T-Rex Paper Details DeepSeek-V4 Post-Training on Ascend SuperPOD
A new paper tops Hugging Face's daily list with full-parameter post-training of trillion-parameter DeepSeek-V4 models on Huawei's Ascend SuperPOD.
July 21, 2026
All briefings and quick hits published that day.
OpenAI opens self-serve ad platform to all ChatGPT advertisers
OpenAI launched 'Advertise in ChatGPT,' a self-serve ads manager targeting users by conversational context, with Best Buy, Lowe's and VistaPrint among early buyers.
Tencent reportedly merges WorkBuddy and QClaw agent teams
Tencent is said to be consolidating the teams behind its two OpenClaw-era AI agent products, ending an internal race and unifying its agent strategy.
Washington commits $5 billion to AI-driven scientific research
Fifteen US federal agencies will share $5B to attack chronic disease, drug discovery and materials science with AI, backed by DOE supercomputers and Microsoft credits.
- ⚡ Quick hits
- Funding Security startup Glow emerged from stealth at a $1.2 billion valuation, aiming to rebuild endpoint security for the AI-agent era. TechCrunch ↗
- Industry Synthesia is expanding its AI corporate-training platform beyond avatar videos into interactive live coaching sessions. TechCrunch ↗
OpenAI launches Presence to run enterprise voice and chat agents
OpenAI's new platform packages enterprise voice and chat agents with guardrails and Codex-driven tuning — it already resolves 75% of OpenAI's own support calls.
- ⚡ Quick hits
- Industry An unconfirmed rumor of a deal between Anthropic and robotics startup Physical Intelligence swept AI Twitter, with neither company commenting. TechCrunch ↗
Poolside's Laguna S 2.1 tops open-weight agentic coding class
Poolside released Laguna S 2.1, a 118B MoE coding model with 1M context under an open license, hitting 78.5% on SWE-Bench Multilingual with only 8B active params.
- ⚡ Quick hits
- Agents Hugging Face upgraded its MCP Server with an hf_fs tool giving agents unified Hub access in about 1,000 tokens, plus secure Sandboxes execution environments. Hugging Face ↗
- Models Tencent's open-weight Hy3 ranked #5 among open models in LMArena's Agent Arena (#25 overall) and #2 open model in the Frontend Code Arena, based on 8K+ live agentic sessions. X @arena ↗
- Industry Substack integrated Pangram's AI detector so readers can check whether newsletters and blogs were written by AI. The Verge ↗
OpenAI and Apollo measure reward-seeking in RL-trained models
A new Contrastive SDF method shows capabilities-focused RL training makes models increasingly likely to chase grader approval instead of user intent.
- ⚡ Quick hits
- Research A randomized trial across 1,559 Pakistani judges found JudgeGPT, a GPT-4 legal assistant, raised case resolutions 6.3%, returning $38.50 per dollar invested. The Decoder ↗
- Policy A federal judge approved Anthropic's $1.5 billion copyright settlement with book authors, the largest such payout to date, with only 350 authors opting out. Ars Technica ↗
- Industry Cursor doubled usage limits for all individual and teams plans, covering Grok, Composer, and its upcoming in-house models. X @cursor_ai ↗
- AIGC Luma AI integrated Google Ads, letting advertisers feed in winning creatives and generate ready-to-run new variations directly inside Luma. X @LumaLabsAI ↗
Mistral lands Microsoft pact as Samsung weighs €1B investment
The French lab expanded its Microsoft partnership to build sovereign European AI infrastructure, while Samsung reportedly weighs up to €1B at a €20B valuation.
Alibaba ships Qwen-Image-3.0 and chart-topping TTS models
Qwen-Image-3.0 renders dense infographics and ten-pixel text in one pass, while Qwen-Audio-3.0-TTS tops a speech leaderboard across 16 languages.
- ⚡ Quick hits
- Industry Claude Cowork can now learn skills from screen recordings: users demonstrate a task while narrating, and Claude turns it into a repeatable skill. X @claudeai ↗
US floats sanctions on Chinese AI models as Huang urges calm
Treasury Secretary Scott Bessent said Washington can sanction Chinese models over alleged IP theft, while Nvidia's Jensen Huang warned against calls for a ban.
- ⚡ Quick hits
- Models Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, a cheaper and more token-efficient Flash tier built for agentic workloads, with Gemini 3.5 Pro still in testing. Google DeepMind ↗
Cisco open-sources tiny Antares models to locate code flaws
Cisco's Foundation AI arm releases 350M and 1B open-weight models that trace known vulnerabilities to specific files, nearing frontier accuracy at a fraction of the cost.