🧙🏼 Claude goes online

Howdy, wizards.

Big thanks to everyone who’ve joined context windows in the last weeks.

I plan on doubling down on making this site the best place anywhere to find credible, updated AI case studies. I’ve added 106 new case studies this week as well as launched a suggestions portal where you can post anything from bugs to feature ideas.

IMPORTANT WIZARD ANNOUNCEMENT: This week is your final chance to lock in that juicy 50% discount on lifetime access (code: WBIA-EARLY-BIRD) , so if you’ve been on the fence—it’s time to make the leap. More details at the end of this email.

Here’s what’s brewing in AI this week.

DARIO’S PICKS

1. Claude finally discovers the web

The folks at Anthropic just gave Claude 3.7 Sonnet the ability to do web searches when it needs to. It provides inline citations, so it’s possible to fact-check the results.

Contrary to ChatGPT, which you need to activate web search on when sending your message, Claude determines when to browse the web itself (after it’s been activated in your settings). Hit it with a question about breaking news or current events, and it'll start pulling from sources on the world wide web.

Web search with Claude is currently available for all paid users in the US, but Anthropic is planning an international roll-out soon.

‎ Why it matters‎ ‎ Web search closes a considerable feature gap Claude’s been having with ChatGPT and Gemini.

Letting the model determine when to go online and when not to could be a smart move towards a more intuitive UI — most of us don't want to perform decision calculus on whether our question warrants a trip online or not.

There’s important pitfalls of this approach, though. First, web search with AI is far from flawless (recommend this brilliant article for context). Second, AI companies ignoring website owners’ efforts to block them is resulting in new ways to confuse the AI (see the next news pick for how Cloudflare is doing this).

FEATURED GPT

The Ecommerce AI Website Advisor helps you kickstart your Shopify store with personalized recommendations on the right Shopify plan, apps, and themes.

Check it out here

Want your custom GPT featured? Apply here.

DARIO’S PICKS

2. Cloudflare is routing AIs down a labyrinth

AI companies have been treating websites like an all-you-can-eat buffet. Most web publishers have had their data scraped and regurgitated in ways that effectively reduces their revenue since the user doesn’t visit their site—even when using robots.txt blocking, an honor system that is the internet’s equivalent of a KEEP OUT sign.

Cloudflare—a “middle man” company whose current job includes making about 20% of the world’s websites more secure—has introduced a new solution. Instead of trying to block the AI bots, they'll send them down a rabbit hole of decoy, AI-generated content.

Here’s how it works: when Cloudflare sees suspicious bot activity, it doesn’t try to block access, but rather serves up a maze of AI-generated content instead of the proprietary one. That’s a slow, expensive goose chase for the AI model and users are likely to get poor responses. To avoid contributing to misinformation, Cloudflare aims to provide factually accurate content, just not the proprietary one.

The feature is currently an opt-in for Cloudflare customers.

‎ Why it matters‎ ‎ I’ve been echoing this sentiment for a while—the web urgently needs mechanisms to avoid the information ecosystem collapsing into a self-referential mush of AI-generate garbage.

There’s two distinct approaches developing here: licensing deals and defense methods, like Cloudflare’s hall of mirrors. If the labyrinths succeed in making AI web searches expensive + poorer in quality, I think AI companies could become more a significant notch more motivated to partner with publishers. Yes, we as AI users might temporarily suffer through some bad responses—but at least we’ll be on the path to a more sustainable internet.

UP CLOSE

How AI helps me run my business, part 2

I run two websites and this newsletter next to a full time job. I use AI in my own workflows to handle a bunch of things more efficiently. This is a behind the scenes look at some of those things.

Last week’s part 1 was about how AI helps me with reaching out to sponsors, branding and coding.

Here’s more things I use AI for in my projects:

Writing

AI helps me with writing two very different ways:

1. Writing to communicate factually

When explaining technical issues or documenting workflows, AI generates precise, well-structured communications that would have taken forever to craft manually. When I was stuck with the beehiiv API recently, ChatGPT drafted my support email, and even helped solve the problem when I pasted back the team's response. Pure magic for technical communication!

2. Writing to spark connection

I use AI as part of my creative process—a thought partner—when writing anything important for my websites or this newsletter. For some good tips on this, OpenAI’s article Writing with AI where they interview different creatives remains very useful. I don’t use dedicated AI writer tools at all, I just use Claude Sonnet 3.7 or GPT-4.5.

