- An AI engineer says most people don't understand what AI can really do and what it can't, yet.
- Those working in AI understand when startup claims are realistic and when they're mostly hype.
This essay is based on a conversation with an AI engineer who currently works for an AI legal startup and asked not to be identified because he is not authorized to speak about his work experiences. Insider has verified his employment.
Generative AI is just way, way overhyped right now and that means that many of the AI startups getting VC funding today are going to fail. This is going to be similar to what happened in crypto. There are going to be selective applications from startups that work well and can build businesses, but maybe 70-80% of them are going to die in the end.
Even OpenAI has been experiencing declining usage lately, which may indicate that generative AI chatbots won't take over the world.
I've been working on AI systems for almost a decade and can tell you that we've seen all of this before. This is what happened to the self-driving car industry.
As good as ChatGPT is at chatting, or Dall-E is at creating art, what these programs are doing is mimicking information they have ingested from the past.
Today's AI can't really do what so many startups say their apps can do because AI can't reliably predict things.
How we got here: three waves of machine learning
There's been, so far, three waves of machine learning in AI development and each of them created a lot of startups. The three waves are: supervised learning, unsupervised learning and reinforced learning.
Supervised is when you are teaching the AI model how to do something like identify a pen. You hire a bunch of people to manually label photos of pens (and startups were born to do this) and you are training the model to answer the question, is this a pen or not?
Unsupervised is when you are writing a rule in an algorithm and telling AI to detect what the object is. An example is pixel detection to identify colors: red, green, yellow, etc.
Reinforced learning is when you are training it to identify, for instance, an apple, and then reinforcing if it got it right or not. Is this an apple? Yes. Okay, great. Is this an apple? No, you got it wrong.
Then AI engineering got into this steep learning curve based on all of it. It was reinforced and unsupervised combined together.
What I'm trying to get people outside the AI community to understand is that this is really essentially just a probability game. Like what is the highest probability that something will occur in the future? For instance, a self-driving car uses a bunch of deep learning models. When the model notices "oh, here's a human to the right of us. One second before I get to a nearby position, I need to predict if this person is going to move across the street or if this person is going to stay still."
And so it calculates all that based on postures: like if they're on their phone just standing still, most likely that person is not going to move. The probability of this person moving across the street is like maybe, 0.001%.
But accidents end up occurring.
So that's what deep learning is: it's like a probabilistic prediction centric thing. And this is the day that we live in today. People say, "Hey, we're creating something new." But all that is, even for something like OpenAI, is that they are feeding it so much data and then they're basically saying, replicate it and create something based on the previous information.
What you can and can't rely on AI to do
Yes, if you use that kind of AI correctly, it's super powerful. But AI wholly depends on the information fed to it. The information could be biased, or spitting out something that isn't creative but basically plagiarized, or based on old and outdated information.
So, how can you tell if an AI startup's technology will work or if it and the company will likely fail? If what it's doing is reusing static information and doesn't rely on having to predict an outcome, it's on safer ground. Like routing a map in a controlled environment for warehouse robots. Unlike self-driving cars, warehouse robots work in a controlled environment, right?
Or call center triage: so anything that comes into the call center for an identified reason, machine learning can analyze that and route it to the correct person.
But startups that require heavy prediction will struggle to achieve their claims, just like most of the self-driving car startups from the last decade have not become big businesses (and we aren't all being driven around in them).
In this bucket are startups with tech that relies on people changing their behavior to trust a machine instead of another human being like, for instance, virtual human apps: AI assistants that are supposed to replace a human to manage an executive's office, or AI bots that are supposed to replace salespeople. Another category is anything that requires strategy, like in the legal world, developing a defense. And another category is anything that requires the AI to understand and predict what people are feeling, like a concierge service.
So how can we bridge the gap between today's limited AI and a day when we can totally rely on it? Well, there is an intermediate solution. We can use AI today to do consistent, repetitive work – let's call that "prework" – and include human beings in the process.
But there's a catch even for this because you also need to account for the true cost of AI adoption. Take, for example, Amazon Go. As of June, Amazon closed its ninth Go store, including flagship stores in San Francisco, Seattle, New York. Who were they up against? They were trying to provide a better user experience by replacing a $20/hour laborer (about what Whole Foods cashiers make) with smart tech.
But that involved paying for expensive engineering talent, building proprietary technology, and then paying for the ongoing support costs for complex, vision-driven computer networks – not to mention the cost of maintaining and retraining their models, which is seriously expensive. Self-checkout ended up being a much more cost-effective way for other retailers like Costco, Walmart, grocery stores and still solved the same issue of improving the efficiency of checkout. No AI required.
So, if you are removing repetitive, predictable administrative tasks, that's a good use for generative AI. But if you are trying to create something that requires predicting something that's going to happen in the future? How does that technology work? That's the shiny new thing. And I would not work for, or invest in, that company today.