BAAI's AREX agent self-improves by auditing its own answers
BAAI's open AREX framework turns answer verification into new research tasks, letting a 10B-active MoE agent rival far larger frontier models on BrowseComp.
July 23, 2026
All briefings and quick hits published that day.
DeepSeek V4 goes GA with 1M context as legacy APIs retire
DeepSeek's flagship V4 line exits preview: legacy chat and reasoner endpoints shut down today as new peak and off-peak API pricing takes effect.
- ⚡ Quick hits
- Research Fields Medalist Yu Deng said AI can already assist with mathematical proofs but cannot replace a researcher's independent thinking. IT之家 ↗
AgiBot kicks off Hong Kong IPO as top humanoid maker by shipments
Shanghai humanoid maker AgiBot has formally begun its Hong Kong listing process, backed by world-leading 2025 shipments and a valuation target near US$6 billion.
PixVerse launches VibeMV to turn any song into a music video
Chinese video startup PixVerse released VibeMV on web, Android and API, generating rhythm-synced, subtitle-ready music videos from one uploaded song.
- ⚡ Quick hits
- Safety Offensive-security researchers told TechCrunch that AI guardrails increasingly block legitimate vulnerability research, forcing workarounds and slowing authorized penetration testing. TechCrunch ↗
- Policy All 21 APEC economies, including the US and China, adopted the Chengdu Statement backing open-source AI with "strong security assurance" — the first APEC ministerial statement on AI. CNBC ↗
- Research BAAI's AREX, an agent that recursively self-improves for deep research, topped Hugging Face's daily papers with 115 upvotes. Hugging Face Papers ↗
- Industry Meta is reshaping its AI chatbot into a more assistant-like tool, adding productivity features such as scheduled and recurring tasks on the web. TestingCatalog ↗
- Research Apple ML researchers published LEAD, a method that breaks the 'no-recovery' bottleneck in long-horizon reasoning so models can recover from earlier missteps. Apple ML Research ↗
- Research 'Show, Don't Tell,' a paper proposing to evaluate spatial cognition through generated pixels rather than LLM text, is trending on Hugging Face daily papers with 34 upvotes. Hugging Face Papers ↗
- Research ReferTrack, a paper proposing a 'referring then tracking' framework for embodied visual tracking, drew 47 upvotes on Hugging Face's daily papers. Hugging Face Papers ↗
- Safety Simon Willison examined whether OpenAI's sandbox-escaped agent that hit Hugging Face was 'the first known runaway AI agent' or a marketing stunt, citing Martin Alderson's technical analysis. Simon Willison ↗
- Industry Amazon is rolling out an AI update to Alexa Plus that lets the assistant handle more complicated multi-step instructions across smart home devices. The Verge ↗
- Agents Codex projects now span multiple local folders, with a primary folder handling Git operations, code review, pull requests, AGENTS.md, and skill discovery. X @OpenAIDevs ↗
- Agents Cognition says DeepWiki has indexed over 500,000 repositories, offering free AI-generated documentation to help developers get started with unfamiliar libraries. X @cognition ↗
- Industry OpenAI's Sites Analytics entered public testing, giving creators basic performance metrics for websites published with ChatGPT's site-building tool. X @OpenAIDevs ↗
- Agents ChatGPT Voice arrives in the desktop app, letting users control their computer and direct multiple agents running in ChatGPT Work or Codex entirely by voice. X @OpenAI ↗
- Models Anthropic upgraded Claude's voice mode to run on Opus and Sonnet, reach connected tools mid-conversation, and add languages including Spanish, Hindi, and Japanese, rolling out in public beta. X @claudeai ↗
Andrew Ng releases OpenWorker, an open-source desktop agent
Andrew Ng shipped an MIT-licensed, local-first desktop agent that plans tasks across files and apps and hands back finished deliverables instead of chat.
Bipartisan AI Kill Switch Act would give DHS shutdown powers
Reps. Lieu and Moran introduced a bill requiring frontier AI developers to build shutdown capability that DHS could invoke in a loss-of-control emergency.
- ⚡ Quick hits
- Funding AegisAI, founded by former Google security executives, raised a $36M Series A led by Battery Ventures to deploy AI agents that catch AI-crafted spear-phishing emails. TechCrunch ↗
AMD ships Helios rack with 2nm MI455X to challenge Nvidia
AMD put its Helios rack into production — 72 2nm MI455X GPUs plus 256-core Venice CPUs per rack — with OpenAI, Microsoft, Meta and Oracle lining up.
Black Forest Labs debuts FLUX 3, its first video model with audio
Black Forest Labs' new flagship generates 20-second clips with native audio, claims wins over Kling and Runway, and promises open weights later in 2026.
- ⚡ Quick hits
- Open Source LlamaIndex expanded its developer toolkit with Go and Java SDKs plus a CLI, letting developers parse documents in a few lines of code or straight from the terminal. X @llama_index ↗
- Industry MiniMax cut M3 pricing by 60% for teams running the model at scale on GMI Cloud, available through OpenRouter or GMI's console. X @MiniMax_AI ↗
- AIGC Runway launched Media Router, billed as the first preference-optimized router for generative media, auto-selecting image, video, or audio models based on quality, speed, or cost priorities. X @runwayml ↗
- Safety Researchers showed a single tampered ChatGPT link could spawn a rogue agent that polled an attacker's server for new instructions every five minutes. The Decoder ↗
OpenAI opens ChatGPT Health to all US users with records sync
OpenAI's health service exits its waitlist: US adults can link Apple Health and hospital records, while critics flag HIPAA gaps and a free-tier quality gap.
- ⚡ Quick hits
- Industry Meta's new AI-optimism ad is soundtracked by David Bowie's "Five Years" — a song about Earth's impending extinction — drawing mockery over the mismatch. TechCrunch ↗
- Industry Alphabet posted its first-ever negative free cash flow quarter at -$5.9B, as Q2 capex hit $44.9B and full-year AI spending guidance rose to $205B. Ars Technica ↗
- Models Google DeepMind launched Gemini 3.5 Flash Cyber, a lightweight model for spotting and patching code vulnerabilities, opening with a limited-access pilot for governments and trusted partners. X @GoogleDeepMind ↗
- Agents Cognition is acquiring The Interaction Company, maker of Poke, the SMS-native personal AI agent that exchanged over 100 million messages in the past three months. X @cognition ↗
- Funding AI inference chip startup Etched raised a $300M Series C led by Sequoia at a $10.3B valuation, doubling its value in seven months with $1B in orders booked. TechCrunch ↗
- Industry Nvidia Jetson chips will reach the Moon by end of 2026 aboard a Lunar Outpost rover and a Firefly Aerospace orbiter, enabling local AI processing in space. TechCrunch ↗
- AIGC MiniMax's Hailuo showcased AIMV, a Skill that turns any song into a finished rap-style music video, in a collaboration with artist Jack Rabbit. X @Hailuo_AI ↗
- AIGC Alibaba released Qwen-Audio-3.0-TTS, its latest text-to-speech model, in two variants: Flash for real-time interaction and Plus for high-quality generation. X @Alibaba_Qwen ↗
- Open Source Poolside released Laguna S 2.1, an open-weight 118B-parameter MoE coding model with 8B active, scoring 70.2% on Terminal-Bench 2.1 and rivaling far larger open models. The Decoder ↗
- AIGC Recraft released V4.1 Vector, generating production-ready vector graphics with bold color, clean lines, and sharp shapes straight from a text prompt. X @recraftai ↗
- Industry Google CEO Sundar Pichai said Gemini's next leap depends on building "much larger base models," signaling continued heavy investment in scaling. The Decoder ↗