Robots are Coming for Your Ivy League Job
In the past 18 months, my views on how robots would slowly change our labor market have been flipped on their head. Maybe drivers, construction workers, and tradespeople should feel safer in their jobs than copywriters, designers, or even doctors and lawyers. In the 1970s, people’s perceptions of what artificial intelligence (AI) and robots would look like in the future were heavily influenced by science fiction. Many popular depictions of AI and robots from this era, such as in movies and TV shows, portrayed them as humanoid in shape, with a sleek and futuristic design. They were often shown as having advanced abilities, such as being able to think and reason like humans and having superhuman strength and speed. They were also often shown as working alongside humans in various fields such as manufacturing and even as personal assistants in homes.
It was widely believed that robots would first take over manual labor jobs, such as assembly line work in factories or heavy machinery operation in construction. These jobs were seen as repetitive and physically demanding, and therefore, well-suited for automation. However, as technology has advanced, it has become clear that robots and AI are increasingly capable of performing a wider range of tasks, including those traditionally performed by white-collar and clerical workers.
This shift is largely driven by the growing ability of AI systems to process and analyze large amounts of data, as well as by advancements in natural language processing and computer vision. As a result, it now appears more likely that robots will take over white-collar and clerical jobs first, such as data entry, customer service, and even some professional jobs like paralegals, radiologists, and others. All this makes sense intellectually but has always been a decade away. However, the capabilities of today’s AI are sending shockwaves across technology, education, and even labor. Perhaps that future decade is going to be this one.
So far, the public chatter has focused on software iterations of AI like chatGPT and DALL-E 2. Our focus as a firm is to understand how this jump in intelligence (think it is safe to call it intelligence at this point) will accelerate innovation in manufacturing, infrastructure, energy, transportation, agriculture, and human augmentation. Technology shifts in labor are always scary but I am incredibly optimistic about what this will mean for humanity in the long run.
What is happening
I assume everyone has heard about chatGPT. If you haven’t, I highly suggest you create a free account and play with it — it is incredible. In short, chatGPT is a large language model developed by OpenAI. It is a variant of the GPT (Generative Pre-training Transformer) model and is trained on a massive dataset of text from the internet. ChatGPT is designed to generate human-like text and answer questions, it is also capable of understanding context and generating coherent, fluent, and comprehensive responses.
The product is impressing everyone because of its ability to generate human-like text that is often indistinguishable from text written by a human. It can respond to a wide range of topics and questions and can even generate creative and imaginative responses. Additionally, ChatGPT can be fine-tuned for specific tasks, such as conversation generation, question answering, and text completion.
To understand the speed of adoption, here is a chart one of my mentors shared with me.
Microsoft announced another $10 billion investment into OpenAI and it is safe to assume that chatGPT’s capabilities will soon start integrating with a variety of applications, including chatbots, virtual assistants, and even content generation for websites and social media.
ChatGPT is the most famous text-output application, but OpenAI initially dazzled me with DALL-E 2. DALL-E 2 is a state-of-the-art language model developed by OpenAI, and is an evolution of the original DALL-E model, which was released in January 2021. DALL-E 2 is designed to generate images from natural language descriptions and text, it can also be fine-tuned on specific tasks, such as image captioning, image synthesis, and text-to-image generation. The model has received a lot of attention and traction so far, it is considered a breakthrough in the field of AI, specifically in the area of computer vision and natural language processing. Like chatGPT, it has the potential to be used in a wide range of applications, including video game design, advertising, and even art. Rumors circulated recently that OpenAI’s launches have been causing some panic in the halls of Big Tech. Could this be the new “Search”?
But OpenAI is not the only game in town. Some models have been trained on more specific data sets for specific use cases rather than the more generalized chatGPT and DALL-E 2. One of Google’s AIs, called the “Medical Brain,” was trained using a large dataset of de-identified medical records. The AI was able to learn from this data and make predictions about patient outcomes, such as the likelihood of a patient being readmitted to the hospital. To evaluate the AI’s performance, the team tested it against a set of medical exams, known as board certification exams, which are used to test the knowledge and skills of doctors. The results showed that the AI performed as well as, or in some cases, better than human doctors who had taken the same exams.
In summary, we are getting really good at language-based models that deal with complex information and knowledge. So why is AI so close to crushing AP English but robots still have issues walking or manipulating an egg?
