#361: Foxglove w/Founder + CEO Adrian Macneil

Foxglove is the company founded and led by Adrian Macneil. In this episode, it appears to be the main guest company and likely the focus of the conversation.

Audi is the automaker the host references when asking whether the early Cruise vehicles were Audi-based. In this context, it refers to Cruise's earlier autonomous test vehicles before the company moved on to other platforms.

The Nissan Leaf is an all-electric hatchback that Cruise was working on in this early autonomous-vehicle context. It’s mentioned here as one of the platforms in the garage during Cruise’s development phase.

GM stands for General Motors, the company mentioned as a key partner in Cruise's early development. The reference shows the timeline around GM's involvement and the arrival of a pre-release vehicle.

The Chevy Bolt is Chevrolet's early long-range EV, and here it was a pre-release car sent to Cruise. The speaker is describing a transition from using the Leaf platform to the Bolt platform, which signals a shift in hardware and development approach.

A pre-release vehicle is an early production or prototype-level car provided before public launch. Companies use these cars for testing, development, and validation before the model reaches customers.

In automotive terms, a platform is the underlying hardware architecture a vehicle is built on. Here, the speaker is talking about moving development from one car's underlying structure and systems to another's.

3D printed parts are components made by additive manufacturing rather than traditional machining or molding. In this context, it suggests an early, experimental build environment with custom fabricated pieces.

'Driver out' refers to a self-driving system operating without a human safety driver behind the wheel. It's a milestone in autonomous vehicle development because it means the car can run its route on its own.

Developer infrastructure is the underlying software and systems engineers use to build, test, and deploy products. In autonomy, it includes the tools needed to manage data, workflows, and deployment at scale.

This refers to deploying autonomous systems broadly and reliably across many vehicles or robots, not just in a one-off demo. The speaker is emphasizing the need for infrastructure that can support large-scale autonomous deployment.

Self-driving is the broader category of vehicles or robots that can operate with little or no human input. Here it’s used as the baseline example for why the tooling could matter beyond cars.

AI, or artificial intelligence, is used here as part of the technology stack that has improved since 2016. In autonomy and robotics, AI helps systems perceive, decide, and act.

Compute refers to the processing power available for AI and software workloads. The speaker is noting that advances in compute since 2016 have changed what’s possible technically.

Cruise is the autonomous vehicle company that was developing self-driving ride-hailing and robotaxi technology. In this segment, the speaker is reflecting on Cruise's rise, deployment scale, and eventual setbacks.

An autonomy stack is the full software-and-hardware system that lets a vehicle perceive its surroundings, plan actions, and drive itself. It usually includes sensors, mapping, prediction, and control software.

A robotaxi is a self-driving taxi service with no human driver behind the wheel. It is one of the main commercial use cases for autonomous vehicle technology.

AV stands for autonomous vehicle. The AV era refers to the period when the industry was heavily focused on developing and commercializing self-driving cars.

Autonomous vehicles are cars or trucks that can drive themselves using sensors, software, and machine learning instead of a human driver controlling everything. In this segment, the speaker frames them as an early major use case for physical AI, especially because they have to operate safely around people in the real world.

Safety-critical systems are systems where a failure could cause injury, death, or major damage. For autonomous vehicles, this means the software has to be extremely reliable because mistakes happen on public roads around people.

Autonomous driving is the use of sensors, software, and control systems to let a vehicle drive itself with little or no human input. In this segment, it’s described as the first major real-world success for applying AI to physical systems.

Autonomous vehicle companies are businesses building self-driving systems for cars, trucks, and other vehicles. The speaker is referring to a newer wave of AV startups that are applying lessons learned from earlier efforts.

AI-native means a product or company is built around artificial intelligence from the start, rather than bolting AI on later. Here it describes autonomous vehicle startups designed around AI-first workflows and architecture.

Wabi is an autonomous vehicle company mentioned as an example of the newer AI-first generation of AV startups. It’s part of the broader discussion about companies learning from earlier self-driving efforts.

