AI Titus's Morning Wire

AI NEWS TODAY

MICROSOFT SWITCHES OFF THREE COPILOT FEATURES TOMORROW, DOWNLOAD WHAT YOU WANT TO KEEP TODAY
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadDeepSeek open sourced the whole runtime its agents run on, and every part is swappable
MIT licensed, built on the Cordis framework. Models, tools, skills, sandboxes and the agent loop itself are all plugins you can replace. If you have wanted to see inside a working agent stack instead of reading about one, this is the whole thing.
📚 LearnA 1B model matched much bigger ones using only data it had permission to use
DFM Mimir v1 was post-trained on 161 permissibly licensed datasets and still beat the original 1B model it was built from across 20 benchmarks. If your blocker on AI is where the training data came from, this is the first result that speaks to it directly.
You can talk an AI reviewer into a better score without changing the science
Across 4,200 conference submissions, rewriting how evidence and novelty were framed moved the scores an AI reviewer gave. Low scores went up and high scores came down. Anywhere a model is grading work, the writing is part of the grade.
★ Must-ReadThe models everybody argues about are not the models anybody runs
Hugging Face counted it: models under 1 billion parameters are 83 percent of all downloads ever, and models over 100 billion are 1 percent. Quantized GGUF repos grew 464 percent. The real world is running small models on modest hardware.
Nobody can reliably spot AI-generated disaster video, and neither can the detectors
On a new benchmark of 17,886 videos, no detector family held up outside its training data. The clips that fool people fool machines too, and sharing on social media makes it harder, because compression strips the evidence.
AI research agents are good engineers and not yet researchers
Seven frontier models across 36 long tasks recombined techniques they already knew instead of inventing new ones, and the same task gave noticeably different results run to run. A useful calibration if you are being sold autonomous discovery.

📺 Watch · latest videos

📚 LearnVercel accuses Cloudflare of stealing
Featured by Latent Space.
Theo - t3․gg
📚 LearnWF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF
Featured by Latent Space.
AI Engineer
📚 LearnAgents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song
Featured by Latent Space.
AI Engineer
FIXING Opus 5: PROOF that Prompt Engineering IS NOT DEAD
Opus 5 is one of the BEST models ever released, and one of the WORST to talk to. Verbose, "load-bearing" everywhere, torching your output tokens. So w
IndyDevDan · 721 views · 51 likes
The Ultimate Guide to Making Your Entire Development Cycle AI Native
Latest from Cole Medin.
Cole Medin · 475 views · 42 likes
AI News: ChatGPT Ultrafast, Grok 4.6, 3 New Open-Source Models, and more!
Latest from Matthew Berman.
Matthew Berman · 57.3K views · 1.5K likes
★ Must-WatchPreviewing Ultrafast mode: GPT‑5.6 Sol at up to 14X the speed
Today we’re previewing Ultrafast mode, a new service tier in the OpenAI API for GPT‑5.6 Sol. Powered by Cerebras, Ultrafast runs up to 14x faster than
OpenAI · 26.5K views · 627 likes
How Web Data Infrastructure Powers the Next Generation of AI — Patricija Žemaitytė, Oxylabs
Minutes into a call to demo a search API rebuilt to answer in under a second, the system got blocked, badly, in front of the client. Patricija Žemaity
AI Engineer · 1.8K views · 28 likes
Never Miss a Customer Call Again (AI Receptionist)
Latest from Nate Herk.
Nate Herk · 5K views · 123 likes
★ Must-WatchWhat does AI actually know about you?
The information you share with AI only travels as far as you let it. Zoe from the Anthropic education team explains what happens to your information w
Claude / Anthropic · 2.6K views · 81 likes

🗣 Voices & Blogs

📚 LearnSimon Willison On The Local Model That Thinks Itself To Death By Default
He likes Qwen 3.8 27B, a 17GB model with real vision and tool calling, with one catch worth knowing before you install it: leave reasoning effort at the default and it burns tokens deliberating over trivial questions. Turn it to low.
★ Must-ReadNathan Lambert On How Chinese Labs Keep Pace, And Why It Is Not Copying
Six concrete reasons, including release cycles measured in days rather than months and a deliberate choice to skip multimodal so the rest is easier to optimize. Ends on a warning that hacking ability is spreading into small models.
★ Must-ReadBen Thompson On Why A Payments Company Would Want The Layer That Picks Your Model
Reads the Stripe and OpenRouter deal as a bet that models become interchangeable and the router owns the customer. Paywalled after the opening, and the opening carries the argument.
📚 LearnSebastian Raschka Builds An AI Text Detector, Then Explains Why Detection Loses
The full build, a fine-tuned DistilBERT classifier with an API and a UI, followed by the honest part: the next model simply stops showing whatever pattern you trained on, and the false positives land on human writing.
Per-developer environments were the goal. Agents moved the goalposts.
Multi-tenancy has moved in one direction for 60 years: the tenant keeps getting smaller. Mainframe time-sharing carved a single machine The post… · The New Stack
Generative design of bacteriophages with genome language models
Agentic AI Summit 2026 Recap | Five Shifts Defining the Future of Agentic AI | Key Developments Across the Agentic AI Ecosystem · 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 · ★240.6K · JavaScript
NousResearch/hermes-agent
The agent that grows with you · ★231.8K · Python
firecrawl/firecrawl
The context API to search, scrape, and interact with the web at scale. 🔥 · ★168.4K · TypeScript
langchain-ai/langchain
The agent engineering platform. · ★144.4K · Python
browser-use/browser-use
🌐 Make websites accessible for AI agents. Automate tasks online with ease. · ★109.5K · Python

