6 and amid xAI legal tussles, momentum and scrutiny grow as platform competition intensifies.
NVIDIA became the first $4T public company, cementing GPU leadership as enterprises scale AI workloads across data centers, edge deployments, and new agentic applications.
The U.S. State Department unveiled a 2026 AI strategy focused on ethics, global cooperation, and workforce training, aligning diplomatic tools with rapid advances in AI governance and security.
Italy’s privacy regulator banned Clothoff for generating fake explicit images, setting a strong precedent against AI-enabled exploitation and strengthening protections for minors.
Enterprise AI infrastructure heats up: Fujitsu + NVIDIA expanded their partnership, while Hitachi + OpenAI target greener, power‑efficient data centers to meet soaring AI energy demands.
Microsoft Sentinel added a unified data lake and agentic AI to automate threat detection and response, consolidating security workflows for organizations of all sizes.

Andrew Ng announced a new deep learning course, refreshing fundamentals for practitioners navigating today’s shifting architectures and training regimes.
A practical roundup highlighted seven capable models that run locally on laptops—useful for private prototyping, offline work, and avoiding cloud costs.
Guides demonstrated 10–100x faster Bayesian inference in Stan using JAX on commodity GPUs, making advanced probabilistic modeling more accessible.
DSPy’s GEPA automated prompt optimization for difficult classification tasks, reducing manual iteration and boosting accuracy on edge cases.
Explainers clarified the Qwen model family, helping teams pick variants, tooling, and inference strategies suited to vision, long context, or constrained hardware.

OpenAI Sora 2 Pro wowed testers with crisp 15‑second clips and faithful storyboard‑to‑video, while raising fresh questions about provenance and copyright.
KLING 2.5 Turbo delivered cinematic sequences and smoother motion, suggesting accelerating quality across competing video generators.
Luma’s Ray 3 and HDR 3 entered head‑to‑head community trials, giving creators practical comparisons on style, detail, and speed.
Moondream showed single‑frame thermal fault detection for pipelines, hinting at industrial monitoring gains beyond generative media.
Video Agents enabled live, conversational analysis over video streams, illustrating how multimodal agents can handle interactive surveillance, coaching, and support workflows.

Experts argued there’s no single “god model”; choosing the right tool for the task beats chasing monolithic state‑of‑the‑art rankings.
Emerging evidence suggests teaching reasoning earlier in pretraining—and using supervised fine‑tuning strategically—can reduce catastrophic forgetting and improve generalization.
A formal limit on information learned per episode from policy gradients reframes expectations for long‑horizon RL, motivating hybrid methods and better curricula.
Proposals to train “secret‑keeper” systems and analyses of PPO/GRPO’s human‑like interpretation aim to stress‑test safety and better understand reward‑shaped behavior.
Synthetic data best practices: moderate rephrasing around 30% often helps; textbook‑style synthetic data tends not to—underscoring careful curation over sheer volume.
Builders emphasized memory, proactive assistance, and thoughtful UX as the real differentiators for everyday AI utility, not parameter counts alone.

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