🏢 🚙 🤖 Issue #38: autonomous progress? media manipulation and old folks

You can follow me on Twitter here

🤖🚗 Autonomous disengagements - Last week US based AV companies raised approximately $3bn from investors - companies such as TuSimple ($95m), Nuro ($950m) and Aurora ($530m). like all technologies, autonomous vehicles, are working their way through the Gartner Hype Cycle - currently L4 automation is bang in the middle of the ‘peak of inflated expectations’ and the ‘trough of disillusionment’,

The aforementioned fundings happened to have coincided with California’s DMV releasing numbers on companies testing autonomous vehicles, and how often human backup drivers had to intervene - an imperfect proxy for industry progress. It’s important to note that the data was context-incomplete, unstructured and lacks a consistent format, thus there many vague reported reasons for the disengagements. First some numbers around registered vehicles and miles of testing:

Waymo has a clear advantage purely in terms of vehicle miles logged in California. Though this doesn’t take into account Tesla, which as we’ve previously covered has the enviable advantge of collecting L2/3 data over-the-air from consumers. Critically, a lot of this data can be from long-tail events i.e highly unpredictable real-world road scenarios.

This shows a significant outperformance for Waymo and a woeful one for Apple. Though, there is a more interesting story here which speaks to the lack of standardization of this data and the gameability of these numbers (tech companies gaming numbers? surely not!). Apple data shows they reported two types of incident, raw disengagements and important disengagements, over differing periods of 2018. When re-calibrating the data to ‘important disengagements’, they show a disengagement every 2000 miles (28 disengagements in 56,135 miles).

The numbers here go to show that AVs have a long way to go in order to live up to their promise of dramatically safer, more efficient road journeys. Currently, in the US, humans have one accident per 165,000 miles driven, a 16x improvement on Waymo.

This data also nods to the fact that California has enjoyed forward thinking regulation when it comes to vehicle testing. What it doesn’t speak to is comparative vehicle testing in China, which has been markedly aggressive in chasing down frontier technologies - doing so with far fewer ethical boundaries. Whilst companies in California are making leaps towards L3/4 - I suspesct China will be aggressive in moving past current US benchmarks of success.

Deeptech - aka nerd porn

It seemed a majority respondents to my reader survey wanted to see a bit more deeptech. All of the below pieces are focused on the manipulation of media using deep learning, so here goes:

The case for open-sourcing OpenAI’s GPT (generative pre-training transformer) language model. Most of you have likely already seen OpenAI’s new model which can teach itself to accurately generate text (original release post here -worth a read if you haven’t). OpenAI have deliberately not open sourced the model due to concerns around dangerous misuse, such as impersonation of individuals and organisations. As humans we have an obsession with destructive forces and death - to a point where we regularly overstate new threats. My conclusion on open-sourcing this:

Isolating audio using CNNs - this convolutinal neural network seperates specific audio sources, such as vocals on a Beatles song or da bass on a Bieber track (god forbid!). This has wide implications for music production, remastering or voice call quality once the model is commercially deployed.

This model is also not dissamilar to Pixel Player who used a ResNet in order to do sound source seperation on images - the below example shows the model differentiating the sound of a violinist versus a guitar on a video:

Text-to-image generation - this paper proposes an AttnGAN, an attention based generative adversarial network which can, edit images based on text input. For example:

A number of companies have been products around media manipulation such as Let’s Enhance, which looks upscale images using a GAN, or Runway, an application platform which open sources the use of such models to creatives.

___________________

Elon on Tesla’s autonomy - this week Elon Musk was on Ark Invest’s podcast talking about Tesla’s autonomous aspirations. Without a hint or irony he claimed that Tesla would be at Level 4/5 autonomy by year end

Why retirement communities might be perfect for AVs - Voyage’s CEO Oliver Cameron, discusses their move to begin operating autonomous vehicles inside the worlds largest retirement village. This gives Voyage a closed beta test of sorts, with consistent testing, low speed and a growing demographich who’ve been the most adverse to the idea of autonomous

which brings me to….

Elder transportation and ridesharing services - whilst we’ve seen a number of horizontal ridesharing services, agnostic to consumer demographic, its interesting to think about which vertical network operators might service interesting niches

Sam