Language skills won’t be enough to stay ahead of China in the economic sectors that matter.
China will release its next five-year plan next month. These plans are important because China pours billions of dollars into, and otherwise gives favorable treatment to, the sectors targeted in these programs. The U.S. will need to use a whole new class of artificial-intelligence models — built for the world of science and math, not language and images — to stay competitive with China as it puts the pedal to the metal in critical sectors.
Beijing’s last five-year plan included a focus on biotechnology, as an example, and China is now rising quickly in novel biopharma IP and clinical trials, often outpacing Western rivals. This new plan, the 15th under Communist Party rule, will continue with a focus on AI and quantum technology and double down on other key areas in the physical world. In particular, the plan will focus on novel material science, which will build on China’s dominance in critical minerals. Novel materials include new metal alloys as well carbon-based materials that can be stronger and lighter than traditional materials such as steel. Novel material science also includes de novo chemical combinations for batteries that can store more energy with faster charge cycles and lighter weight. Applications include hypersonic missiles, advanced chip manufacturing, military ships and batteries for energy storage.
AI models trained on social media and Pinterest images are poorly suited for these applications, which need subatomic-level physics. We need new AI models for the real world — quantitative models trained on lab data and equation-based outputs that let them leapfrog current technology. Hallucinations aren’t an option for mission-critical applications. The country that has the best quantitative AI models and quantum technology will dominate the rest of this century. The Chinese know this and are focused accordingly.
Areas of real-world AI include pharma, semiconductors, energy and financial services — four sectors that represent more than $25 trillion of global economic output and that thrive on numbers and equations, not language.
The AI with which American users are familiar is large language models, which can pump out text, songs, images, videos and other digital media. AI for the quantitative world is something else entirely, focusing on creating novel medical treatments, de novo material science, and advanced risk management and portfolio construction. The fields of robotics and autonomous driving are also moving quickly and need quantitative models that help these platforms navigate the world.
Ilya Sutskever has recently been adding his voice to the “beyond LLM” camp, stating that we won’t reach artificial general intelligence with language models alone. Gary Marcus has been heralding this theme for years, now in good company. World Labs and AMI are two new companies focused on these quantitative models.
There is consensus in the AI world that the model builders have vacuumed in almost all possible data that one can find accessible online. New data that will change the course of AI models won’t come from the Web, but from novel sources that are sector-specific. For the development of new medical treatments, we need data sets created from automated labs that run thousands of experiments. These robotic labs are emerging in the West and in China. The lab output is then augmented with synthetic data based on the equations of physics and chemistry to expand the data sets further. These data sets are then used to train proprietary models with novel architectures that are adapted to the sector employing them.
The need for new models is especially acute in defense and energy. China is outpacing the U.S. in developing hypersonic missiles. These missiles travel at Mach 5 or faster and are much more difficult to track and shoot down than conventional versions. Novel materials help make these missiles lightweight and strong. China also controls the battery market, with 92% share of battery-grade lithium and 85% of the market. To compete, we have to move beyond lithium-based batteries to novel chemistries.
Similarly, new materials and processes are needed for advanced chip manufacturing. As chips pack in more transistors, we are approaching physical barriers. Novel materials and designs are needed to keep Moore’s Law progressing. Quantitative AI models are needed to design these new materials and topologies for chips. There are many possible combinations of elements to make these materials, only a few dozen of which might be viable to take to scale. Computing the optimal mix of elements is a task far beyond traditional tools.
Computing risk and optimization in financial services is also critical to our economy. The complexity of fast-changing markets and trading strategies is outstripping the capabilities of traditional risk tools. In the world of insurance, for instance, the same risk models have been used for decades, but they don’t capture the temporal and high-dimensional resolution needed for this market. Insurance companies are flying blind when they want to compute the blended risk of taking on additional pools of policies from other brokers.
In asset management, a crisis is growing as managers gain more assets, now often in the trillions. The size and scale of these managers means that small movements on their part can move the entire market. More complex structured instruments as well as increased leverage in the private markets that are now available mean that the real-world footprint of many assets are many times their notional value. Language models aren’t appropriate tools for these critical tasks, and the viability of our financial markets depends on getting this right.
China is intent on dominating key sectors of the global economy and has proved it can do so in areas such as critical materials. The laser focus in its new five-year plan puts the U.S. and its allies on notice that we will have to leapfrog with AI for the real world if the West is to edge out China.
Mr. Hidary is CEO of SandboxAQ.
Read more on The Wall Street Journal
