July 17th-July 23rd // Estimated Reading Time: 8 minutes

The most popular AI bootcamp in the marketing industry is BACK! The AI Accelerator helps marketers level up on all things AI for marketing and advertising so they can make a meaningful impact on their business, fast.

We’re thrilled to announce Blis, the T-Mobile Advertising Solutions omnichannel DSP, as our exclusive sponsor for this run!

Enroll now using promo code EarlyBirdsLearnAI for 20% off!

Chinese AI model Kimi K3 catches up to Claude and ChatGPT
Source: Associated Press
July 17th, 2026

Summary: Beijing startup Moonshot released Kimi K3, an open-source AI model with performance on par with Claude and ChatGPT’s latest models. K3 costs more to run than past Chinese models, but it's half the price of OpenAI's top model, GPT-5.6 Sol. The release follows Zhipu's GLM-5.2 last month, another open-source Chinese model that developers say also performs close to top US models at a lower price. The Information reported that Microsoft is testing Kimi K3 for Copilot, which currently runs on Anthropic and OpenAI models.

Opinion: US export controls forced Chinese AI developers to build for efficiency, leading many US companies to adopt the Chinese models to cut AI spend. That backfired! The underlying code and weights may be freely available, but US lawmakers have flagged concerns about data security and Beijing's influence over models that could end up embedded in American software and infrastructure. For marketers considering tools or vendors, weigh the cost savings against those risks, especially if the tool touches sensitive client or company data.

OpenAI models escaped secure testing environment and hacked a company to cheat on a test🔒
Source: Fortune
June 21st, 2026

Summary: OpenAI was running a test to see how good its models are at hacking. OpenAI had turned off some of the model's usual safety restrictions and locked it in a closed sandbox environment with no internet access. GPT-5.6 Sol and a more powerful, unreleased model found a previously unknown flaw in the software, allowing it break out and connect to the internet. Once out, the model figured out that Hugging Face, a company that hosts AI models and datasets, probably had the test's answer key stored somewhere on its servers. So it hacked into Hugging Face's systems and took it to cheat on the test. Hugging Face caught the break-in on its own before OpenAI said anything, and when its team investigated, they used a Chinese AI model instead of a US one, because the US models' built-in safety restrictions kept blocking the questions their investigators needed to ask.

Opinion: Give a model a narrow goal with no guardrails, and it will find the fastest path there, rules be damned. It's what happens whenever a system optimizes for a score instead of the behavior the score is supposed to measure. Marketers using AI for lead scoring, ad buying, optimization, or performance targets should consider a similar question: What would this model do if "hitting the number" and "doing it the right way" conflicted?

Alphabet and Tesla test Wall Street’s patience as AI spending overshadows growth
Source: CNBC
July 22nd, 2026

Summary: Alphabet and Tesla are pouring massive sums into AI infrastructure, enough to push both companies' free cash flow into negative territory in Q2. It’s the first time Google has ever had negative cash flow as a public company. Alphabet now expects to spend $195-205B this year, up from its earlier forecast of $180-190B. Tesla's spending jumped 142% to $5.79B for the quarter. Alphabet's free cash flow turned to negative $5.9B, compared to nearly +$25B a year ago, while Tesla's fell to negative $1.1B from +$146M last year. Both companies beat revenue estimates, including an 82% jump in Google Cloud revenue. Shares still dropped after hours, with Tesla down 4% and Alphabet down more than 3%. Could be a warning sign for other big AI spenders releasing their earnings reports in the coming weeks, including Meta, Microsoft, and Amazon.

Opinion: The “negative cash flow” headline is misleading, especially in Google’s case. AI is a generational opportunity, Google is clearly capitalizing on it, and in order to continue capitalizing on it, it must continue to invest. That’s what free cash flow should be used for anyways; smart investment. AI infrastructure investment is as smart a bet as anything out there. The real headline here should be Google’s eye-popping growth numbers, especially in areas like Cloud (+82%), Search (+17%), and YouTube (+13%).

Google launches Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

What It Does: Google released three new Gemini models built for balance efficiency, speed, and quality. 3.6 Flash is the general-purpose workhorse, better for coding and knowledge work while using 17% fewer tokens than its predecessor. 3.5 Flash-Lite is made for speed and high-volume tasks like document processing. 3.5 Flash Cyber is a specialized model for finding and patching security vulnerabilities, but it's only available to governments and vetted partners for now.

