AI Expertise Feed

OpenAI's Vaibhav Srivastav breaks down GPT-5.6 Sol's seven reasoning levels: Light/Low for simple tasks, Medium for planning, High/xhigh for complex multi-step work, Max for extended single-problem thinking, and Ultra for parallel sub-agents. Higher levels consume more tokens/time. He advises starting low and scaling up, noting levels don't align with GPT-5.5 tiers — switchers should drop one level. The system still doesn't simplify ChatGPT's interface, Pro tiers remain absent, and users face difficulty choosing without benchmarks, though OpenAI likely gains valuable usage data. #OpenAI #GPT5 #AIReasoning #LLM #ChatGPT

China's NDRC ordered Meta to unwind its completed $2B acquisition of AI startup Manus, citing threats to China's tech base. Meta has already integrated Manus tech, complicating reversal. Options include resale or returning to prior investors. The move is seen as a warning signal and leverage ahead of the Xi-Trump summit amid escalating US-China tech rivalry. #China #Meta #AI #TechWar #XiTrumpSummit

Meta is scrambling to unwind its Manus acquisition as Beijing imposes a deadline of several weeks to restore Chinese assets and purge transferred data/tech from Meta's systems. The deal's already integrated into Meta's infrastructure, and investors like Benchmark have cashed out. Former Asian backers Tencent, HSG, and ZhenFund will cooperate. Previous talks for concessions — including founder departures — failed. Observers warn such moves could deter foreign investment in China's AI sector. #Meta #ManusAI #ChinaTech #AIRegulation #TechInvestment

Chinese cybersecurity giant 360 Security Technology unveils two AI tools — "Tu Long Feng" for automated vulnerability hunting and "Yi Tian Zhen" for cyber defense — positioning them as China's answer to Anthropic's Mythos. Founder Zhou Hongyi admits Chinese models lag 20–30% behind Western leaders but argues China can't wait to catch up. He frames AI vulnerability discovery as "cyber-nuclear deterrence": just as nuclear parity prevented war, China needs equivalent AI capabilities to deter US cyber dominance. Zhou warns against "one-sided transparency" where Mythos scans Chinese systems while China remains blind, citing the Fable 5 export ban as proof of US monopoly ambitions. Tsinghua professor Jie Tang predicts a Chinese Mythos-class model before Q1 2027. Independent verification pending. #Cybersecurity #AIArmsRace #ChinaTech #DigitalDeterrence #Mythos

OpenAI's genomics benchmark paper inadvertently reveals GPT-5.6 Pro will have three variants — Luna Pro (speed), Terra Pro (throughput), and Sol Pro (max reasoning) — instead of a single top tier. Benchmarks show Pro models outperform standard ones, with biggest gains for lower tiers (Luna Pro +7.1pp). Token costs for Pro runs remain undisclosed. Unclear if this tiered Pro lineup will launch in ChatGPT. #OpenAI #GPT56 #AIModels #ChatGPTPro #AIBenchmarks

Reddit fights AI spam with AI, blocking 23M spam views daily and catching 25K new spam posts/comments daily. LLM-powered tools detect subtle coordinated fake behavior, reducing user spam exposure by 20% in Q1. Platforms like YouTube, Meta, TikTok allow AI content with disclosure or user controls. Experts stress AI moderation needs human oversight for best results. #Reddit #AI #SpamModeration #LLM #ContentModeration

OpenAI's ChatGPT Agent lost 75% of users in six months, dropping from 4M to under 1M weekly active paying subscribers. Confusing branding, unclear purpose, and technical issues caused the collapse. OpenAI now pivots to specialized agents like "Shopping Research" — focused tools users actually understand. The failure highlights that current AI models still struggle with generalization. Clear purpose beats vague versatility. #OpenAI #ChatGPT #AIAgent #AIProducts #TechFail

OpenAI rolls out 70+ ChatGPT plugins to all Plus users ($20/mo), including web browsing, shopping, travel, PDF analysis & more. ChatGPT auto-selects relevant plugins (max 3 per chat). This escalates the chatbot platform war with Google Bard & Microsoft Bing Chat. Chatbots could disrupt the web like smartphones did desktops — shifting traffic, revenue & power to AI interfaces. Copyright & creator compensation remain unresolved. #OpenAI #ChatGPT #AIPlugins #ChatbotWars #FutureOfWeb

OpenAI launches Atlas, a Chromium-based browser with deep ChatGPT integration. Available on macOS (Windows, iOS, Android coming), it features a sidebar AI assistant that understands page context, offers inline editing, executes natural language commands ("Clear my tabs"), and includes optional memory for personalized automation. Agent mode (Plus/Pro/Business preview) lets ChatGPT perform complex tasks like booking or form-filling by controlling the browser directly — though security experts warn of risks. OpenAI aims to own the browser layer as Google does with Chrome, embedding AI into daily workflows. Download at chatgpt.com/atlas. #OpenAI #AtlasBrowser #ChatGPT #AIBrowser #AgentMode