🌏 The Wire · Drudge / Breitbart

STRIPE REPORTEDLY BUYS OPENROUTER FOR $7 BILLION, FIVE TIMES ITS MAY PRICE
OpenRouter raised at $1.3 billion in May. Neither company has confirmed. If your app routes model calls through it, your vendor is becoming a payments company.
GOOGLE SWITCHES OFF THREE IMAGEN 4 ENDPOINTS TODAY
Anything still calling imagen-4.0-generate-001, ultra or fast stops working. The migration target is gemini-3.1-flash-image and it is not a drop-in replacement, so budget time rather than a find and replace.
GROQ RAISES $350 MILLION AT HALF ITS OWN 2025 VALUATION
$3.5 billion, down from $6.9 billion, after Nvidia licensed its technology and hired its chief executive. Nvidia is now an investor in the round.
DYNATRACE PAYS $915 MILLION FOR AI EVALUATION COMPANY ARIZE
$815 million of it in cash. Monitoring vendors buying AI evaluation is a signal about what enterprises will insist on before they let an agent near production.
ANTHROPIC RAISES ITS OWN MISALIGNMENT RISK RATING FROM VERY LOW TO LOW
The same report discloses that human-feedback vendor traffic ran without its blocking bio classifiers from May 2025 to April 2026, across 133 million exchanges. A company grading itself down in public is unusual enough to note.
HIGGSFIELD TAKES $400 MILLION AT $5.4 BILLION ON $700 MILLION ANNUALIZED REVENUE
DST Global led. The AI video company says 390 of the Fortune 500 already use it, which is a real revenue number rather than a promise.
A16Z PUTS $40 MILLION INTO THE COMPANY THAT GRADES EVERYONE ELSE'S MODELS
Vals AI valued at $400 million, revenue up eightfold over 2025, and its results already show up in model cards from OpenAI, Anthropic, Google, Meta and xAI. Its new tool builds a coding benchmark out of your own repository.

🤖 Trending Models · Hugging Face

Qwen/Qwen3.8-2.4T-A95B
text-generation · ★1K · 9.5K dl

📈 Markets

NVDA 225.16 ▼0.1%
MSFT 495.40 ▼0.3%
GOOGL 345.90 ▼0.1%
AMZN 262.65 ▼0.9%
META 589.85 ▼0.9%
AMD 514.39 ▲6.5%
AVGO 392.99 ▼5.9%
PLTR 174.04 ▼2.8%
SPCX 140.00 ▼0.9%
TSLA 342.27 ▲0.7%

The BriefMicrosoft is folding its two Copilot apps into one, and on Tuesday August 18 three features stop working: Podcasts, Group Chat, and Deep Research. Anything you made with them has to be downloaded before the switch or it goes away with the feature. The one that will sting is Deep Research, the tool that reads a lot of sources and writes you a report. It is being replaced by something called Researcher, and Researcher sits behind a paid Microsoft 365 subscription, so a job that was free yesterday now has a price. If you run Microsoft 365 for other people, today is the day to tell them, because tomorrow it is a support ticket instead of a heads up.

Level UpSpend fifteen minutes today on the one Copilot change that costs money. Open Microsoft's own page on Deep Research and read what moved into Researcher and what tier it now needs. Then do two things: export anything you or your clients still want out of Deep Research, Podcasts or Group Chat, and check which of your Microsoft 365 licenses actually include the replacement. That second check is the one people skip, and it is the one that turns into an angry call on Wednesday. Deep Research in Microsoft Copilot, and what replaces it

Caught Up?
  • Google shipped Gemini 3.7 Flash at half price through December 31, then it doubles on January 1.
  • DeepSeek turned on surge pricing Sunday. Peak hours cost double and some rates are up fourfold.
  • Alibaba's Qwen passed 3 billion downloads and is now the most-used open model family, ahead of Meta and Google.
  • DeepSeek open sourced its whole agent runtime under the MIT license, every part swappable.
  • Google shuts off three Imagen 4 image endpoints today, and the replacement is not a drop-in swap.