Welcome, humans.

This was a huge week in AI news and it all basically happened on Tuesday. We’re still trying to wrap our minds around the landscape shift...

To sum it up:

DeepSeekV3.1 is now basically the smartest non-reasoning model…

Gemini 2.5 is now the best overall AI (because the popular way to benchmark AI now is with video games, here’s 18 demos of what it can do)…

And OpenAI is now the top image AI generator—things got so poppin’ that Sam had to put on rate limits to stop the GPUs from melting.

Also, everything is Studio Ghibli style now. Example:

As amazing as it would be if the ACTUAL Studio Ghibli made LOTR, it’s worth keeping in mind that the studio’s leader, Hayo Miyazaki, probably hates this.

In fact, this moment is like the complete opposite of his world view come to life (he famously called an AI animation “an insult to life itself”). He puts a TON of work into making each Studio Ghibli film, as this documentary shows. And it’s worth it.

In true Miyazaki form, we can’t help but wonder if there’s a better balance to be struck between empowering individuals with incredible digital power at scale, and respecting the hard work of the artists whose work these machines were trained on.

After all, animation is an incredibly grueling field (especially anime, which is full of exploitation) where artists are already underpaid and overworked.

Animators could probably use more creative tools to make their own work easier and faster—so why aren’t AI companies building more tools artists actually want?

Here’s what you need to know about AI today:

• We dive into the rise of AI shopping.

• OpenAI improved 4o for paid users.

• Google announced five AI-enhanced travel planning tools.

• Anthropic dropped two new papers that explain how AI think.

AI shopping assistants are taking over your Shopping Cart…

Amazon, Perplexity, and Adobe all agree: AI shopping assistants are about to change how we buy stuff online. Like, dramatically.

Adobe Analytics just released a bombshell report showing traffic to U.S. retail websites from generative AI sources has jumped a mind-boggling 1,200% since July 2024. That’s data based on 1T visits to US retail sites.

During last year's holiday shopping season, there was a 1,300% increase compared to the previous year—with Cyber Monday seeing an insane 1,950% year-over-year spike.

Adobe also surveyed 5K consumers directly to find out how they use AI.

The data shows AI is becoming our go-to research buddy:

• 39% of U.S. consumers have already used generative AI for online shopping.

• 53% plan to do so this year.

• 92% of those who've tried it say it enhanced their shopping experience.

• 87% are more likely to use AI for larger or more complex purchases.

Once shoppers arrive from AI referrals, they're more engaged—viewing 12% more pages per visit, with an 8% higher engagement rate and a 23% lower bounce rate than visitors from traditional channels.

Not all categories are created equal, though. Conversion rates are highest in electronics and jewelry, but lowest in apparel, home goods, and groceries.

This makes sense—TVs and watches have more objective specs (“find me a 65-inch TV under $1,000”), while clothing and groceries are more subjective and sensory.

So what tools are they using? Adobe’s report didn’t say, but here’s a few of the tools we know (and here’s a few others we just found):

• Perplexity Shopping: Offers one-click “Buy with Pro” checkout directly within search results, with shopping queries up 5x since November and partnerships with 150+ merchants in the works (apparently it now works with Firmly).

• Agora: An e-commerce search engine specifically for 25K small retailers (like Shopify store-fronts)

• ChatGPT: People are also using GPT Search and Deep Research to help them shop online (including get coupon codes).

Not to be left out, Amazon's also building an entire ecosystem of AI shopping tools:

• Rufus: Amazon's shopping chatbot that answers product questions, compares items, and makes recommendations based on your browsing history.

• Interests: A new AI feature that continuously scans inventory for new products matching what you're passionate about, alerting you to relevant deals and restocks.

• Health AI: An AI assistant to answer wellness questions and suggest products with “clinically verified” badges for answers reviewed by licensed clinicians.

Amazon's ultimate plan is to integrate these AI assistants with Alexa+, creating a seamless ecosystem where your voice assistant can instantly call the appropriate AI tool based on your request.

Our take: The entire product discovery process is changing from endless scrolling to concise, personalized recommendations.

While Google dominated product search for decades, these new AI assistants from Amazon, Perplexity, and others threaten to reshape the entire e-commerce landscape—and the data shows consumers are ready to embrace the change.

Prompt Tip of the Day

The best AI users don't hunt for perfect prompts—they build systems to manage them. In this 16-minute tutorial, Jeff Su shows how to create a Notion database that captures, organizes, and categorizes your AI prompts without disrupting your workflow.

Treats To Try.

1. Qwen2.5-VL-32B interprets your images with better reasoning and more detailed analysis than previous models (along with improved mathematical reasoning).

2. Ideogram 3 is an image generator with a new Style References feature that lets you upload images to control generation aesthetics and a Random style exploration feature to pick from billions of presets.

3. T3 Chat is an $8 a month chatbot to run fast, cheap AI models like DeepSeek and Gemini flash.

4. Zen is a web browser alternative to Chrome thats free and open-source.

5. Stripo helps you create customizable email templates with drag-and-drop building that you can easily export to your email platform

6. If you use Ableton to make music, someone made a tool to connect it to Claude so you can make music with basic prompts (demo).

7. DeepLearning.AI from Andrew Ng is a great resource for AI courses—just go to courses, and check “short courses” for “beginners.”

See our top 51 AI Tools for Business here!

Around the Horn.

• OpenAI released a new version of 4o for paid users that’s better at following instructions, should be better at coding and creative tasks, and also uses fewer emojis—we’ll report back if it ends up being any good!

• NVIDIA released Project G-Assist 0.1, a local AI assistant for GeForce RTX PCs that optimizes performance through voice commands using an on-device Llama model.

• Google announced five new or updated tools across Search, Maps, Lens and Gemini to help users plan and enjoy summer vacations.

• Character.AI tested a feature that gives parents insights into their kids’ chatbot usage, including conversation topics and time spent chatting.

• Butterly Effect, the maker of viral AI agent Manus wants to raise a new round at a $500M valuation.

FROM OUR PARTNERS

How to speed up AI delivery in healthcare—reliably

Building healthcare AI solutions? You'll want to read this case study. This healthcare company used Vellum's release tools to reliably ship AI updates 40% faster and earn clinician trust. 

Get it here.

Intelligent Insights

1. Anthropic just published two new papers revealing how they can now trace Claude's internal “thoughts.”

   1. Paper 1 = Circuit Tracing, which describes how they replaced Claude's neurons with simpler puzzle pieces to show what it pays attention to.

   2. Paper 2 = On the Biology of a Large Language Model, which applies this method to Claude 3.5 Haiku.

2. Anthropic also released more data on how people are using Claude (you can search the data here).

3. A new study found that CheXzero, a popular AI model used to scan chest x-rays for diseases, failed to detect potentially life-threatening conditions more often in marginalized groups (particularly Black women).

4. We love this take on where vibe-coding is going, titled “DIY apps are the New Myspace”—because software is now cheap enough to develop without it needing a $1B exit to pay off, a wave of niche, custom software is now possible.

5. We also like this blunt take from physicist Sabine Hossenfelder, warning that we’ll see how vastly overvalued language model providers are once a fundamentally different AI approach inevitably outperforms them.

6. This is a great (if slightly technical) overview on AI agents from Latent Space. Also, Swyx (of Latent Space) noted that the time between major state of the art model releases is shortening…

7. CoreWeave’s IPO today will be a good test of AI hype vs reality—buckle up!

A Cat's Commentary.