An AI engineer explains why most AI startups die.

• AI engineers say most people still don’t understand what AI can and can’t really do.
• Those working in AI understand when startup claims are realistic and when they are hype.
• He shares how he evaluates which startups to use, invest in, or work with, and which to avoid.

This essay is based on a conversation with an AI engineer currently working for an AI legal startup, but has requested anonymity as he is not authorized to discuss his work history. Insiders confirmed his employment.

Generative AI is currently hugely overrated, which means that many of the current VC-funded AI startups will fail. This will be similar to what happened with cryptocurrencies. There will be select applications from startups that work well and can build businesses, but probably 70-80% of them will eventually disappear.

Even with OpenAI The amount used has decreased recently.which may indicate that generative AI chatbots will not conquer the world.

I’ve been working with AI systems for nearly a decade, and I can say that I’ve been through all of this before. This is what happened to the self-driving car industry.

As much as ChatGPT is good at chatting or Dall-E is creating art, what these programs do is mimic information they ingest from the past.

Today’s AI can’t really do what many startups say their apps can do because AI can’t predict things reliably.

How we got here: three waves of machine learning

There have been three waves of machine learning in AI development so far, and each wave has spawned a number of startups. The three waves are supervised learning, unsupervised learning, and reinforcement learning.

Supervised is when you teach an AI model how to do things like identify pens. We’ve hired a bunch of people to manually label photos of pens (and startups were born to do this), and we’re training a model to answer the question, “Is this a pen?” .

Unsupervised is when you write the rules in your algorithm and tell the AI ​​to detect what the object is. An example is pixel detection to identify colors such as red, green, and yellow.

Reinforcement learning is about training, for example, to identify apples and reinforce whether it was correct. is this an apple? yes. Ok, great. is this an apple? No, you are wrong.

AI engineering then entered this steep learning curve based on all of that. It was hardened and bonded unsupervised.

What I want people outside the AI ​​community to understand is that this is essentially just a probability game. For example, what is the highest probability that something will happen in the future? For example, self-driving cars use many deep learning models. When the model says, “Oh, there’s a human to our right. He needs to predict a second before he reaches a nearby position, whether this person will move across the street, or if this person will stand still.” I have.””

And it’s all calculated based on posture. For example, if someone is just standing around using their cell phone, they are likely not moving. The odds of this person moving across the street are probably 0.001% for him.

But in the end, an accident will occur.

This is deep learning. So it’s like something that revolves around probabilistic prediction. And this is the day we live in. People say, “Hey, I’m building something new.” But even with something like OpenAI, they’re just saying that they’re feeding it tons of data and basically cloning it and creating something based on previous information.

What AI can and cannot rely on

Yes, this kind of AI is very powerful if used correctly. But AI is completely dependent on the information it is given. Information may be biased, uncreative but basically plagiarized, or based on old and outdated information.

So how do you know if an AI startup’s technology will work, or if the technology and company are likely to fail? It is in a safer position if it does not rely on predicting outcomes. It’s like routing a map in a warehouse robot controlled environment. Unlike self-driving cars, warehouse robots operate in a controlled environment.

Or call center triage: everything that comes into the call center for a specific reason is analyzed by machine learning and routed to the right person.

But just like most of the self-driving car startups of the last decade haven’t become big companies (and we’re not all swayed by self-driving cars), it does require careful forecasting. Startups will struggle to live up to that claim.

This bucket includes, for example, start-ups working with technologies that rely on humans changing their behavior to trust machines rather than another human, like virtual human apps. That is, AI assistants and AI bots that are supposed to manage the boardroom instead of humans. It should replace the salesperson. Another category is anything that requires a strategy, such as building defenses in the legal community. And another category is anything that AI needs to understand and predict people’s emotions, such as concierge services.

So how do we bridge the gap between today’s limited AI and the day when we can rely entirely on AI? Well, there is a middle ground. Today, we can use AI to do consistent, repetitive tasks. Let’s call this “pre-work”. – And include humans in the process.

However, even this has its pitfalls, as the real costs of AI deployment must also be considered. For example, consider Amazon Go. As of June, Amazon closed its ninth Go store. Including flagship stores in San Francisco, Seattle and New York. Who were they up against? They were looking to provide a better user experience by replacing their $20 an hour worker (about the income of a Whole Foods cashier) with smart technology.

But that includes paying for expensive engineering talent, building your own technology, and paying the ongoing support costs of a complex, vision-driven computer network. Not to mention that the cost of maintaining and retraining the model is also very high. Self-checkout has become a much more cost-effective method for other retailers such as Costco, Walmart, and grocery stores, yet it still solves the same problem of improving checkout efficiency. No AI needed.

Generative AI can therefore be put to good use when it comes to removing repetitive and predictable administrative tasks. But what if you’re trying to create something that needs to predict what will happen in the future? How does that technology work? And I wouldn’t work for or invest in that company now.