Brian O’Malley

Whereas something that's more quantitative, like filing someone's taxes on their behalf, if there's an error deep in there that's ultimately going to get me audited, that's going to cause a bigger problem.

So in that case, 95% accurate really is not 100%.

And so I think there's some truth to the fact that it will take a little while to get there.

When I think about Waymo in San Francisco, it took them longer to deliver the product than expected.

But once it was out there, it's both satisfied customer needs, the extent that they can charge a premium, and also the safety incidents.

are way down.

Now, what I don't think people are fully looking at with AI yet, which is also true with Waymo, is just what are the intrinsic costs of training a new city?

And from a pure economic standpoint, what's the breakeven time period to get back on those initial investments?

And how do the economics ultimately work?

Because from what I understand, the cost to train San Francisco for Waymo was very expensive to the extent where you couldn't just automatically do that across the rest of the country, unless capital was seemingly free for a very long, long period of time.

I think we're in that same place with AI where that last 5% of training, there will ultimately be a question once gross margins matter, once these companies need to show a path to profitability, whether that cost actually lines up with the benefit.

And that's where right now, I think we're seeing that having a human in the loop for that part, not only does it complete the needs of the end customer, but it also might be a more cost effective solution than trying to get 100% of the way there anytime soon.

To use the autonomous example, once every couple of months, media would get up in arms, you know, Tesla car gets in an accident.

Of course, it didn't report on the 10 times, 20 times more accidents that happened in the non-Tesla vehicles.

So it's kind of this like impossible straw man where you're comparing something that causes accidents every 10,000 rides with something that causes accidents every 5,000 rides.

And there's no relative kind of comparison there.

And I think you brought up a good example, like email and how many errors do human beings do on emails?

when I put emails into AI that AI, at least today, does a much better job than I do in terms of drafting and responding to emails.

So there's not only the error rate, but it's also the relative error rate that I think gets lost on people, especially in consumer products.

At the end of the day, these products need to be adopted by people in society.