Nvidia wants to make sure that human drivers don’t blame autonomous ones for their fender benders. 

The company is seeking to patent a system for “allocating responsibility” in interactions related to autonomous or semi-autonomous machines. Nvidia’s tech allows self-driving cars to determine “levels of responsibility” for different objects in the surrounding environment, allowing them to make better decisions. 

Nvidia noted that autonomous cars often use “static assumptions” to estimate or model the behavior of human drivers, meaning they assume drivers will behave predictably. “Since drivers do tend to make unnecessary or unexpected maneuvers in certain circumstances, these static assumptions may not be adequate,” Nvidia said. 

Nvidia’s tech uses neural networks trained on real-world data of interactions between vehicles and other objects and the “parameters,” such as speed or direction, associated with those interactions. This teaches the system to understand the dynamics of different interactions it may have with other vehicles, pedestrians, or objects. 

Nvidia’s neural network can then be used in real-time to estimate levels of responsibility in interactions with its environment. The system uses that data to determine controls, such as velocity, acceleration, or turning rate, helping the self-driving vehicle better navigate the environment and account for what’s around it. For example, if it notices a nearby vehicle is driving erratically on the freeway, it may move lanes to give the other driver space. This also allows the vehicle to more accurately comprehend the motion of other drivers, avoiding “overly conservative maneuvers,” the company said. 

In the event of a collision, this system can also indicate levels of responsibility for the crash to figure out who was the “largest contributor to the occurrence,” Nvidia said. However, the company noted that these fault contributions shouldn’t be “conclusory,” but instead used as an “informative tool or estimate of the responsibility allocation.” 

Nvidia has filed various patent applications to ensure that its autonomous machines avoid as many collisions as possible, both on the road and on factory floors. The company is a leading provider of self-driving vehicle technology, with revenue from its automotive sector reaching a record high in the last fiscal year and growing 11% year over year in the first quarter. The company also forged a partnership with Foxconn late last year to develop “AI factories,” or data centers dedicated to developing self-driving cars.

In an investor call in February, Nvidia CFO Colette Kress said that “nearly every automotive company working on AI is working with Nvidia.”

Its strength in the AI market generally makes Nvidia a power player in the autonomous car market, said Bob Bilbruck, CEO of consulting firm Captjur. The company is the No. 1 purveyor of AI chips, easily beating out competitors like AMD and Intel. Its first-quarter earnings cemented its AI dominance, with revenue tripling to $26 billion and net income rising seven times over to $14.8 billion year over year. 

“They own the AI chip technology that’s going to power most of these systems,” said Bilbruck. “They have a huge lead.” 

However, self-driving cars still have a public perception issue, as many consumers simply don’t trust autonomous vehicles. And National Highway Traffic Safety Administration investigations into companies like Tesla, Google-owned Waymo, and Amazon-owned Zoox likely don’t help.

But as self-driving vehicles get smarter, Nvidia’s tech looks at the safety problem from an entirely different angle, aiming to protect autonomous vehicles and their occupants from the recklessness of human drivers, said Bilbruck. 

One other consideration: Tech like this also stands to revolutionize a process that everyone loves going through when they get into a fender bender: insurance claims. Nvidia noted in the filing that its tech should be used to inform decisions – not make them. However, a system that uses an AI algorithm to assign fault to one party or another could be tempting, Bilbruck said, both to insurance companies seeking to automate operations and autonomous vehicle firms looking to clear their names.

Shopify wants to help you touch up your ads – even if you’re not great at Photoshop. 

The company filed a patent application for “tuning AI-generated images.” This patent would seemingly expand on the company’s existing AI tools, allowing merchants to personalize AI-generated images used in advertising. 

While generative models may create images rapidly, editing these images can prove tedious, Shopify noted. “Where the images generated by a model are determined to be deficient, users may need to re-run the model by, for example, changing text prompts multiple times until the output is satisfactory,” the company said. 

Shopify aims to offer “a “post-processing” quality assurance layer” for this very purpose. To start, Shopify’s system collects a bunch of data from a merchant’s online storefront, such as product images and categories, as well as customer interaction data, such as purchase data, “dwell time” and clickthrough rate. 

