A 1,200-line Waymo system prompt reveals heavy guardrails; Salesforce loses faith in LLMs; Google may put AI answers behind a paywall; OpenAI's voice engine can clone voices from 15s of audio — huge potential and big risks.
OpenAI reportedly considering in-chat ads that could erode neutrality. Waymo is testing Google's Gemini as a conversational in-car assistant — leaks spark safety and hallucination concerns. A Stanford/Harvard paper pressures agentic AI claims, and a budget soundbar review highlights clearer voice interfaces.
John Carreyrou and other authors sue major AI firms over low settlement offers; OpenAI warns prompt-injection attacks may be unsolved; Alexa Plus gains deep booking integrations (Angi, Expedia, Square, Yelp); DeepMind's Hassabis calls out GPT-5 hype.
Alleged GPT-5 solves an open math problem, Yann LeCun and Demis Hassabis clash over AGI, OpenAI warns agentic browsers may be vulnerable to prompt-injection, and Amazon’s Rufus rethinks shopping — breakthrough claims, heated debate, and real security concerns.
Anthropic's Claude Opus 4.5 claims a ~4h49m autonomous "time horizon." Nvidia's NitroGen learns from 40,000 gameplay hours to build vision-action agents. Kuaishou's Kling 2.6 adds voice control and motion. PagerDuty warns Shadow AI and agent autonomy create SRE and governance risks—human oversight required.
Anthropic’s open Agent Skills gain traction (Microsoft adoption) while OpenAI pushes a ChatGPT App Store. Palona AI uses vision and workflows to run kitchens—spotting undercooked pizzas—with a memory called “Muffin.” GPT-5.2 aces benchmarks but struggles with real research, framing open-standard vs. walled-garden stakes.
Google’s Gemini 3 Flash becomes the default—speedy and cheap but with a ‘reasoning tax.’ Nvidia’s Nemotron 3 introduces Mamba for long contexts. Visa + Akamai enable verified shopping bots. Plus CrewAI + Gemini pipelines for cheap, autonomous research agents—huge potential and new risks.
AI is shifting from single-model showdowns to orchestration. Zoom’s federated “traffic controller” outscored rivals on the so-called “Humanity’s Last Exam.” OpenAI launched GPT Image 1.5 for faster, precise edits, Zenflow builds model pipelines for coding, and Adobe opens Firefly to FLUX — integration wins.
Nvidia acquires SchedMD (Slurm) and launches Nemotron 3 open models. Korean startup Motif’s 12.7B model outperforms GPT‑5.1 by rethinking data and long‑context infrastructure. Yann LeCun doubles down against scaling LLMs, and Mirelo raises $41M to add synced foley to AI-generated video.
Reasoning models are crushing the CFA (Gemini 3.0 Pro scored 97.6%). Adobe brings Photoshop, Acrobat and Express into ChatGPT for in-chat edits. ZDNET’s robot vacuum can pick up laundry, and AI AR glasses add contextual overlays — with privacy and real-world limits to consider.
OpenAI unveils GPT-5.2 (Instant, Thinking, Pro) with a 400k-token context, big benchmark wins and steep Pro pricing. Google fires back the same day with a Gemini 3 Pro–based research agent. Breakdown of performance, costs, enterprise focus, missing image features, and the agent race.
Quilter’s AI designed and manufactured a Linux PC board in one week and it booted on the first try. Google’s new FACTS benchmark reveals roughly a 70% factuality ceiling, with vision tasks failing often. ChatGPT topped the US App Store and Google launched managed MCP servers to simplify AI agent integrations.
Mistral unveils Devstral 2 (123B) and a powerful 24B Small 2 plus Vibe CLI, while differing licenses split free indie use from revenue‑capped enterprise access. OpenAI, Anthropic and Block join the Linux Foundation to standardize agents. SAP reveals stark human bias against AI-labeled work; Google bets Gemini on XR.
Z.ai drops GLM-4.6V (106B) and a 9B Flash under MIT with native function-calling and long-context video—huge for on-prem agents. Also: Chrome’s agent guardrails, Booking.com’s modular AI wins, and OpenAI’s ‘Code Red’ as Google and open source tighten the market.
OpenAI pulls ad-like app suggestions after backlash. Engineers warn AI coding agents are brittle, insecure, and often add technical debt. Meanwhile open-source GeoVista narrows the gap with commercial geolocation models — powerful but privacy-risky.
Google's Gemini 3 Deep Think adds visible reasoning. A new GAM memory architecture promises long-term context without token bloat. OpenAI tests a 'confession' channel to surface uncertainty. Plus Gong data: AI-powered sales teams drive 77% more revenue and shadow AI use is booming.
Gemini 3 Pro soars in blinded trust tests, AI models prove adept at finding and exploiting smart-contract bugs, Google unveils Workspace Studio to build Gemini agents inside Docs/Sheets, and Meta poaches Apple design lead Alan Dye — big shifts in trust, security, and UX.
DeepSeek's MIT‑licensed breakthroughs, OpenAGI's Lux desktop agent, US-built Trinity models, on-device small-model blueprints, and AWS–Visa agentic commerce — how open-source, agents, and efficient models are reshaping AI infrastructure.
A roundup of AI breakthroughs and challenges: GPT-5's touted math prowess, ontology-based guardrails for enterprise agents, Meta's Matrix for decentralized synthetic data, GAM memory to combat context rot, and the learning trade-offs of coding assistants.
A roundup of major AI moves: the U.S. 'Genesis' closed-loop research push, Black Forest Labs’ Flux 2 image advances, OpenAI’s data residency and voice UI updates, a music-label AI deal, troubling safety findings about ChatGPT, plus new autonomous agents and cost-aware LLM routing.