Clear communicators aren't lucky. They have a system.

Here's an uncomfortable truth: your readers give you about 26 seconds.

Smart Brevity is the methodology born in the Axios newsroom — rooted in deep respect for people's time and attention. It works just as well for internal comms, executive updates, and change management as it does for news.

We've bundled six free resources — checklists, workbooks, and more — so you can start applying it immediately.

The goal isn't shorter. It's clearer. And clearer gets results.

Get the free toolkit

Hey there,

Hope you had a great Christmas! We’ve got industry giants making surprising moves and some changes coming to our favorite tools that… well, we might not love.

Let’s dive in.

Today:

• Groq Teams Up with Nvidia 

• ChatGPT Might Get Ads Soon

• Qwen Fixes AI Image Consistency 

• Nvidia to Ship Top Chips to China 

• TikTok Owner Bets $23 Billion on AI

ChatGPT responses… with ads inside them? 😬

A report says OpenAI is actively discussing ad formats for ChatGPT — including sponsored content woven directly into answers, plus possible sidebar ads next to the response. One example in the report: ask for mascara recommendations, and a sponsor like Sephora could surface in the “best options.” 

What’s especially spicy: another idea mentioned is ads that appear after you ask follow-up questions (like clicking into travel suggestions), and OpenAI reportedly told The Information they’re exploring how ads might work “without compromising user trust.” 

And the trust angle is the whole story here. The same report points out Sam Altman has previously called ad-shaped AI responses “dystopian,” particularly if the recommendations are influenced by your prior chats — but the report suggests OpenAI may be looking at ads powered by ChatGPT’s memory (targeting based on conversation history).

Why this matters:

• If ads get “blended” into answers, the line between helpful recommendation and paid placement becomes a constant question.

• The real risk isn’t “ads exist” — it’s whether users can clearly tell what’s sponsored, and whether personalization crosses into creepy territory.

Groq ↔ NVIDIA: a licensing deal that says “inference is the battlefield” ⚙️

Groq announced (Dec 24) it signed a non-exclusive inference technology licensing agreement with NVIDIA aimed at scaling “high-performance, low cost inference.” 

The headline detail that jumped out: Groq says Jonathan Ross (founder), Sunny Madra (president), and other Groq team members will join NVIDIA to help advance and scale the licensed technology — while Groq remains independent.

They also note a leadership change: Simon Edwards steps into the CEO role, and GroqCloud continues without interruption.

Why this matters:

• This is another sign that inference speed + cost is where the real competition is right now — not just who trains the biggest model.

• “Non-exclusive” is key: it’s not a full acquisition vibe, more like NVIDIA grabbing a proven approach and industrializing it at scale while Groq keeps operating.

Qwen’s image editing update: better identity consistency (finally) 🖼️

Qwen dropped an updated image editing model (Dec 24) called Qwen-Image-Edit-2511 on Hugging Face, positioned as an upgrade to Qwen-Image-Edit-2509. The big claim: it holds facial identity much better during edits — including group photos with multiple people. 

They also mention broader improvements (lighting control, camera angles, product design, geometric calculations) and say popular community LoRAs are baked into the base model now. It’s released under Apache 2.0, and there’s a Hugging Face demo plus free testing via Qwen Chat.

Why this matters:

• If you’ve ever tried “change outfit / change setting / keep the same person” and gotten a totally different face… yeah. This update is aimed directly at that pain point.

Apache 2.0 licensing makes it easier for developers/teams to experiment and build workflows around it.

🧠RESEARCH

SemanticGen: Video Generation in Semantic Space

Current video AI creates clips piece-by-piece, which is slow and computationally heavy. This new method, "SemanticGen," first sketches the video’s overall meaning and layout before filling in the visual details. By separating the big picture from the fine print, it generates higher-quality, longer videos much faster than today's best models.

Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies

Researchers found that AI models process information in distinct stages: early layers explore ideas while later layers refine the final answer. They developed "Bottom-up Policy Optimization," a training technique that strengthens these foundational early layers first. This approach helps the model build better reasoning skills, significantly improving its performance on complex puzzles.

SpatialTree: How Spatial Abilities Branch Out in MLLMs

This study organizes how AI sees the world into levels, ranging from basic perception to complex navigation. The authors discovered that forcing AI to "think" too much can surprisingly hurt its ability to simply see clearly. Their new "auto-think" method lets the model intelligently decide when to pause and reflect, fixing this issue.

🛠️TOP TOOLS

Each listing includes a hands-on tutorial so you can get started right away, whether you’re a beginner or a pro.

Banana AI – GPUs For Inference - inference hosting platform built for high‑throughput AI workloads.

BandLab Songstarter – Make Music Online - cloud‑based digital audio workstation (DAW) and music community where you can record, edit, mix, master, and share songs on the web or mobile

Banter AI – AI Business Phone Calls - automates inbound and outbound phone calls with an AI receptionist that answers questions, books appointments, and routes calls using your business’s own details.

📲SOCIAL MEDIA

🗞️MORE NEWS

Nvidia Chips in China Nvidia plans to start shipping its powerful H200 chips to China by mid-February, though it still needs final government approval. This move follows a policy shift that allows these sales with a 25% fee, reversing a previous ban on the technology. However, it remains uncertain if officials will give the green light for the deal to go ahead.

ByteDance's Big Spending TikTok’s parent company, ByteDance, is reportedly planning to spend $23 billion next year to build the computer systems needed for AI. Most of this massive budget will go toward buying advanced chips and data centers to power their smart tools. This aggressive investment highlights the intense global race to secure the hardware necessary for future technology.

Italy vs. Meta Italy’s competition regulator has ordered Meta to stop blocking rival AI chatbots from operating on WhatsApp. Officials argue that Meta is unfairly restricting consumer choice to favor its own services. Meta plans to appeal the decision, stating that its messaging platform was not designed to host outside AI tools.

Waymo’s New Assistant Waymo is testing a voice assistant powered by Google’s Gemini to act as a helpful companion in its self-driving cars. Leaked documents show the AI can control features like the air conditioning, but it is strictly forbidden from discussing accidents or pretending to be the driver. The system is designed to be a polite, limited butler that keeps clear boundaries for safety.

Authors Sue AI Giants Six authors have filed a lawsuit against major tech companies, including Google and OpenAI, for allegedly using pirated books to train their models. Unlike a standard group lawsuit, these writers are suing individually to seek much higher damages of up to $150,000 per book. They claim the companies stole their work from illegal websites to build profitable products without paying.