I’m going on record saying that Claude’s new web search functionality is a quantum leap forward in AI’s utility for everyday users. It has, for me, shattered the frame of reference of what LLMs are “good at” and “not so good at”.

How it Was

I used to think of LLMs as talking to someone who has read every Encyclopedia Britannica (or Wikipedia, I guess…) article ever written up until about 2022. That was amazing, but it limited their utility for anything that required real-time information.

Reimagining Capabilities

Now, a new world of interesting use cases has emerged. I mentioned online shopping in a previous post, but this is just one trivial example. Many users will now be using Claude to curate daily news for them, and they’ll be able to use Claude to contextualize that news as well.

I am already seeing significant improvements in Claude’s ability to make near-term predictions. It took some prompt refining, but I was able to get Claude to predict the date of the last frost in my area this spring (April 24th–later than I was hoping but oh well) so I could plan my summer garden. In forming its prediction, Claude drew on multiple different sources, including the latest meteorological data, published expert opinions, and available prediction models from the web.

Structural Payoffs

I predict that web search functionality will, in most cases, essentially “fix” the lingering hallucination problem that everyone seems so keen to cite as a reason not to use AI. With web search, you can not only ask Claude to check its answers against credible online sources, but you can also ask it to cite/link those sources so you can review them yourself. I’ve tried this out with several prompts in various areas, and it does incredibly well.

See here a conversation I had with Claude about the future of solar energy. Very impressive. AI is no longer the black box that it used to be–it can start to support its answers in a way that is intelligible to humans and allows for detailed human fact checking.

Project Knowledge (see here for my earlier post), as a tool, is going to become less important. A lot of the heavy lifting of AI use is about giving the LLM the right amount of context for the discussion you want to have. Much of that context is freely available online, and prompts can now simply ask Claude to survey the available information on a topic before responding.

For example, I used to upload a PDF of any book I wished to discuss with Claude so that it could cite specific examples from the text. With web search, I can simply ask Claude to find an available PDF, read it, and use it as reference material. This may sound small, but the implications across more complex tasks are significant. Also, there are some serious limitations around what is available in a format Claude can digest, but now more online sources specifically formatted for and targeted at AIs will emerge.

How it Almost is…

It’s almost like Claude has read Britannica/Wiki up to today. I’m no longer talking with a genius who has been living under a rock for the past 3 years.

…but it’s not like that. Not yet, at least. Web search is a specific module/capability within Claude, much like the Analysis tool and other similar widgets. You don’t always have to specifically prompt Claude to search the web, but unless it does, you’re still talking to the same old Claude as before.

But that was already pretty good! And this is better! And you can now simply ask it to do a quick web search to get up to date information. No doubt, Anthropic will make that functionality native soon, but until then, it’s really not a pain to ask.

Conclusion

I am not an AI expert, so I will state it as inelegantly as I please: the game has very much changed. This latest development calls for a reevaluation of what you want from AI in light of what it’s now capable of. I highly encourage folks to revisit old tasks that they were previously unsuccessful getting AI to assist with–try again!

Let’s also be clear, the internet has always mostly been full of cat pictures and lies. Accordingly, much of the information Claude has gained access to is worthless or worse. But a paradigm shift has now occurred wherein AI will be our first line of defense against misinformation and bias–first Claude can vet the sources and then, only when it’s satisfied or fooled, we can.