Keep in mind: co-writing with AI is fundamentally different from things like coding with AI (where your primary goal is just “it works”). I’ve tried both ways and the difference in engagement when I write stuff myself is massive. As cheesy as it sounds, the ability of your authentic voice to spark connections with the right people is important.

Transforming information

AI turns unstructured data into exactly what I need. Every bit of cognitive energy not wasted on reformatting data is brain power I can redirect toward actual creative thinking.

Some ways I apply this:

• Making sense of my scribble: My handwritten notes are chaotic — text squeezed into margins, arrows everywhere, really unhinged stuff. I snap a photo, send it to ChatGPT and ask for a transcription and coherent outline.

• Helping me fill out forms from hell: I've used this for everything from tax applications to applying for a visa in a foreign language. Instead of figuring out "this goes there, that goes here," I show AI my source documents (business details, passport info, etc.) and the form. And then I watch the cognitive tax of matching information evaporate!

• Creating tables: People love a good table—it enables you to get an overview of a lot of information quickly. I often use AI to turn bits and pieces of disparate information into a table.

Here’s a specific example of how AI helps me transform information (feel free to skip if you already got the point here):

a16z has a new article with a list of the top GenAI apps. In this article, they show an image of the logos of top 50 products in ranked order, followed by a wider breakdown of consumer AI trends.

Let’s say I want a structured overview with insights about the companies mentioned in the article body, with their respective rank from the top 50 list. That would require taking unstructured data (an image and an article) and combining them to create a structured output (a table).

You can adapt this approach to many scenarios:

"Here's [paste/upload something and describe your unstructured content]. Context: [brief background if needed]. Please transform this into [describe desired output structure with specific fields].”

And a pro tip: Say “give me the output as tsv” at the end. You’ll be able to just hit this button and paste the output directly into a spreadsheet:

For creating tables like this, I’ve started using reasoning models like o1 or Claude 3.7 Sonnet with extended thinking, because they feel more precise (like they double check things), and that seems effective when transforming information from one format to another.

It’s also a bit of a balancing act to not overload them in a single one prompt, but do this a few times and you begin to understand how much they’re able to handle at a time.

Data enrichment

I often have a dataset which lacks some information and, as long as it’s freely accessible online, AI has a good chance of finding it. In the screeshot below I started out with a list of company names, and asked Perplexity for their websites, social media account, industry and number of employees:

Enriching data will be a growing use case I think as the top models start integrating proprietary data sources better (like Perplexity with Crunchbase and Factset) + making browser tools widely available (Operator, Computer Use, etc) which lets users instruct the AI to log-in to sites.

Beware of hallucinations when using web search, though. Models tend to get confused part of the time. Whatever information you’re collecting should either be easily verifiable by you, or a scenario where accuracy isn’t that important.

- - -

If you like these tips & tricks on how I personally use AI, hit reply to let me know—and I’ll make more of it.

CONTEXT WINDOWS

Fire your AI consultant

If you're working with a company looking to implement AI in any capacity, be it as an external advisor or driving innovation internally — you're navigating a fast-changing foggy landscape where mistakes are expensive.

That’s why I created context windows, the most complete database of 600 700+ AI implementation case studies on the web. It gives you a categorized and up-to-date view of case studies released by all major AI labs.

It helps you save tons of time on research, spot trending use cases in your industry long before your competitors and — through proven, real-world examples — enables better decision making for your AI strategy.

Why is most content locked? Simple. I'd rather you pay me directly than watch a consulting firm charge you their monthly coffee budget ($20,000+) to serve you the same ideas wrapped in corporate jargon and slides with misaligned text boxes.

There’s only two ways to get access:

• Refer 6 friends to this newsletter (they’ll thank you for it) and earn lifetime access. Check you referral count below.

• ⏰ LAST CHANCE: Get lifetime access with the code WBIA-EARLY-BIRD for $49.99 (50% off).

The forecast shows a 100% probability that this deal will be gone by the end of the week, so grab it while it’s here.

THAT’S ALL FOLKS!

Affiliate disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes link to products and other newsletters. Please assume these are affiliate links. If you choose to subscribe to a newsletter or buy a product through any of my links then THANK YOU – it will make it possible for me to continue to do this.