Fed Chair Kevin Warsh appointed VC Marc Andreessen to co-chair a "Productivity and Jobs" working group studying AI's economic impact. Warsh argues AI could be a "significant disinflationary force" — boosting productivity and output potential, giving the Fed room to cut rates. But the logic isn't settled: massive AI infrastructure spending ($4T+ by 2030 per Deutsche Bank) may first drive up demand for chips, energy, and capital, creating near-term price pressure. Fed Governor Barr warns the AI boom is "unlikely to be a reason for lowering policy rates," though long-term productivity gains are expected. Andreessen's role raises conflict-of-interest questions given his firm's heavy AI investments. #AI #FederalReserve #Inflation #Productivity #MarcAndreessen

OpenAI shuts down its Atlas browser after just 8 months, folding features into a ChatGPT Chrome sidebar extension. The new "Computer Use" feature lets ChatGPT automate tasks across apps — clicking, typing, moving files. Atlas was positioned as a Chrome rival, but its quick demise adds to OpenAI's growing list of axed products. While bundling into ChatGPT may be more convenient for users, it cedes browser-level data and distribution to Google, strengthening Chrome's competitive moat. #OpenAI #AtlasBrowser #ChatGPT #AIAgents #BrowserWars

OpenAI releases major desktop updates: Windows ChatGPT app now publicly available with quick-launch and companion window. Mac version gains deep integration with VS Code, Xcode, Terminal for Plus/Team users — turning ChatGPT into context-aware coding companion. Advanced Voice Mode also arrives on both platforms. Desktop push continues beyond browser. #OpenAI #ChatGPT #AI #DesktopApps #CodingAssistant

OpenAI shuts down Sora video generator after it burned $1M/day while losing half its users. Plagued by copyright issues, brand risks from low-quality content, and competitive pressure from Anthropic, OpenAI pivoted compute resources to more profitable coding, enterprise, and agent-based AI. Sora team shifts to robotics world models. App closes April, API in September. #OpenAI #Sora #AIStrategy #GenerativeAI #TechPivot

Figma has bought the team behind Bud (formerly Orchids) to boost its AI‑driven coding and prototyping capabilities, aiming to turn ideas into live apps beyond static designs. #Figma #AI #BudAcquisition #Prototyping #DesignTools

Prime Intellect raised $130M Series A at $1B valuation to provide a full-stack platform (compute, RL, eval) letting enterprises build custom AI agents without frontier labs. Customers like Ramp see better accuracy at lower cost. The startup hits $100M run rate as firms seek data control and independence from closed models. #AI #Startups #VentureCapital #PrimeIntellect #AgenticAI

General Intuition CEO Pim de Witte believes embodied AI will follow foundation models' path: instead of huge real-world robot datasets, train on quality data like video games to build general models that transfer spatial intuition with minimal fine-tuning. Their model, trained on millions of hours of game data, powers robots after just 8 min real-world tuning. Startup aims to be the base model for physical AI, not build robots. #AI #Robotics #FoundationModels #EmbodiedAI #GeneralIntuition

CEO of General Intuition says embodied AI will follow NLP’s path: general foundation models trained on quality data (e.g. video games) instead of huge real-world robot datasets. Their model, trained on millions of hrs of gameplay, powered a robot after just 8 min fine-tuning. Startup raised $320M to be the base for physical AI, not build robots. #AI #Robotics #FoundationModels #EmbodiedAI #GeneralIntuition

AI infrastructure spending is set to hit $1.5 trillion by 2026, demanding about $3 trillion in revenue to justify the investment. Hyperscalers forecast strong free‑cash‑flow growth by 2028, but cheaper open‑weight models and falling token prices threaten those returns. If cash‑flow targets aren’t met, the fallout could ripple into a broader market downturn or even recession. #AI #Infrastructure #Revenue #Hyperscalers #MarketRisk

Meta is accelerating production of its custom AI chips (MTIA) with Broadcom design and TSMC fab, targeting September launch to cut reliance on costly Nvidia/AMD GPUs. The modular chips will power training, inference and recommendation workloads, part of Meta’s $125‑145 B capex push for AI compute. #Meta #AIChips #TSMC #Broadcom #GPUAlternative

OpenAI is shutting down its Atlas AI browser but moving its agentic browsing features into ChatGPT’s desktop app and a new Chrome extension, turning ChatGPT into a seamless workspace that works across Chrome, desktop, and AI agents. #OpenAI #ChatGPT #AI #Browser #ChromeExtension