AI’s dramatic reduction in development costs for software has led to a shift in innovation and investment focus. In most embodiments this means companies working on hardware, and more particularly, means companies working at the intersection of AI and hardware.
For those with a seasoned background in tech, this comes in stark contrast from past years, when software companies with lower capital costs and faster development times were seen as a more appealing bet. This transformation has brought with it a change in optics, where those same challenges previously considered to be headwinds for launching hardware are now seen as a moat of protection.
Here we touch on some of biggest applications that have come out of this confluence between AI and Hardware. These are topics that we will cover in detail later in this series.
AI’s First Commercially Proven Application - Image Processing
Machine Learning applied to the field of Image Processing is arguably the first application of AI in a commercial environment. Electronic Manufacturers have been using Automated Optical Inspection (AOI) systems for over a decade now, with this technology considered to be a staple for any electronic production line. Other applications including Object Detection, Face Identification, and Character Recognition found equal levels of success, further solidifying that AI was ready for commercialization and here to stay. Fast-forward and the usage of Neural Nets for image processing has only continued to become more prolific, leading the path in applications such as Tesla’s Autopilot.
AI’s Biggest Succes – LLM’s for Software Development
To date, software written by Large Language Modules (LLM’s) such as Claude Opus and GPT Codex has the highest ROI of any AI application. This has had huge impacts on hardware teams needing to operate on two technical fronts by allowing efforts to be dedicated to the most salient engineering challenges. Noteworthy, much of the software developed in this vertical has helped accelerate integration between physical and digital systems. The benefit of highly connected hardware is something that Manufacturers have long since benefited from since with the advent of Industry 4.0. These same benefits in efficiency, speed, and quality from ‘smart’ hardware are now becoming more common, getting us to a place where highly connected physical systems are the norm.
LLM’s Role in Knowledge work Acceleration
LLM’s most direct impact on solving problems in deep tech can be distilled down to one concept: Cross Disciplinary Knowledge. A defining characteristic of Deep Tech is that it is addressing problems that span across multiple engineering and scientific disciplines. One great example is Lab-on-a-Chip. It is a consolidation and miniaturization of laboratory processes onto a single chip, a problem where experts in microfluidics, molecular biology, optics, and MEMS fabrication are needed throughout development. AI, described by some as a ‘Specialist on Demand’ allows every team member to gain an understanding of neighboring disciplines, leading to better questions, collaboration, and innovation.
The Big Gamble - Reinforcement Learning (RL) for Robotics
Reinforcement Learning has been the main approach for training robots to edge ever closer towards that human-like flexibility. This is probably what most people think of when they imagine AI and Hardware, with videos of dancing humanoids and robotic surveillance dogs at the forefront of news for robotics. However, it is also the area that has been slowest to demonstrate real world value.
With how effective language models provide insightful responses across a wide range of questions, it’s easy to wonder why robotics has not seen the same level of success with performing generalized tasks. In our opinion, the lag behind LLM’s is not a fundamental problem with the approach or technology; it’s simply that navigating through the physical world comes with highly complex challenges that are not present in text-based interactions. This is reflected in the architecture of the human brain where roughly 8x more neurons are dedicated to handling vision related tasks than language related tasks.
We do believe that reinforcement learning will continue to allow robots to navigate through unstructured situations, but for the time being it’s probably best to temper expectations when comparing them to their biological counterparts.
Takeaway
One of my favorite aspects of AI & Hardware, is that these ‘wins’ are as much a product of what Hardware has to offer AI, as it is how AI is advancing Hardware Technologies. It’s a situation where physical systems, equipped with low-cost sensors, capable motion systems, and extensive digital connectivity; and AI with a proven track record of LLM’s & massive investment focus are coming together to be the next stage of technology progression. Truly exciting to see where things head too next.