AI Titus's Morning Wire

AI NEWS TODAY

IT ADMINS CAN NOW SWITCH ON CLAUDE'S APP CONNECTORS FOR A WHOLE COMPANY AT ONCE. ANTHROPIC'S ENTERPRISE-MANAGED AUTH WENT LIVE TODAY THROUGH OKTA, SLACK, NOTION AND DATADOG INCLUDED
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadAlibaba's Wan3.0 turns documents, slides and web pages into 30-second videos
Launched Monday, one day after Alibaba closed its $10.2 billion share sale. Clips run twice as long as the last version, and the model takes documents, spreadsheets, slides and web pages as input, not just text prompts. It ran in public beta since August 6 and is already used for ads, short drama and tourism videos. Feeding it a sales deck and getting a video back is the part worth testing.
★ Must-ReadQwen3.8-27B is the only small model in Code Arena's WebDev top 10
It landed at #9 with 1595 points, six ranks behind its giant sibling Qwen3.8-Max. A 27B open-weights model you can run on one good GPU is now trading blows with frontier systems on real web development tasks. If you have been waiting for local models to get good enough for daily coding work, this is the signal.
OpenAI launched AI Futures, a research blog on who holds power in an AI economy
Posted August 24 by OpenAI's new Strategic Futures team. The team calls concentration of power the biggest and hardest long-run risk: how does a free society keep individual rights when a few actors control transformative AI? Planned output is posts, papers, videos and podcasts. Worth bookmarking whichever lab you favor.
Headlong, a persistent agent that runs continuously and debugs itself for $1 to $2 an hour
Andy Konwinski showed an agent that keeps running instead of stopping at the end of a task, fixing its own failures along the way, with a 48-minute self-debugging run as the demo. The always-on agent that watches its own work is becoming a product category, and the hourly cost is already close to nothing.
📚 LearnAn $800 mining GPU was unlocked to 64GB and runs Qwen3.8-27B at 84 tokens a second
A kernel patch unlocks the full 64GB of HBM on NVIDIA's CMP 170HX, a card sold cheap for crypto mining. The author reports 84 tokens a second at short context and 57 at 200K, with prefix caching cutting a 200K prefill from 157 seconds to 2.5. Numbers are the author's own, but the repo is public and the recipe is documented.
📚 LearnPipette: 10,000+ verified benchmarks of which models actually run on which devices
Liquid AI and Artificial Analysis released a benchmark covering 35 models, 7 quantization levels and multiple runtimes across 4 devices. It answers the question every local-AI person asks: what will this model really do on my phone or laptop, not on a lab server. Check it before you buy hardware or pick a model for the edge.

📺 Watch · latest videos

📚 LearnBuilding pi in a World of Slop — Mario Zechner
Featured by Latent Space.
AI Engineer
📚 LearnAI Research Legend’s Honest Assessment of Where We Are
Featured by Latent Space.
Unsupervised Learning: With Jacob Effron
📚 LearnWF2026: Autoresearch & Keynotes ft. Anthropic, Google DeepMind, Amazon AGI, Sonar, Arena, Recursive
Featured by Latent Space.
AI Engineer
The 3 AI Agency Mistakes Keeping You From $20K/Month Retainers
Latest from Nate Herk.
Nate Herk · 893 views · 70 likes
How Boris prompt 1000+ agents overnight...
Get free Claude Code AI Coding course: https://clickhubspot.com/2dd2fb 🔗 Links - Step-by-step workshop in AI builder club: https://www.aibuilderclub.c
AI Jason · 1.1K views · 55 likes
AI, AGI, and ASI
Latest from Matthew Berman.
Matthew Berman · 1.4K views · 60 likes
★ Must-WatchMake information visual with ChatGPT
Turn dense information into interactive visualizations with the Visualize skill in ChatGPT and Codex. Katia Gil Guzman shows how to transform meeting
OpenAI · 42.8K views · 1.6K likes
Intelligence EXPLOSION: Harness Engineering with Pi Agent, Deepseek, and Gemini
The Intelligence EXPLOSION is HERE. 🔥 Five plus model releases in five days, LLM pricing wars in full effect, and most engineers are still locked insi
IndyDevDan · 24K views · 873 likes
The Only Codex Course You Need in 2026 (4.5 Hours)
Latest from Nick Saraev.
Nick Saraev · 15.5K views · 739 likes
The Agent Behind the Curtain: Building the Oz Cloud Agent Platform — Safia Abdalla, Warp
Warp open sourced about three months ago and went from roughly 20,000 GitHub stars to over 60,000, with thousands of pull requests and hundreds of con
AI Engineer · 2.5K views · 40 likes

🗣 Voices & Blogs

★ Must-ReadSimon Willison reads the FT: Anthropic's best model is a hard sell when cheaper tools are this good
August 23. His notes on FT reporting about Fable 5 adoption and revenue: the top model earns its keep on hard problems, but most users route most work to cheaper models. Read together with Drew Breunig below, the two posts explain this month's pricing mood.
📚 LearnDrew Breunig: Fable and the end of the free lunch
For years it was silly to optimize your prompts and harness, because the next model would arrive smarter and cheaper and wash your work away. Fable 5's premium price, with GLM 5.2 at roughly a ninth the cost, ends that. Context engineering and harness quality now pay real money. The best short read of the weekend.
Jason Rebholz: enterprise-managed auth is a start, not a finish line
The security take on today's siren, from the former Corvus CISO. Central on and off switches through Okta are real progress, but there is still no fine-grained control over what a user can do inside a connector once it is on. Turn it on with your eyes open.
A laid-off Indeed engineer built a job-search rival with Claude Code, and shares the numbers
Dreamwork: 4,300+ users, 91 paying, and 3 confirmed hires in its first 4 weeks, built on 1,100+ merged PRs with a pipeline that classifies and enriches about 15,000 job listings a day. A concrete picture of what one person plus an agent stack ships now.
OpenAI built a chip in nine months. Then it let AI rewrite the code.
When OpenAI unveiled Jalapeño, its first custom inference chip, in June, the company made some big promises. The chip, developed The post OpenAI… · The New Stack
Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
Building the Systems Behind AI Agents | Rethinking the Future of Software Engineering | Key Developments Across the AI Frontier · via Berkeley RDI · Agentic AI Weekly

