Good morning. We’re taking a deep dive today into Waymo’s newly-released safety statistics, which sound good but don’t quite tell the complete story (which, long story short, is that we don’t have nearly enough data).

— Ian Krietzberg, Editor-in-Chief, The Deep View

In today’s newsletter:

Warp: Payroll and Compliance for Startup Founders

Steve Jobs is quoted as saying, “Design is not just what it looks like and feels like. Design is how it works.” There are dozens — hundreds, maybe — of payroll providers to choose from. So why would a startup attempt to build another?

Simple. Every other payroll solution out there aims to compete with features, instead of actively simplifying workload for the customer. Traditional payroll software has users paying for 50 features they won’t use and fails to do the basics (such as paying people and staying compliant with different state tax agencies) well.

Warp has said no to all the superfluous features that other software providers are focused on, and instead opts for a dead-simple interface that allows startups to get in, do the thing, and get out.

Startups should spend more time in Notion and less time on the phone with the Pennsylvania Department of Revenue.

Get started now with Warp and get a $1,000 Amazon gift card when you run payroll.

AI for Good: NASA’s high-tech satellite upgrades

Even as technology has evolved dramatically over the past few decades here on Earth, up in space, there’s a lot of tech that has become outdated, according to NASA. 

What happened: In light of this, Irvine Sensors Corporation is working with NASA to develop computer tiles that will enable space-based remote sensors to process data far more quickly. 

“SMARTIE would have endless applications,” James Yamaguchi, Vice President of 3D Electronics and Mass Storage at Irvine said in a statement. “It could provide autonomy to single satellites or satellite constellations using AI, enable distributed sensors where parts of the instrument are set in different spacecraft, and perform complex operations usually done on the ground to reduce data throughput.”

AI drug discovery company is trying to challenge DeepMind

Earlier this week, we talked about Google DeepMind’s AlphaFold and AlphaProteo biological AI models. On Tuesday, a new startup emerged — Chai Discovery — that says it does the same thing, just a little bit better.

The details: The model, a multi-modal foundation model for molecular structure prediction, is called Chai-1. The company released the model for free for both commercial and non-commercial use. 

Chai, which was founded only six months ago, has raised $30 million in funding from Thrive Capital and OpenAI, according to Bloomberg. Jack Dent, one of the company’s co-founders, said Chai is making its first model free and hasn’t yet discussed plans to commercialize its tech. 

The caveat here is that Chai-1 and similar models can speed up the drug discovery process by offering up candidates that then need to go through years of rigorous, human-led testing. This, according to one health VC, is why the drug discovery side of the AI/health intersection is a much more challenging investment than it seems.

Tech Workshop: Build a Voice AI Agent with Deepgram & Groq

Ready to write the code to build your own voice AI agent? Then this workshop is for you!

Our experts will guide you through each step and complex challenges in AI agent development like interruption handling, end-of-speech prediction, and low-latency function calling.

Plus, receive:

Sign up before it's too late! Get 20% off with code: DV20

📆 September 20 @ 9AM-12PM PT

If you want to get in front of an audience of 200,000+ developers, business leaders and tech enthusiasts, get in touch with us here.

Part 1: Waymo unveils new safety data

If you’re as chronically online as I am, you might have seen that Waymo, the robotaxi startup that actually seems to be pulling it off, recently published a new safety hub, tracking all of its safety statistics against human benchmarks. The data implies that, across a few different metrics, self-driving cars (at least, Waymo’s self-driving cars) are safer than human drivers, a point that Twitter took and ran with.

But the reality here is just a little more complicated than that. Part of this has to do with the scope and scale of the data at hand, which we don’t have enough of. Part of it has to do with the realities of the underlying AI architecture. 

Let’s start with the data: Through June of 2024, Waymo’s total fleet has clocked 22 million vehicle miles without a human driver. This breaks down to 15.4 million in Phoenix, AZ and 5.93 million in San Francisco, CA. 

According to journalist Tim Lee, Waymo has reported 200 total crashes. Of the 23 most severe accidents, 16 involved another car rear-ending the Waymo.

Part 2: The complicated statistics behind ‘safe’ self-driving cars

Here’s where it gets complicated. If I stopped there, things sound really positive. The problem is the available data doesn’t really make for an even comparison, and this comes across in a few different points. 

In San Francisco, Waymo operates across a 55-square mile region (roughly half of the size of the Bay Area). Waymo operates across 315 square miles in Phoenix, an area that, in total, covers close to 500 square miles. 

As Waymo itself notes in its safety hub, “all streets within a city are not equally challenging. A limitation of these presented benchmarks is that spatial and temporal differences between the Waymo and human benchmarks within these deployed cities have not been accounted for in the rates.” Waymo did note that its vehicles tend to operate in the denser parts of its deployment areas. 

An element of this disparity also involves quantity. Waymo told me that it has a total fleet of 700 vehicles, with around 300 in San Francisco and 200 in Phoenix. The company said that not all vehicles are on the road at the same time.

By comparison, In Arizona alone, there are more than eight million registered vehicles. Arizona drivers in the Phoenix Metro Area — with a population of 5 million — drive about 10 billion miles each year, which equates to roughly 27 million miles each day. 

In Phoenix, as in everywhere else, there are many, many more human-driven cars, covering far more miles, all the time. If everything scaled appropriately (similar mileage, number of vehicles and geographical area), it is not at all clear that Waymo’s safety stats would scale in kind. 

As Missy Cummings, a former safety advisor to NHTSA and a mechanical engineering professor at George Mason University, has said: we do not have nearly enough data to “make a statistically strong claim. We’re not even close.” (In the above talk, Cummings explains the major limitations of the current neural networks on display here).

I believe in self-driving cars. As Waymo points out, lacking a human driver removes human fallability and chronic distraction. Humans are not great drivers, and in my own experience, they seem to be getting worse. (Even with that, NHTSA has reported a fatality rate of 1.26 per 100 million miles in the U.S. in 2023). 

But at the end of the day, what’s happening here is pattern recognition, not reasoning. That’s what’s dangerous. And it’s dangerous for the same reason it’s dangerous for cops and schools to integrate generative AI; these systems are not reliable.

Months ago, we reported on a study that suggested autonomous vehicles are likely safer than human drivers. But it noted the caveat that they are more likely than humans — by significant rates — to get in crashes at dawn, dusk and during turns. 

Waymo will continue to scale up its operations. I hope that its safety numbers scale in kind, but I find this unlikely, especially if and when they push east into states that deal with worse weather. The problem is that the rest of us are, willing or not, participants in this experiment.

I think we’ll get there. But it’s not as close as it seems. And we need to be skeptical and cautious in order to make it happen safely.

🤔 Your thought process:

Selected Image 2 (Left):

💭 A poll before you go

Thanks for reading today’s edition of The Deep View!

We’ll see you in the next one.

Here’s your view on AI in finance:

A third of you don’t care if your finance professional uses genAI; 22% just don’t want to know about it and 18% use AI as your financial professional.

15% aren’t on board.

Something else: