The vast scale of text data available on the internet, combined with ever-expanding computing power and advancements in algorithms, has led to the breakthrough of ChatGPT. Since then, we have seen continuous improvements across various benchmarks, along with an expanded range of input types, including images and audio.
After two years of rapid and significant investment in this field, some concerns arose regarding whether progress could continue due to a potential shortage of training data. It is widely believed that models like GPT-4 were trained on almost the entirety of the internet. Given that these models function as distilled versions of their training datasets, a reasonable concern was whether they could distinguish novel insights from errors or anomalies, anything not present in their training data.
However, in 2024 a new paradigm called ‘reasoning’ emerged, showcased by models such as DeepSeek R1 and OpenAI’s O-series. This breakthrough enabled AI to achieve unprecedented performance in fields where the accuracy of outputs could be reliably verified, such as coding and theorem proving.
In these domains, AI can explore multiple hypotheses, different versions of a computer program, for instance, even if they were not explicitly represented in the training data. A verification system, such as software test cases, can then validate correct solutions and reward the model accordingly.
Programming and solving mathematical problems fall within the realm of digital outputs, but can this approach be applied to real-world scenarios? Robotics and automation provide excellent examples. For instance, a robot can be instructed to pick up an object and be rewarded upon successfully completing the task.
Recent AI advancements have enabled robots to fully understand products, storage bins, containers and other essential elements required for warehouse operations. By relying on computer vision, robots can determine the properties of items and the best way to grasp them. AI can also differentiate between bundled and individual items when counting stock. With reasoning models extending into robotics, we believe it will soon be possible to achieve 100% picking coverage for stock keeping units (SKUs) in warehouses. This means that every item in a warehouse could be picked by a robot, bringing us closer to fully automated, ‘lights out’ warehouses.
To achieve this, highly specialised robotic grippers will be needed. These, in turn, require advanced perception and control capabilities to handle items effectively. Given the current pace of technological progress, these developments should be within reach in the next few years.
One natural question that arises is whether these advancements will lead to widespread adoption of humanoid robots? As we discussed in a previous blog post, we do not believe humanoid robots will be the most cost-effective solution for future factories and warehouses. These environments can be designed with efficiency in mind by using dedicated robotic arms and manipulators.
From a maintenance perspective, it is far easier to keep spare parts for a robotic arm with a maximum of 10 joints than for a humanoid robot, which would require at least 100 joints.
AI is a powerful enabler, and while it will allow us to achieve fully automated warehouses, this is only the beginning. The next step is automation in semi-structured environments such as hotels, restaurants and hospitals, where robots will gradually take over tasks currently performed by humans. I expect this shift to occur within this decade, although humanoid robots may not dominate, wheeled and four-legged robots offer significant advantages in mobility and should see widespread adoption.
Looking ahead to the next ten years, as more robots are manufactured and deployed, their cost will decrease due to economies of scale. This will have several implications:
What exactly societies will produce ‘more of’ remains uncertain, just as it was difficult 20 years ago to predict the full impact of the internet and smartphones.
On a personal note, I am excited about the creative potential unlocked by these powerful new tools. A personal AI assistant could become the ultimate teacher, with infinite patience, offering everyone the opportunity to be a lifelong learner.
Get in touch with Nomagic to see how our innovative technology can take your fulfillment processes to the next level.