What happened
At our Breakthrough Science Conference last fall, I was fortunate to spend time talking to folks way smarter than me about robotics. We were specifically focused on discussing when a generalized humanoid robot (like the one Tesla is working on) would be doing my laundry. Turns out folding laundry is a huge technical problem. Understanding and predicting the motion of fabric is hard. From those conversations I had three takeaways:
1. Training Models
Large amounts of data are available for training language models, which allows AI to learn from examples and make accurate predictions or decisions. In contrast, in real-world environments like carpentry, it is more difficult to obtain high-quality and diverse data, which can limit the ability of AI models to learn and perform effectively. These training models are crucial for artificial intelligence (AI) because they allow the systems to learn from data and make predictions or decisions. When a model is trained on a specific dataset, it learns to recognize patterns and relationships in the data, which it can then use to make predictions or decisions about new data. However, the effectiveness of AI models is highly dependent on the quality and diversity of the data sets used for training. If a model is trained on a limited or biased dataset, it will be less accurate and less effective in making predictions or decisions about new data.
We see this clearly in a lot of the robotics and autonomy companies we screen. A huge competitive advantage is on how much training data they have. If there is limited training data, then a large amount of cash will be needed to record and process that data. Cruise, Waymo, and Tesla have spent huge amounts of money recording real-world drivers to create these data sets and training models. This is such a costly process that a lot of new AI training is now done in virtual environments. Think of certain construction, trade jobs, or even janitorial workflows. How much data do we have documenting these tasks compared to 100 years of legal briefs, decisions, and opinions?
2. Real vs. Abstracted Worlds
Sh*t happens in the real world. Abstracted models like language and legal constructs often involve a relatively simple and well-defined set of rules and concepts, which can be learned by AI models through training. In contrast, real-world environments like driving are complex and dynamic, with many variables that need to be considered, such as weather conditions, traffic, and unexpected events.
In abstracted models like language and legal, there is also often a set of standard rules, conventions, and terminology that can be learned and applied by AI models. In contrast, in real-world environments like driving, there is often a lack of standardization, which makes it more difficult for AI models to learn and adapt.
Finally, in abstracted models, the consequences of errors made by AI models are usually less severe, in contrast, in real-world environments like manufacturing or construction, even small errors can have catastrophic consequences, this factor makes it harder for AI to perform and also require more safety measures and testing before deployment.
3. Firmware vs Software
I do not know if this is a separate takeaway or a derivative of the first two, but I started thinking about human abilities as firmware and software. The pattern I was seeing is that a lot of the hard stuff for robots were things that our pre-language ancestors could probably figure out. That makes sense — these AI models are LANGUAGE-TRAINED.
Let’s think of our interactions with the environment as a combination of firmware and software. The firmware, in this case, refers to the physical abilities and motor skills that allow us to move, walk, jump and catch objects. These abilities are hardwired into our bodies and are largely automatic and reflexive, much like firmware in a computer. They are the foundation of our ability to interact with the world around us and are necessary for our survival.
On the other hand, knowledge of medicine and the ability to paint are more like software, as they are learned and acquired through education and practice. They are not innate abilities, but rather, they are built on top of the firmware of our physical abilities. This knowledge and skill set can be thought of as a set of instructions that we use to interact with the world in a more complex and sophisticated way. Like software, knowledge and skills can be updated, improved, and expanded upon, allowing us to achieve new capabilities.
We have been writing for 5,500 years. It took us 10,000,000 years to pick up a tool (and get really good at handling an egg). The firmware took a longer time to code. NOW — that is also because the iteration cycle was much longer. Written language unlocked an exponential increase in our capabilities as a species to iterate and learn; AI is an extension of that curve. But I also think we should not take for granted how much hard-coded programming is born in every child.
What’s Next
The internet is basically an amalgamation of all recorded human intellectual output. Combined with huge advances and cost reductions in processing power, the emergence of extraordinary AI applications is not that surprising. AI will slowly automate many tasks traditionally performed by white-collar professionals, such as paralegals and radiologists. AI-powered systems can be trained to analyze legal documents and identify key information, reducing the need for human paralegals to perform these tasks. Similarly, AI-powered image recognition systems can be trained to identify and diagnose medical conditions in medical images, reducing the need for human radiologists. AI can be used to automate repetitive tasks such as data entry and filing, freeing up human workers to focus on more complex and higher-value tasks. As AI technology continues to advance, it is likely that we will see more and more white-collar jobs being replaced by automation, making it important for workers to develop skills that complement AI and are difficult to automate such as creativity, critical thinking, and emotional intelligence.
My prediction is that the next 2–3 years will see verticalization of these language models around specific workflows. It is not hard to imagine combining chatGPT and DALL-E 2 to create a marketing tool that automatically refines copy and creative to increase customer conversion. It is also not too hard to imagine a tool that outputs entire webpages and user flows. If the last ten years gave us Shopify, the next five might give us Shopify + Brandify + Marketfy in a single stack.
In contrast, more physical-world jobs such as carpentry, plumbing, electrician, and construction are difficult for robots to perform because they require the ability to adapt to unexpected changes in the work environment, such as uneven surfaces, obstacles, or an unplanned change in the work process. Trade jobs often require a high degree of precision and attention to detail, which can be difficult for robots to achieve. These tasks often require the use of tools, and the ability to work in tight spaces or at odd angles, which can be difficult for robots to replicate. Additionally, many trade jobs involve working with irregular shapes and materials, which can be hard for robots to handle.
