All Updates
Page 1067 of 1570
March 26, 2026
Apple Siri to Integrate Third-Party AI Chatbots
Apple's iOS 27 update will allow users to link third-party AI chatbots from the App Store, such as Google's Gemini or Anthropic's Claude, to Siri for fetching replies. This 'Extensions' system enables toggling integrations on iPhone, iPad, and Mac, building on the existing ChatGPT integration. The feature expands Siri's AI capabilities beyond Apple's ecosystem.
Apple Ends 512GB Mac Studio Sales
Apple has discontinued the 512GB unified RAM option for Mac Studio M3 Ultra due to memory supply crisis. Maximum now 256GB, impacting high-end local AI workloads. Speculation suggests M5 series may cap at 128-256GB.
Apple Axes Mac Pro for Mac Studio
Apple discontinued its high-end Mac Pro desktop on Thursday. The machine targeted video editors, photographers, and advanced pro users. The company now favors the compact Mac Studio alternative.
Gemini Imports Chats Flawlessly from Other AIs
Gemini enables seamless transfer of chats from other AI apps like ChatGPT. The process works flawlessly, allowing Gemini to learn from past conversations. This feature revolutionizes switching to Gemini while preserving history.
Apple's AI Playlist Playground Bad at Music
Apple's new AI Playlist Playground beta in Apple Music poorly handles specific music prompts, such as atmospheric instrumental black metal, recommending mismatched tracks with vocals, field recordings, ambient electronic, and doom jazz. The author expresses skepticism about AI's music recommendation capabilities. YouTube Music's AI generator performs better on the same prompt.
Apple Bonuses iPhone Designers vs OpenAI Poaching
Apple awarded rare bonuses to iPhone hardware designers this week to counter departures to AI startups like OpenAI. These startups are building their own devices, fueling a talent poaching wave. This retention effort highlights escalating competition for hardware expertise.
Gemini Adds Chat Import from Rivals
Google released tools for Gemini allowing users to upload chat history and context from other AI apps. This feature targets users of rivals like OpenAI’s ChatGPT and Anthropic’s Claude to ease switching.
WhatsApp iPhone Dual Accounts AI Photos
WhatsApp's March 2026 update enables two accounts on one iPhone. It adds smarter storage controls, cross-platform chat transfers, and AI-powered photo editing for power users.
Eval Traces Over Outputs for Local LLM Agents
Local LLM agents can yield correct finals despite internal flaws like wrong tools, loops, or extra steps. Author urges trace evaluation: tool choice, steps, loops, reasoning quality. Shares GitHub rubric-eval for local checks using Ollama judge.
OpenClaw Creator Urges US to Copy China AI
OpenClaw's creator argues the US can learn from China's rapid AI adoption. The AI tool has exploded in popularity in China but faces mixed reactions in the US.
Claude Cowork Tests Computer Use & Dispatch
Claude Cowork and Claude Code now support Computer use and Dispatch features. The article details hands-on testing to assess their practical capabilities. It highlights both strengths and key limitations revealed in real use.
1M Tokens/s Qwen 3.5 on B200 GPUs
Achieved 1.1M total tok/s serving Qwen 3.5 27B (FP8) on 96 B200 GPUs using vLLM v0.18.0. DP=8 outperforms TP=8 by 4x; MTP-1 boosts utilization, MTP-5 crashes. 97% scaling efficiency, but Inference Gateway adds 35% overhead.
Qwen 3.5 27B Reaches 1.1M tok/s on B200s
Achieved 1,103,941 tok/s on 96 B200 GPUs across 12 nodes using vLLM v0.18.0. Optimizations include DP=8 over TP=8, 4K context, FP8 KV cache, and MTP-1 speculative decoding boosting utilization. Near-perfect scaling at 96.5% efficiency without custom kernels.
Kensho's LangGraph Multi-Agent for Finance Data
Kensho, S&P Global’s AI innovation engine, built the Grounding framework using LangGraph. This multi-agent system provides a unified access layer for trusted financial data retrieval. It tackles fragmentation issues at enterprise scale.
AI Coding Tools Spike Vulnerabilities
Usage of AI coding assistants has surged, yet vulnerabilities in AI-generated code have risen correspondingly. AI tools are introducing more security flaws as adoption grows.
Google Live Translate Lands on iOS Globally
Google's AI-powered Live Translate earphone feature now supports iOS devices. Previously limited to Android in the US, India, and Mexico, it's expanding to iOS and Android users in 12 countries including Germany, UK, Japan, and Thailand.
Hard CoT Interp Tasks Released
Researchers introduce nine objective chain-of-thought interpretability tasks where black-box LLM monitors fail OOD. They open-source datasets with in-dist and OOD evals, baselining probes, TF-IDF, and LLM monitors—non-LLM methods outperform OOD. Testbed to benchmark community progress beyond simple CoT reading.
Apple to Open Siri to External AI Assistants
Apple plans to allow external AI assistants access to Siri in an upcoming iOS update. This follows integration with OpenAI's ChatGPT and extends to competitors. The move aims to position iPhone as a stronger AI platform.
Retail Investors Sell Nvidia First Since July
Individual investors dumped Nvidia shares for the first time since July on Wednesday. This marks a potential shift after Nvidia's meteoric rise to S&P 500's largest company. The selling coincides with market pullback from AI-fueled rally amid Iran war tensions.
EBMs Beat MLPs on OOD Boundaries
Reddit discussion compares Energy-Based Models (EBMs) to ReLU-MLPs, revealing EBMs superiorly handle OOD points near training data boundaries without 'spandrels'. Using toy datasets like split circle, twist, and kissing pyramids, MLPs extrapolate linearly creating artifacts, while EBMs avoid such assumptions. EBMs assign scalar energy to configurations, minimizing it for inference.