Wave is another autonomous vehicle company referenced as part of the newer AI-native AV ecosystem. It’s mentioned alongside Wabi as an example of a startup taking a modern approach to autonomy.

Boxsoft appears to be a platform or tool used by autonomous vehicle companies to build more efficiently. In context, it’s presented as infrastructure that helps AV teams move faster and leaner.

This refers to applying autonomous systems beyond cars, into real-world machines that operate in physical environments. The speaker is broadening the discussion from self-driving vehicles to robots and industrial equipment.

Bedrock is mentioned as a company building autonomous construction machinery. It’s used as an example of autonomy moving beyond passenger cars into industrial equipment.

Waymo Trucking refers to Waymo’s trucking-focused autonomous vehicle effort. It’s mentioned as the source of people who later moved on to build autonomous construction machinery.

Autonomous construction machinery is heavy equipment that can operate with minimal human control, using sensors and software to perform tasks on job sites. It’s an example of autonomy being applied outside road vehicles.

Consumer robots are robots designed for everyday personal or household use rather than industrial settings. Here they’re mentioned as another area where autonomy and AI are expanding.

Autonomous defense capabilities are defense systems that can operate with varying degrees of automation, potentially including drones or other unmanned systems. The speaker is referencing growing investment in this area as part of the broader autonomy market.

Drones are unmanned aerial vehicles that can be remotely piloted or operate autonomously. They’re mentioned here as one of the real-world machine categories benefiting from autonomy software.

An infrastructure layer is the underlying software and systems that other applications are built on top of. In autonomy, this can include data handling, telemetry, simulation support, and tooling for development and testing.

A data flywheel is a feedback loop where more usage creates more data, and that data improves the product, which then attracts more usage. In autonomous driving, it means more miles driven can lead to better detection of rare events and better system performance.

In data science, the long tail refers to rare, unusual, or edge-case events that don’t happen often but matter a lot for system performance. For self-driving cars, these are the weird situations that are hard to predict but important to handle safely.

Simulation is the use of a virtual environment to test how a system behaves before or alongside real-world deployment. In autonomous driving, it lets engineers evaluate software against many scenarios without needing to drive them all on public roads.

AV stands for autonomous vehicle. The AV industry refers to companies and researchers building self-driving cars, robotaxis, and related software and hardware.

To debug means to investigate a problem in software or systems and figure out why it happened. Here it refers to engineers analyzing robot or AV data to understand failures and fix them.

A data set is a structured collection of information used to train or evaluate a system. In this discussion, the data set is being expanded with interesting driving or robot events so the model can learn from them.

Training is the process of teaching a machine-learning model using data so it can make better predictions or decisions. In AV systems, training improves how the software handles driving situations.

A robotics startup is a young company building robots or robot-related software and hardware. Here it’s the broader business context for the discussion about detecting problems in autonomous systems.

In machine learning and robotics, low confidence means the system is not very sure about the decision it just made. That uncertainty can be used as a signal that something may need human review or a fallback action.

Safety systems are built-in mechanisms designed to detect problems and reduce risk before something gets worse. In this segment, they’re described as systems that can flag potential incidents automatically.

Data mining is the process of searching large datasets to find patterns, anomalies, or similar events. Here it refers to looking through incident data to find other cases that resemble a known problem.

A search problem here means finding specific examples inside a large dataset. The speaker contrasts it with building a model, emphasizing that sometimes the challenge is just locating the right incidents to learn from.

A mining problem means extracting useful patterns or rare examples from a large body of data. In this context, it’s about surfacing unusual driving events that can improve autonomy systems.

An autonomy company is a business focused on self-driving or driverless vehicle technology. The speaker is describing how these companies need major infrastructure, data, and research teams behind the scenes.

Tesla is an electric vehicle manufacturer that also develops advanced driver-assistance and autonomous driving systems. In this context, it’s being grouped with other autonomy companies that need heavy infrastructure and research investment.

