How to Compete with AI — and Win: Dharmesh Shah (Transcript)

On the list, they say it’s like tennis for people that think running is overrated. Perfect for me, I think that would be awesome.

Falling in Love with AI

As you can probably tell, I am in love with AI. Not in a frightening, romantic way like the movie Her, and not even in a friendly, platonic way like the movie Big Hero 6, but more in a geeky, systematic, algorithmic kind of way. I love AI for what it helps me do, and what I do is create software, and AI has raised the level of abstractions for software developers everywhere. So now I can solve a much broader class of problems than I’ve ever been able to solve before.

This is a story of how I fell in love with AI. It goes way back to 2 BC, and by BC, I mean before ChatGPT. You know, the dark ages when we had to Google things ourselves. So picture me in my natural habitat. I’m working at home in my pajamas, clacketing away on a mechanical keyboard that sounds like it was designed in the 1900s, largely because it was designed in the 1900s, and I’m doing the thing that I do, clicky-clacketing away, and I get a message from Brian Halligan, who’s my friend and co-founder of HubSpot, and Brian asked “Dharmesh, want to join me on a call?”

And my immediate reaction is, wait, what? Because as an introvert, I hate phone calls with the heat of a thousand suns, and my friends and my family know that the absolute best time to call me is email. The second best time to call me, also email.

Anyway, so back to the call. In this case, I said yes, and the reason I said yes was the call was with Sam Altman, the founder and CEO of OpenAI, and they had just gotten started, and they were working on this new AI technology based on Google research called GPT. They didn’t have an application, there was no user interface, all they had was an application programming interface built for developers like me to embed AI into their own applications.

Shortly after the call, late that night, I got early access to the GPT API, and I coded a simple application just to test out the API, and I had my first conversation, first real conversation with AI. It was love at first chat. The fact I could have a natural language conversation with software was just game-changing, and I had attempted to do this for 20 years in prior attempts, and this was the first time it actually felt real. And this was not even the ChatGPT, this was two years before that.

I was wholeheartedly humbled, and I became a boy standing in front of a journey of pre-trained transformer hoping you would love him. So over the ensuing weeks and months, I spent many quiet contemplative nights pondering what this all meant. What had we just unlocked? AI is possibly the most powerful tool that humanity has ever created.

Understanding AI: How It Works

So now you’re wondering, all right, that’s all very highfalutin, but what does that mean to me? And that’s why we’re here. So I’d like to try and answer three questions. One is, how does AI actually work? Because by understanding some of the basics of it, I think you can extract more from it, and it will also reduce your anxiety if you’re feeling some anxiety. What should you do with it? What have I learned about working with AI within the company and helping advise a bunch of startups and other people? And then finally, where do I think things are going from here? It’s all about transforming you into the you that’s powered by AI.

Let’s talk about how AI works. We’re going to decompose the GPT that you probably typed in a thousand times. It stands for generative pre-trained transformer. So just decompose it. The G stands for generative, and this is the technology that’s used to create the large language models that power things like GPT. And the reason generative is so important is because that’s what allows the AI to actually create things, to generate text, to write an article, to write a limerick. This is the first time we’ve seen software reasonably be able to do that. And that’s why this age of AI, because AI and machine learning have been around for decades, is called the age of generative AI, because for the first time, we can actually generate things.

Now, some experts in the field who are brilliant have sort of dismissed large language models as being glorified autocomplete, which I will concede is the truth, but it’s not the whole truth. There’s much more to it than that.

So let’s dig in. So let’s say we pass the large language model this sentence fragment, “the horse jumped over the blank.” What the large language model is programmed to do, designed to do, is predict the next token. The token is just an aspiring word. It’s just a term geeks use. We can think, predict the next word. And what it does is based on this neural network, based on all this AI software, it comes up with a probability list of here are the likely words that come next based on what you just told me. So here it’s saying about 45% chance that the word “fence” would fit in here. There’s no perfect answer, but this one is the most likely it thinks to be correct.

Now the cool thing is, so we can do that, but we can keep calling the LLM recursively so we can just take the sentence that passed us and pass it back in and it would generate the next word and the next word and the next word for as long as we want to go. And that’s why the large language model, which still does one word at a time, can write a 1200 word article on whatever topic you choose. It’s just looping through that construct, predicting the next word, predicting the next word, predicting next word.