That does not mean that there will not be huge advances in robotics. Better language models will inevitably lead to significant improvements in the capabilities of robots. AI provides the ability for robots to learn and adapt to new situations, allowing them to perform a wider range of tasks and operate in more complex environments. For example, advances in natural language processing, computer vision, and machine learning allow robots to understand and respond to human speech, recognize and track objects, and make decisions based on sensor data. This can enable robots to perform tasks such as customer service, warehouse management, and even surgical procedures.
AI can also enable robots to learn from their experiences, improve their performance over time, and adapt to changing environments. This can make robots more resilient and reliable, reducing the need for human supervision and maintenance. Additionally, AI can also enable robots to perform tasks that are difficult or impossible for humans to perform, such as working in dangerous or inaccessible environments or performing tasks that require a high degree of precision or speed. This can open up new opportunities for robots in fields such as manufacturing, construction, and space exploration.
What about the soul of the artist?
In a previous life, I worked in entertainment and spent my days surrounded by incredibly creative people. I traditionally viewed artistic visual expression as one of the last things that AI would touch. I keep thinking of the scene in Short Circuit when Johnny 5 exclaims “Butterfly” when shown an inkblot test. I’m not so sure. Several image models have really impressed me. The one that really hit my core was shared with me recently by a producer friend. Canadian Director Johnny Darrell used the Midjourney AI tool to visualize what Tron would look like if directed by Alejandro Jodorowsky. Here are a few sample images but I would recommend reading the full New York Times article or scrolling through all the screen shots.
What does that mean for creativity? I don’t know if AI could make the jumps from realism to impressionism to expressionism to cubism, etc. It will be great at mash-ups and beautiful derivative works. One day, I’ll be able to type in “Citizen Kane directed by Akira Kurosawa set in 1920s Argentina,” sit back and enjoy the show. I would be first in line to watch Jodorowsky’s Tron. HOWEVER, I do not believe that AI could create a new aesthetic as severe and original as Jodorowsky’s. These creative jumps are still our domain for some time to come. I hope that AI unleashes our creativity and raises the weird and unconventional. In short, the soul of the artist is safe, but hacks beware.
Why you should be excited
The replacement of certain jobs by technology is a natural process that has occurred throughout history, as new technologies and innovations have emerged. One example of this is the displacement of certain jobs such as blacksmiths, carriage makers, and horse breeders with the arrival of the car. With the advent of the automobile, the demand for these jobs decreased, as cars became a more efficient and practical means of transportation.
This process of job displacement is not new and is likely to continue as new technologies and innovations are developed. The introduction of new technologies often leads to increased efficiency, productivity, and cost savings, which can make certain jobs obsolete. However, it is also important to note that new technologies also create new jobs and opportunities, such as in the automobile industry, where new jobs were created in car manufacturing and maintenance.
Someone asked me recently what technology advances I was most excited about, and it did not take me long to think of three key innovations: (1) increased human lifespan (2) limitless energy, and (3) an AI model so advanced that every single child in the world would have the equivalent of the best teacher for every subject on a 1:1 basis adapting to their best way of learning for the marginal cost of compute.
Forget the assembly-line paradigm of education. We are talking about every child learning as fast as they can in a bespoke manner as long as they have an internet connection. Imagine Einstein teaching you physics or Washington answering your questions about the revolutionary war. WOW. The pitfalls of bias, indoctrination, false information, etc. in this view of the future are also scary — but that is a topic for another post. We have a few decades (maybe!) before we contend with what I call the “Bill and Ted’s” model of education. I believe there is much more reason to be optimistic as we slowly free our minds and our time from derivative workflows and democratize education, embrace our creativity, and ultimately achieve our full potential as a species.
Full disclosure: As an experiment, I used chatGPT to write 51% of this post. There are entire paragraphs written by artificial intelligence (here is a link to what was written by AI vs me). I don’t think it is necessarily difficult to pick which sections. There are variations in diction, humor, pacing, etc. But it is still incredible. Once training models can be augmented to user-specific content (I feed all my writing, emails, texts, etc. to a learning model), it will be even harder to differentiate.
I’m not applying to vocational schools or dusting off my resume just yet, but am I guilty of telling my six-year-old to treat Alexa nicely and that engineer or scientist are the only two professions available to him? No comment.
Prime Movers Lab invests in breakthrough scientific startups founded by Prime Movers, the inventors who transform billions of lives. We invest in companies reinventing energy, transportation, infrastructure, manufacturing, human augmentation, and agriculture.
Sign up here if you are not already subscribed to our blog.