Novel models are new machine-learning models or approaches that haven’t been used before. The speaker is describing how top autonomy teams focus on inventing new methods rather than just operating existing systems.

Novel architectures are new structural designs for machine-learning systems. In autonomy, architecture choices can strongly affect how well a system processes sensor data and makes driving decisions.

Fleet infrastructure refers to the systems needed to manage a large group of vehicles, including dispatch, maintenance, tracking, and operations. In AV companies, this is a core operational layer rather than just the cars themselves.

First principles means solving a problem by breaking it down to its most basic truths instead of copying an existing playbook. The speaker is saying they had to build their AV systems from the ground up because there wasn’t an established template.

This appears to be a transcription of Zoox, the autonomous vehicle company, though the transcript says 'Zoo.' The speaker is listing AV companies they spoke with, so this likely refers to Zoox in context.

A multimodal data visualization tool combines different kinds of data—such as video, sensor inputs, maps, and logs—into one interface for analysis. In AV development, this helps engineers understand what the vehicle saw and how the system behaved.

Barrier to entry is the level of cost, expertise, or resources required to compete in an industry. Here the speaker is arguing that AV development becomes impractical if only companies with billions of dollars can participate.

A logging file format is a structured way to store recorded data so it can be searched, replayed, and analyzed later. In robotics and software systems, logs are crucial for debugging what happened and why.

A bespoke file format is a custom data format built by one company instead of using a standard one. In this context, the speaker is criticizing how different companies each created their own way to store robot/log data, which makes sharing and tooling harder.

This is a metaphor for a standard file format for robot data, similar to how PDF became a universal document format. The speaker is saying MCAP is meant to be the common, portable way to package robot logs and data.

MCAP is a data container format used in robotics to store and exchange logged sensor and system data. Here it's being described as the standardized, portable format the industry should use instead of each company inventing its own file format.

SaaS stands for software as a service, where software is delivered over the internet as a subscription. The speaker is drawing a comparison between robotics data standards and how the software industry standardized around common tools and formats.

This appears to be the company or product name the speaker is referring to, likely a transcription variant of Foxglove. It is mentioned as the organization working on the robotics infrastructure layer.

This is the idea that physical components become cheaper, more standardized, and less differentiated over time. In automotive and robotics contexts, it often means the real competitive advantage shifts from the hardware itself to software, integration, or services.

A log file format is the structure used to store recorded data from software or hardware systems. The speaker is contrasting useful product work with low-level data formatting work, which can be necessary but not the main business value.

The trough of disillusionment is a stage in the technology hype cycle when excitement fades and weaker companies fail. The speaker is using it to describe the market cleanup that left stronger products and companies in a better position.

This refers to the widely reported October 2023 Cruise safety incident that led to major scrutiny and operational changes. The speaker uses it as a key reason Cruise became a separate case in the autonomous vehicle industry.

A robotics company builds machines that can sense, decide, and act in the physical world, often using AI and automation. Here, the speaker is comparing the difficulty of starting such a company in 2016 versus today.

AI inference chips are processors optimized to run trained AI models efficiently, especially in real time. They matter here because more capable chips let robots run local AI models onboard instead of relying on the cloud.

Batteries are the energy storage units that power electric systems, including robots and EVs. The speaker is pointing to battery improvements as one reason robotics is improving.

Local models are AI models that run on the device itself rather than in the cloud. In robotics, that can reduce latency and make the system more independent.

Actuators are the components that convert electrical or hydraulic energy into motion. In robots, they’re what move arms, hands, joints, and other mechanisms.

Dexterous hands are robotic hands designed for fine, human-like manipulation. They’re important in automation because they let robots handle more complex real-world tasks.

Sensors are the inputs that let a robot perceive its environment, such as cameras, lidar, force sensors, or encoders. They’re a core reason modern robots can operate more intelligently.

Open source vision language models are AI systems that can interpret images and text together, and whose code or weights are publicly available. The speaker is saying people are adapting these models for robot control tasks.

