English | EN 
Since the release of ChatGPT by OpenAI, the tech industry has been in a frenzy, playing catch up and assessing the business impacts of AI chatbots. While the success of Large Language Models used in AI chatbots promises innovation in a variety of applications, it may also mask a significant shortcoming of the AI technology. As AI quickly finds its place in programming, development and design workflows, it remains markedly absent from real world physical applications.
Constrained Intelligence: AI in a Box
Like robotic technologies, AI struggles to find use cases outside of controlled environments. In fact, ChatGPT is a great example of this. A beta release of ChatGPT’s underlying technology, GPT-3, was available to developers two years prior to the release of ChatGPT. The biggest change between the initial release of GPT-3 and ChatGPT is that ChatGPT packaged GPT-3 in a rule-constrained environment.
When we use the term artificial intelligence, we invoke Alan Turing’s “thinking machines” and robots featured in Isaac Asimov’s sci-fi collection I, Robot. Since around 2015 tech journalism has been awash with stories about Amazon delivery drones and Tesla full self-driving vehicles. Yet, examples of AI deployed at scale in unconstrained real-world environments are sparse.
The Challenges of Breaking AI Out of its Cage
In 20151, Elon Musk predicted Tesla, Inc. would achieve complete autonomy in their production cars by 2018. More than 8 years later, we have yet to see fully autonomous self-driving outside of limited test trials. Currently, there are no fully autonomous production cars on the market. Now to be fair to Musk, the development time of innovative technologies is unpredictable by nature. Many of the true challenges an innovative project must overcome will only be revealed through brute trial and error.
The real world remains one of the biggest challenges in computing and indeed in engineering more generally. The more one ponders the problems presented by the real world the more the real world seems computationally intractable. However, I am not so pessimistic. In the context of autonomous technologies, it is essential to remember that vehicles, drones and robots will still operate within delimited environments.
The Importance of Delimiting the Environment
Explicitly delimiting the environment, or problem space, of a technological solution is critical when building for the real world. However, I believe many engineers understand this intuitively, it runs counter to many research and development workflows. In my experience, many R&D projects start by identifying the most significant challenges, prioritizing these challenges by difficulty, and then working through them one by one. This approach is great for solving many types of problems. However, the effectiveness of this approach relies on a team’s ability to correctly identify the most significant challenges up front.
It is often the case that when we build for the real world each problem solved produces new problems. This can create a recursive loop development cycle. This often leads to waste as development teams develop tunnel vision. The focus on the specific problem of the day leads engineers to lose sight of the bigger picture. With missed deadlines, projects in a recursive cycle become pathological. Inevitably, the project dies when it exceeds its budget.
By delimiting the application space first, a team can avoid some of the difficulties created by an iterative challenge-oriented approach. A “delimitation first” approach enables the development of hybrid and modular solutions. By carving up a problem space, we can eliminate spaces with existing solutions and distribute the remaining spaces amongst teams for focused development.
The Risks of Universal Solutions
With full self-driving, it appears that Tesla is attempting to create a universal all-in-one solution. The car itself is responsible for all tasks. Yet, there are some choices that seem odd. Like Musk’s insistence on a vision-only based system. This choice appears to stem from Musk’s approach of cutting all the fat and a little bit of muscle, then working back to a solution from a deficit. This is a very effective approach at developing highly cost-effective solutions. However, self-driving could be a problem that cannot be optimized for cost.
It perplexes me that no one is advocating for nationwide roadside infrastructure like sensor networks to assist autonomous vehicles. The open public infrastructure would offload much of the complexity faced by autonomous vehicles. This would significantly reduce the scope of the challenges faced by automakers. While cost would be a significant factor, there is precedent for this. With nearly 40,0002 fatal motor vehicle crashes per year in the United States (as of 2021), it should be framed as a public health and safety measure. While this is an order of magnitude, less death than that caused by smoking in the US according to the Centers for Disease Control (CDC)
The allure of creating universal solutions has been an ambition of many of Silicon Valley’s progenitors. Yet, a solution that must succeed in every possible situation will encounter every possible problem. History shows overly ambitious targets without delimitation mostly lead to over-extension, wasted resources, and burnout. The lesson from Tesla’s ongoing journey to full self-driving capability is emblematic of this challenge. As a technology enthusiast, I hope for Tesla’s success, but with full self-driving, I sense they are making a tough problem needlessly more difficult.