This data is used to train a deep learning generative model, specially tailored for that merchant. To make the model more robust and ensure that it’s accurately depicting products, Shopify would train these models on large sets of “regularization images” of products in the same category. For example, if a merchant has a leather jacket on its website, it would train the model on other images of leather jackets as well.

Finally, Shopify’s system would allow merchants to generate and edit photos of their products. The tech may rank images on indicators of “photorealism,” such as human poses or structural and lighting anomalies, and assign “aesthetic scores” to generated images. 

Shopify started launching generative AI tools in early 2023, with its first offering aiming to help merchants generate descriptions for their products. The company has since added a tool called “Shopify Magic,” which allows merchants to use generative fill for photo backgrounds, as well as AI-powered conversational search. 

While Shopify has a number of big-name clients on its roster like Gymshark, Redbull, and Allbirds, many are small business owners or merchants with limited time and marketing budgets. In these cases, generative AI tools like this can help these companies create ads and product images at a quality they previously couldn’t achieve. 

But adding these tools stands to benefit more than just their merchants, said Ryan Doser, VP of Inbound Marketing at Empathy First Media. The company’s moves line up with the broader AI push in both the tech and e-commerce industries. And after the company saw its stock price plummet after reporting losses in the first quarter, AI may help drum up investor excitement. 

However, as always, there are two sides to the AI coin. For one, It’s not hard for an image model to succumb to bias, said Doser, with Google Gemini’s historically inaccurate image fiasco being a prime example. “You really have to think about who are these companies behind the generative AI software,” said Doser. “Whether it’s OpenAI or Microsoft or Google or Adobe, there’s always going to be some sort of bias depending on which tool you’re using.” 

Additionally, while it may be tempting to zhuzh up your product images with generative AI, merchants may run the risk of false advertising, whether from an overly polished image or AI hallucination. That’s why these tools are often better used for ideation, Doser said. “I don’t know if I would recommend using it for an actual polished campaign that’s ready to go and put ad dollars behind.” 

Adobe wants to be kept in the loop. 

The company filed a patent application for “predicting video edits from text-based conversations” using neural networks. As the title of the patent suggests, Adobe’s tech essentially aims to pre-emptively edit videos for users by guessing what needs to be done from their messages. 

The goal is to streamline the editing process by reading messages between two or more parties and figuring out if those messages contain video edits. Adobe noted that these messages can be between two users either within the video editing system or in an external messaging system. 

After collecting this data, the system uses neural networks to map content from the messages to different editing operations and to predict the “parameter values” of the editing operations to the video sequence. For example, if one user says in a conversation that the brightness of a shot is too low, Adobe would use that information to make the shot brighter. 

The company noted that, while existing text-based editing solutions offer the ability to modify image data, these tend to be time and resource-intensive for editing videos. These systems also tend to require detailed and explicit instructions, and generally don’t pick up context from conversation “which generally do not contain explicit textual descriptions of edit operations.” 

It’s no secret that Adobe has been putting the pedal to the AI metal. The company has sought to claim several inventions as its own, filing applications for data visualization tools and AI summarization, while also taking on issues like hallucination and bias.

The company has been loading its app suite with AI integrations, mostly building on top of Firefly, its generative image model released last March. The company has embedded the model throughout its image-focused programs like Photoshop, Lightroom, Illustrator, and Indesign. 

With its AI efforts, the company’s goal is seemingly to democratize design, making it easy for anyone to access with little to no training or experience. “I think generative AI is going to attract a whole new set of people who previously perhaps didn’t invest the time and energy into using the tools to be able to tell that story,” Adobe CEO Shantanu Narayen told The Verge’s Nilay Patel on the Decoder podcast last week. 

And soon, the tech may make its way to video: in mid-April, the company previewed several generative AI tools for Premiere Pro, which aim to “streamline workflows and unlock new creative possibilities,” the company said. New capabilities, coming at some point this year, include extending scenes, adding and removing objects, and text-to-video tools. This patent could offer a glimpse at what may be in store. 

While the company noted that it’s in “early explorations” of bringing third-party generative models into its platform, it’s also building its own video model using Firefly, potentially competing with OpenAI’s Sora and Google’s recently announced Veo.