📦 What's Being Built · GitHub

affaan-m/ECC
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first dev · ★243.1K · JavaScript
NousResearch/hermes-agent
The agent that grows with you · ★236.2K · Python
deepseek-ai/deepseek-harness
DeepSeek Harness: Everything is a Plugin. · ★194.5K · TypeScript
firecrawl/firecrawl
The context API to search, scrape, and interact with the web at scale. 🔥 · ★172.2K · TypeScript

🌏 The Wire · Drudge / Breitbart

HUGGING FACE, THE HOME OF OPEN AI MODELS, EXPLORES A SALE AT $13 BILLION OR MORE
Business Insider surfaced the talks Sunday and TechCrunch confirmed the company engaged a bank to weigh inbound interest. It was last valued at $4.5 billion in 2023, and earlier this year it turned down a $500 million Nvidia investment at a $7 billion valuation to avoid one dominant backer. No buyer is named yet.
NVIDIA IN TALKS TO BACK PERPLEXITY AT A $30 BILLION-PLUS VALUATION AS REVENUE TRIPLES
Perplexity's annualized revenue climbed past $750 million from under $250 million at the start of the year, helped by Perplexity Computer, its agent that automates desk work. Nvidia has held a stake since 2023. Talks are preliminary and neither company confirms them.
XPENG'S ROBOT UNIT RAISES $900 MILLION, THE BIGGEST EMBODIED-AI ROUND EVER IN CHINA
The round values the robotics business at over $6.3 billion post-money, led by IDG Capital with Tencent and Alibaba as strategic investors. The IRON humanoid is slated for mass production by the end of 2026, working XPeng stores and campuses first, then shipping in 2027.
A ROBOT RAN 100 METERS IN 9.39 SECONDS, FASTER THAN USAIN BOLT'S HUMAN RECORD
X-Humanoid's two-legged machine edged Bolt's 9.58 from 2009 at the World Humanoid Robot Games in Beijing, where more than 2,000 robots from 16 countries competed. Two years ago humanoids could barely jog. The slope of that curve is the story.
AMAZON QUIETLY RAISED ECHO, KINDLE AND FIRE TV PRICES UP TO 60 PERCENT AS AI EATS THE MEMORY SUPPLY
The Echo Dot went from $49.99 to $79.99 overnight, the base Kindle from $109.99 to $149.99. Amazon blames significant increases in memory and storage costs, the same chips AI data centers are buying up. The AI buildout is now visible in ordinary checkout carts.
TAIWAN INDICTS NINE, INCLUDING NVIDIA AND SUPER MICRO STAFF, FOR SMUGGLING AI SERVERS TO CHINA
Prosecutors say 74 of 130 export-controlled B300 servers reached Chinese customers through Indonesia, Japan and Hong Kong before customs caught the rest. They want five-year maximum sentences for four defendants, including an Nvidia manager described as the key figure.
XIAOMI UNVEILS ITS OWN 3NM PHONE CHIP, PLUS AI AND CAR CHIPS, ALL BUILT AT TSMC
The Xring O3 packs 24 billion transistors and is the first phone chip to support LPDDR6 memory, debuting in the Xiaomi 18 Fold. Alongside it: the O100 NPU for multi-device AI and the D100 automotive chip. One more big buyer is walking away from off-the-shelf silicon.

🤖 Trending Models · Hugging Face

orcarouter/Qwen3.8-27B-Uncensored-MLX
image-text-to-text · ★1.1K · 68.9K dl
orcarouter/Qwen3.8-27B-Uncensored-FP8
image-text-to-text · ★1.1K · 249.7K dl

📈 Markets

NVDA 208.48 ▼2.9%
MSFT 487.31 ▲0.8%
GOOGL 348.06 ▲0.9%
AMZN 262.07 ▲1.3%
META 559.02 ▲1.7%
AMD 456.75 ▼3.5%
AVGO 358.76 ▼2.6%
PLTR 175.89 ▼2.3%
SPCX 135.00 ▼1.4%
TSLA 348.95 ▼3.8%

The BriefAnthropic made Claude's app connections something an IT team can turn on for the whole company at once, through Okta. Workers get tools like Slack, Notion, Datadog, Asana and Figma inside Claude on their first login, with no setup steps of their own. If you run IT or an MSP, this is the week to decide which connectors your people should get, and to read the security caveats before flipping the switch.

Level UpQwen3.8-27B is the local model everyone is testing right now: 27 billion parameters, Apache 2.0 license, reads images and documents, and it just cracked the top 10 on a real web-dev leaderboard. This week, try running it on your own machine. This guide walks through the hardware you need, the quantized versions that fit consumer GPUs, and the setup steps in plain words. specs, benchmarks and the hardware you need to run Qwen3.8-27B locally