Fine tuning is the process of adapting a pre-trained AI model to perform better on a specific task or dataset. Here, it means customizing vision-language models to understand robot action data.

Robot action data is data collected from robots performing tasks, including movements, sensor readings, and outcomes. It’s used to train models so robots can learn better behaviors.

Reshoring manufacturing means bringing production back to a country where it was previously moved overseas. In this discussion, it’s part of the broader push to automate factories and build domestically.

Automation is the use of machines and software to do work with less human involvement. In manufacturing, it often means robots and control systems replacing repetitive manual tasks.

This is a discussion about domestic manufacturing and the political/economic push to produce goods in the United States. Here it’s being used as part of the case for more automation.

Aurora is an autonomous vehicle technology company focused on self-driving systems, especially for trucking and other mobility applications. It is mentioned here as part of the broader group of AV companies whose talent and ideas are spreading into other industries.

This means the main challenge is no longer the self-driving technology itself, but everything required to run the service at scale. That includes manufacturing, logistics, charging, cleaning, and city-by-city deployment.

This segment is about the challenge of expanding Waymo beyond a few cities. The discussion focuses on the difference between solving autonomy and actually operating a large robotaxi network.

This refers to the technical challenge of making a vehicle drive itself safely and reliably in all situations. The speaker contrasts it with the operational challenge, implying the driving tech may be less of the bottleneck than the business and logistics around deployment.

Edge cases are unusual or rare situations that automated systems may not handle well. In autonomous driving, these are the weird scenarios that can still trip up even very advanced software.

This refers to the production side of scaling a fleet of autonomous vehicles. Even if the software works perfectly, the company still has to build enough cars to support a large rollout.

For an electric autonomous fleet, charging is a major operational task. Vehicles need to be kept powered and rotated through charging infrastructure so they can stay on the road.

Cleaning is part of keeping a robotaxi fleet ready for passengers. Autonomous ride-hailing services need vehicles to be presentable and sanitary between trips.

Validation here means testing and proving that an autonomous system behaves safely across many scenarios before it is deployed. The speaker is emphasizing that even in simpler environments, edge cases still need to be checked carefully.

A Series B is a later venture-capital funding round than Series A, typically used to scale a company after initial product-market fit. Here it’s part of the discussion about the startup’s fundraising history.

A Series A is an early venture-capital funding round, usually one of the first major institutional investments in a startup. The speaker is discussing how much capital the company has raised across funding rounds.

Gatic is a company named in the client list, likely in the autonomous-vehicle or trucking space. It’s mentioned as one of the self-driving-related customers.

AV stands for autonomous vehicle, a vehicle that can drive itself using sensors and software. The speaker identifies autonomous vehicles as a major growth area for the platform.

This is a discussion segment about defense and aerospace as major investment areas for the company’s platform. It’s a topic marker rather than a specific automotive term.

This refers to vehicles that operate on land, as opposed to drones in the air or boats on water. In robotics and autonomy discussions, it usually means cars, trucks, and other wheeled or tracked machines.

These are large AI models designed to serve as a general base for many robotics tasks, similar to how foundation models are used in language or vision. The idea is to train one model that can adapt to different robots, environments, and tasks.

This means a robot’s ability to precisely control its hands, grippers, or tools to pick up and handle objects. It’s a major challenge because real-world objects vary in shape, weight, and fragility.

These are tasks that require many steps over a long period of time, rather than a quick one-shot action. In robotics, it means the system has to plan, adapt, and keep working toward a goal across multiple actions.

This segment is about robots used in warehouses to move packages and handle logistics. It’s a practical application of robotics focused on storage, sorting, and fulfillment.

This topic covers robots that inspect store shelves, check stock levels, and verify pricing in retail environments. It’s a retail automation use case rather than a vehicle-specific subject.

Symbi Robotics is a robotics company mentioned as an example of autonomous systems working in retail environments. They’re described here as using robots to drive through grocery store aisles and inspect inventory.