Quick Take: The AI race isn't just about building the smartest model anymore. With new Chinese AI models entering the picture, it's also about making good-enough models, purpose-built for specific use cases, cheap enough to run constantly. Google, OpenAI, and Anthropic are all chasing that efficiency curve now. If you've deployed AI agents, this round of releases and future releases are worth doing regular price comparisons against whatever models you're using now.

The Setup: Before the average car buyer walks into a dealership, they've already spent 14 hours researching vehicles online. Hyundai wanted to drive purchase intent, but its video ads were often buried within cluttered webpages. Hyundai and its agency, Canvas Worldwide, wanted to test whether better visibility could get more people to actually visit Hyundai's vehicle details page before going to a dealership. 

The AI Solution: Hyundai decided to use Chalice AI's ad decisioning models through OpenX's infrastructure via the OpenXBuild Real-Time Bidstream API. Instead of bidding across every page on auto websites, the AI scored each individual page in real time for quality, factoring in things like page clutter and how visible the ad would actually be. Only high-scoring pages got a bid; everything else was filtered out.

The Results:

• 67% lower CPMs for online video

• 20% lower cost per vehicle details page visit

Why This Matters: Cheaper impressions and better outcomes aren't the same thing. Hyundai's gains came from being pickier about which impressions it bid on, not from casting a wider net. That suggests that impression-level decisioning based on page quality can move performance as much as audience targeting.

Your Action: Chalice AI's model scored individual pages in real time using signals like clutter and viewability, then fed that score into the bidder before any bid was placed. Ask your DSP two specific questions: does it support bidstream-level enrichment from a third-party scoring model today, or only post-bid measurement? If you have log-level access, you could build a simple version of this yourself, at least to start. 

Pull your last 30 days of log-level bid data, filtered to your top 200-300 URLs by spend. Run a headless browser and screenshot each URL as it actually renders, ads and all. Feed those screenshots to a vision model with a simple rubric: how much of the page is ad units versus content, is your slot above the fold, does the layout look like something a human would tolerate or something they'd bounce off of. Score every page 1 to 10.

Now cross that score against your actual performance data. Where is the line? Is there a score below which CPA falls apart or view-through rate craters? Turn your scored list into an inclusion list, and push it to your DSP as a targeting list, refreshed monthly.

If the results are strong, that's your business case for pursuing an enterprise-grade version.

Stagwell is building its own AI media curation marketplace🔒 - Stagwell Curate is a Claude-powered AI platform that curates CTV, video, display, and audio ad inventory marketplaces. The goal: bypass curation fees and reduce time spent managing multiple SSP and DSP marketplaces.

'No longer AI losers': Havas leans into AI-first identity as North America revenue grows🔒 - CEO Yannick Bolloré says Havas's Converged.AI system has helped the agency holding company retain clients. Its organic revenue was up 2.5% in Q2, with North America net revenue up 6.9% for the first half of the year.

Google gives publishers an opt-out for AI Overviews - Google is rolling out a new Search Console control that lets publishers keep their content out of AI Overviews, AI Mode, and Google Discover's AI features, without losing rankings in traditional search.

Anthropic's landmark $1.5B copyright settlement gets final approval - Authors and publishers will get $3,000 per work across roughly 500,000 works. The payout stems from Anthropic downloading pirated books from sites like Library Genesis to train its models.

Google is reportedly designing a new AI chip to make Gemini more efficient - Google is said to be building a new chip, internally called "Frozen v2," that could run its Gemini models six to 10 times more efficiently than its current chips.

The U of Digital AI Edge Newsletter is intended for AI Edge Pro subscribers, but occasional forwarding is okay!

To subscribe visit Uof.Digital/Newsletters/ai-edge or AI Edge Pro, or contact us directly for group subscriptions.

And remember, U of Digital helps teams drive better outcomes through structured education on critical topics in marketing / advertising like AI, mar tech, programmatic, CTV, commerce media, and more. Interested in learning more about how we can supercharge your team? Let’s talk!