Five things that really are different with AI
Chatbots that pass the Turing test
The original Turing test is simple. One test subject talks to two ‘entities’, one human and one machine. A machine that fools half or more of test subjects into believing it is the human (instead of the real human) is said to have passed the test.
LLMs effortlessly pass this test in all kinds of settings. They are as good as humans at just making conversation. We’ve gotten so used to it already. ChatGPT is already two years old. This capability has enabled:
Really great Customer Support bots, tutors, meal planners, recipe suggestors, email writers, etc etc. This is still a large part (the biggest part?) of what people use ChatGPT for.
Real-time, near perfect translation. I thought about putting this down as a separate category, but decided to demote it to a sub-bullet. Debatable, I admit, because LLMs were basically created by Google to improve Google Translate. But at the end of the day, this is just chat in every language.
The weirdness around AI friends/girlfriends/boyfriends. Personally I find it very difficult to empathise with the stories of people who fall in love with AI. But, evidently, it’s happening enough to drive 15k visits per week Reddit communities.
A revolution in speech-to-text transcription and text-to-speech generation
Everyone I know has started talking to their devices. That used to be a super fringe thing to do. Alexa, Google Assistant, Siri, they all sucked so bad at even understanding what you’re saying. Now, they understand. Whether they can do something about it, that’s a different question. But at least they know what you’re saying.
Nearly every meeting I’m in is now transcribed automatically and I regularly use these transcripts to extract action points, notes, or just to remember what was said exactly.
Text to speech used to be a very hard technology to build. The father of my high school girlfriend built it in the 00s, by having real humans record thousands of vowels that they could then use to construct convincing human-sounding speech in narrow domains. Now, the products from companies like ElevenLabs sound absolutely human, no matter what text you input.
Semantic analysis of information
AI now has a level of understanding of text, that in some cases almost feels like a rudimentary understanding of the world. They are pretty good at a whole bunch of surprising fact-finding jobs that previous search solutions were not.
Summarizing and ‘un-summarizing’.
Diagnosing obvious things. I’ve read anecdotes of people going to the hospital at the insistence of ChatGPT, and finding out they were having a heart attack
Web search. I’m always mind blown when I use Perplexity for something like: “I remember reading a post, I think it was on Substack by a guy who runs a marketing agency. He was being very critical about LLMs, and ….. etc” It actually finds the damn post. When this used to happen in the past, it was probably forever gone.
Understanding {code base, contract, report} and explaining it. Who still reads a contract or long report before first exploring it with an AI Chatbot? I sure don’t. Amazing!
How do you connect a google sheet with an API? Good luck figuring that one out with just Google as a non-coder. Now, ChatGPT will just generate the code to paste into the AppScript runner.
Cloud Cognition
This sounds catchy. I use cognition instead of intelligence on purpose. Intelligence has a lot of implications. But cognition I think is a bit more clean. When an AI is autonomously deciding whether to use a search tool or not, or whether it has enough information to generate an image or still needs to ask the user for more input, it is ‘thinking’. Many modern AI applications have forms of memory, where they can store information for later so they don’t forget everything all the time. The arsenal of tools that, say Claude or ChatGPT can use now is growing super fast. They can write code, create excel files and powerpoint decks. The Chatbots are given a list of tools, and they decide by themselves which one to use when, given the guidelines and the user input. When I first built this capability into Magicdoor.ai I was genuinely mind-blown.
People are justifiably excited about ‘Agentic AI’. When we gave our AI Chatbot tools to access our backend for live data about customers, that reduced human support tickets by more than 40% within a few weeks (some teething issues). It’s true that this will result in job losses in the support agent profession, but also think of the customers on the other side of this who get their issues solved in a few minutes instead of hours.
Video and image generation indistinguishable from real
In 2025, image generation crossed over from being hit and miss to being actually useful for marketing people. Video is now there in terms of quality, but not in terms of access yet. It won’t be long. There isn’t enough of it out there yet to know really how this is going to play out, but YouTube, Instagram and TikTok have already changed dramatically. Recently, there was the kangaroo who got a jumpscare from a halloween skeleton on a porch. Reddit sentiment is 95% sure AI. I genuinely could not tell. I’m not surprised it’s AI but also would not have been very surprised if it was real. Here’s another example, of a kid riding a dog (hear me out, and take a look, okay). When the cameraperson first looks forward, the road marking are yellow, and there are no power lines perpendicular to the road. When he looks backwards the road markings are white. Then when he looks forward again, there are power lines suddenly.
I keep saying this, but: detecting AI generated content is hopeless. Human content will be authenticated instead. There might be a market for products guarantee a piece of content was human generated.
There really is new capability
These five things are genuine new capabilities that computers didn’t have until a few years ago. What we will do with these capabilities is up to us, and it is still early.
Just for fun, here is a presentation version that Claude made from this post in 3 minutes without any other input, and no changes from my side. It’s not perfect and I think it AI'ified the content a bit, taking the edge off where I might not want that to happen. I mean, voice assistants really did SUCK!
This was a good one! If Claude could talk to Napkin.ai then the presentation may get even better.