Now let’s tackle the pre-trained transformer part. I’m going to focus on the pre-trained because transformers make my head hurt. Here’s how it works. So there’s this process called pre-training, and it takes months and months, but I could not miss the opportunity for a good movie reference. It’s the little things in life.

So imagine this made up fictional machine and the machine has a bunch of knobs on it, like a bunch of knobs that you can turn up or down, billions and billions and billions of these knobs, like a stereo from the 1980s. And the purpose of the machine is that it will take some text in and it will attempt to predict the next word as its output. That’s what it sends out. That’s what the machine does.

In order for it to be able to do that, it goes through this pre-training process. And here’s how that works. The AI companies collect all publicly available knowledge on the Internet, every web page that ever was published, all of Wikipedia, all of Reddit, everything that it can find. And then it will

Take chunks of that, pass that text into the large language model, into the machine, and then say, did I get it right or not? So if we pass in “the horse jumped over the blank” and it doesn’t guess correctly, it’ll go back and the system will say, I’m going to tweak those billions and billions of knobs until I get the right answer. So it’s optimizing for accuracy.

Now imagine it going through that process of putting an input in, seeing if it’s right or wrong, tweaking knobs and dials, happens billions and billions of times over the course of many, many months, spending millions and millions of dollars. That’s what results in the large language models that we use in products like ChatGPT.

In theory, a large language model is an autocomplete, right? That’s what it’s doing. But it’s like an autocomplete with a PhD in everything! It has built this world model, trying to understand language and all the things that it has in order for it to solve for that accuracy, be able to predict the next word. It needs to know a fair amount in order to be able to do that. That’s why it feels much more magical than a simple autocomplete.

The Limitations of LLMs

Okay, so LLMs are awesome, love them, but there are some problems. So LLMs are limited to whatever data they had at the time they were going through pre-training. So it’s like hiring an intern that has a PhD in everything, very, very well read. They show up on the first day on the job and they know nothing about your business. They don’t know how any of your internal IT systems work. They don’t know where the coffee room is. They know nothing because that was not part of their pre-training. That’s not what they had studied. That’s problem number one.

Problem number two is that the data that’s embedded inside the LLM is frozen in time. So in that hypothetical intern example, even if they had read your company’s website, maybe they read it six months ago and things have changed, it’s not going to know about any changes after that pre-training date.

Problem number three is large language models sometimes hallucinate. And that’s a fancy highfalutin word for they make things up. Now they’re not trying to mislead you. They’re like that friend that is so confident that they don’t know they’re wrong. We all have one of those friends. Some of us are one of those friends.

What You Should Do With AI

So I’m going to start with what you shouldn’t do about AI, which is dismiss it, restrict it, ban it in your household, in your team, in your company. So there was a time when people said, “Don’t use spellcheck, use a dictionary and you’ll learn how to spell better.” Then they said, “Well, don’t use the internet. The internet has mistakes.” Shocking. And now we will have people that say, “Don’t use AI” because life is supposed to be hard, I guess.

Speaking of a hard life. So I’m originally from India. I was born in a tiny town in the state of Gujarat in India. Shout out to my fellow gujus out there if they’re writing. I know there’s like hundreds of millions of you. But in my house growing up, we had no phone, no TV, no refrigerator. And so our milk was delivered daily. So it was like cow to table, right? Before it was fashionable to do those kinds of things.

Now I’m going to contrast that to my son, Sohan. This is him when he was two years old, beta testing the latest iOS. And he learned to swipe before he learned to wipe. He’s been exposed to technology. I exposed him to that early GPT when he was in his early nines before ChatGPT came out. He’s 14 now. And he uses AI in much more creative ways than I do.

Generation AI

So I’ll give you an example. So Sohan is a fantasy fiction manga fan and an aspiring author someday. What he uses ChatGPT for is to do world building, which you do in these very complicated plots. So you set up your world that says, here’s the rules. Here’s the power structure. Here are the characters. Here’s what you can and can’t do. Here are the magical powers. And he has a 2000 word prompt that describes his world. And the way he tests is that he makes it into a role playing game. So he tries to do things to see if the system will allow people to do it, and if it can track the characters and tracks what’s going on. It is brilliant. It’s brilliant.

Basically, what he’s doing is he’s programming in what will be the most popular programming language in the world someday, English, right? He’s just giving the computer instructions as to what he wants.

Now Sohan’s not just using generative AI. He is a member of generation AI, an entire generation of people that will be growing up with AI. They don’t know a world without it, just like many of us don’t know a world before the Internet. Now imagine telling one of these folks that you don’t use AI. They’re going to be like, “Is this like cave age days? Are you like carving things into stone? Like no AI? Seriously, lol.”

So my advice is to be careful, but to stay curious. So at HubSpot, we have 8,000 plus employees, and we not only encourage them to use AI, we expect it. We invest in it. We help educate. When we’re doing hiring, it’s a signal that we look for. You don’t have to be an AI expert, but you at least have to be curious and have an eagerness to learn it.

So one thing I’ve learned is that getting AI adoption isn’t a walk in the park. It’s a hike up a hill. There’s going to be resistance. And the way to make progress is to dream big. You have these big lofty ideas of what you can do with AI. But my advice is to iterate small. Try small experiments, try little things, and move forward.

The Latest AI Advancements

Let’s close with what’s going on right now. The first big advancement is now we can give AI access to tools. So it can use a web browser. It can use your internal IT systems if you give it access. And it’s going quickly from being this kind of system of interaction, where you just type back and forth, to a system of action. It can actually do things on your behalf.

Now, GPT has been great for creating things for humans, text, audio, video. But where we’ve seen the steepest advance, the fastest progress, is AI generating things for computers, namely generating code. And this is a massive development. Because now if you have a problem, a business problem, a science problem, an academic problem, if you can kind of state your problem in terms of code, like, “I wish software existed to do X.” If you can just describe that, there is a good chance that AI is going to be able to generate that product in a thing called vibe coding your way through.

AI Agents: The Next Frontier

So those are all exciting developments. What’s the most exciting is all these things coming together in this new thing called AI agents. We’re not going to spend a lot of time on agents. The simple definition is AI agents are software that does things that require multiple steps. It’s just AI software.

Now here’s a concrete example. I had to squeeze one visual dad joke in there. It’s in my speaker’s contract. Here’s an AI agent example. We could ask for an AI that says, go look for new videos from my CEO and founders. Generate the transcript. See if it’s relevant. Summarize and send me an email. So the agent can go through multiple steps, work on your behalf behind the scenes.

So the thesis is that tomorrow’s teams are going to be hybrid. It’s going to consist of both humans and AI agents working together to accomplish whatever the goal of the organization is. Now this is most often quoted thing when people say, “Oh, is AI going to take my job?” And the answer is, AI won’t take your job. Someone using AI will.

I think about it a little bit differently. I would assume that AI will take your job and give you one that’s much better with much less rote work, much more help from digital systems so you can focus on the things that matter most. And the thing to remember as humans, we are more than the sum of tasks that we do. I get AI help on all sorts of things. Am I less valuable as a human than I was before?

I don’t think so. I think I have more impact. I think I’m more valuable. I think that applies to all of us. Back to our original question. Is this an existential threat or an exponential opportunity? And I think this is not a case of you versus AI. This is a case of you to the power of AI. It’s an exponential amplifier.

Closing Thoughts

As you walk away, my advice to you would be, ask yourself, any time you do something, you’re sitting at the computer about to start something, ask yourself, “How might I use AI to help me with this?” Put aside the skepticism. Just type something into ChatGPT or your favorite tool of choice and see what happens. Just start learning by doing it and give it a shot and you’ll be surprised how often it works.

And the thing to remember is that, and this is a really beautiful irony, is that the better our artificial tools get, the better AI gets, the more human it allows all of us to be. We’re spending more time on the things that make us human because human is not a bug. It is the ultimate feature that we all have.

Thank you. Thank you.

Related Posts

• Mo Gawdat: How to Stay Human in the Age of AI @ Dragonfly Summit (Transcript)
• Transcript: The ONLY Trait For Success In The AI Era – Aravind Srinivas on Silicon Valley Girl Podcast
• Transcript: Mark Zuckerberg on AI Glasses, Superintelligence, Neural Control, and More
• Lessons from Apple ID Hacks: What Transcription & Media Sites Should Do to Secure Their Users’ Accounts
• Transcript: Demis Hassabis on AI, Creativity, and a Golden Age of Science – All-In Summit