In the current AI landscape, creative and talkative large language models may be the star of the moment, but a quieter revolution is underway.
Quantitative AI, grounded in rigorous scientific modeling and high performance computing, is charting a new frontier. While LLMs are great for human language tasks, quantitative AI is tuned into complex tasks in science and healthcare, pushing the boundaries of what’s possible in these fields.
According to SandboxAQ, quantitative AI is not just the future of AI: It’s the AI shaping the future. The company, which started as Alphabet’s AI quantum computing unit (hence the AQ), became an independent startup in March 2022. Thus far, SandboxAQ is generating a lot of attention (and investments) for its large quantitative models, or LQMs. These AI models are trained on proprietary data generated using physics-based methods and have applications in diverse fields, including life sciences, energy, and chemicals.
AIwire sat down with SandboxAQ VP of Engineering and Lead Scientist Dr. Stefan Leichenauer to learn more about the company’s LQMs and how they are advancing scientific discovery.
First, What is Quantitative AI?
Historically, the term ‘quantitative AI’ has been primarily associated with financial services, where it underpins strategies like algorithmic trading and risk modeling. However, SandboxAQ is expanding the concept by applying its principles to scientific and technical domains as well.
“Quantitative AI, for us, is very simple. It refers to the general idea that you are meeting the data and meeting the tasks that you’re trying to solve. You’re meeting it where it is and discussing it on its own terms,” Leichenauer told AIwire.
(Source: SandboxAQ)
Quantitative AI models address problems and data in their natural context, particularly outside the realm of language tasks. This involves leveraging traditional numerical modeling, such as solving equations or simulating physical systems, and enhancing those processes with AI to extract more insights and drive faster discoveries. For example, in molecular modeling, quantitative AI works with the underlying physical properties and rules rather than funneling everything through natural language.
“You want to talk about the molecules using the language that they want to be speaking in, not a human language, like natural language, but the language of numbers: the language of equations, of physics, of chemistry,” Leichenauer says.
By using AI models that are aware of the underlying quantitative systems, it’s possible to interpolate and extrapolate data patterns in ways that respect the inherent rules of those systems, even when those rules are not explicitly defined.
Like in molecular modeling, the exact interactions between molecules might not be fully captured by existing theoretical models. However, quantitative AI can analyze the data and identify patterns or behaviors that align with the underlying physics or chemistry, even if those patterns were not explicitly programmed into the model.
From Molecules to Medicines
One of the key areas where SandboxAQ’s LQMs are driving molecular modeling advances is in drug discovery, a complex task that integrates quantitative AI, computational chemistry, and multiple verification processes.
Drug discovery involves identifying small molecules that bind to target proteins effectively while meeting constraints like manufacturability, safety, and scalability. SandboxAQ is focusing its efforts on the early stages of drug discovery, particularly on identifying small molecules that can effectively target specific proteins without causing adverse effects, Leichenauer told AIwire.
This involves leveraging two critical sources of data: experimental data from past chemical research and computational chemistry techniques. By applying known mathematical equations and verifying their accuracy with experimental verification, SandboxAQ combines these data sources with their quantitative AI approach to streamline the discovery process in identifying potential drug candidates.
To navigate the vast possibilities in molecular drug discovery, SandboxAQ combines off-the-shelf computational tools with proprietary solutions tailored for speed and precision. As Dr. Leichenauer explains, “Sometimes those off-the-shelf tools are too slow to really plug into this kind of process. So, we need to be faster. We needed to invent new tools, so we’ve done that.”
These custom tools allow researchers to quickly evaluate whether a molecule is suitable for a desired purpose, including whether it can be synthesized in a lab, which is a critical consideration for moving from computational modeling to real-world applications.
Verification plays a pivotal role in SandboxAQ’s approach. In quantitative AI, the ability to use computational methods to test and confirm predictions at every step of the process provides a high degree of accuracy. “Verification is a key piece of it, and it’s kind of a key thing that you can do in quantitative cases,” says Leichenauer. “You have a lot of handles and opportunities to verify what you’re doing through all kinds of different computational methods to develop a lot of certainty as you progress through a problem like this.”
Transformers, Tensors, and Tailored Tools
While transformers have become synonymous with large-scale AI models, SandboxAQ takes a more tailored approach, leveraging its diverse set of tools to address specific challenges in quantitative AI. “Transformers are great, they’re not just for language modeling, they’re a very useful kind of tool,” says Leichenauer. However, transformers tend to be best suited for big data problems and are computationally expensive to run, making them less ideal for applications where speed and efficiency are paramount. “We’ll use a transformer architecture when we need to, but we’re not wedded to it,” he adds.
Instead, the company also taps into alternative machine learning models, such as tensor networks. Tensor networks are specifically designed for modeling physical systems like those found in physics and chemistry. Originally developed for solving problems in quantum physics, these models excel at representing complex atomic and molecular systems.
Tensor networks also offer a significant advantage: they allow SandboxAQ to sidestep the need for quantum computers in many scenarios. By scaling up tensor network algorithms using GPUs, the company achieves results comparable to what might be expected from quantum computing. “By scaling up these other methods, like tensor network methods, you can get away without a quantum computer and just do it with GPUs,” says Leichenauer.
“We didn’t invent [tensor networks], but we happen to have really strong expertise. And some of the work that we’ve done recently, for example, is collaborating with Nvidia to accelerate and scale up these tensor network algorithms, which really allow you to push the boundaries on quantitative modeling of complex atomic and molecular systems, to do the kinds of things that you wish you had a quantum computer available so that you could model these very complex quantum systems,” Leichenauer told AIwire.
The Virtuous Cycle of HPC, AI, and Big Data
In the world of AI, the synergy between high performance computing, artificial intelligence, and big data has been described as a “virtuous cycle,” or a self-reinforcing loop where each element accelerates the advancement of the others.
SandboxAQ’s approach to quantitative AI follows a similar principle. Leichenauer explains how the company’s work builds on traditional HPC-driven numerical modeling to create a robust data layer. This data fuels AI models capable of streamlining calculations and predictions, creating a feedback loop where both the AI and the underlying simulations continuously improve.
To illustrate, Leichenauer draws a comparison to weather modeling, where increasingly refined data about Earth helps improve models that predict the planet’s climate and atmospheric conditions. In this context, the focus remains on improving a single model based on progressively better data from one planet.
However, SandboxAQ’s work comes with a twist: in the realms of chemistry and materials science, the task is akin to modeling the weather on countless planets, each with its own unique properties. “The AI model that you trained on Earth data is not immediately going to apply with the greatest accuracy on Jupiter or Neptune,” says Leichenauer.
This adaptability is where SandboxAQ’s quantitative AI stands out. By combining HPC techniques with AI models that respect the numerical and physical rules of each system, they can extrapolate insights across vastly different contexts. The result is a system that isn’t limited by the static nature of its training data but instead generates clean, reliable, and relevant datasets on the fly, the company claims, sidestepping the pitfalls of traditional AI approaches that often struggle with data integrity.
Scaling Solutions and Preparing for Growth
What’s next for SandboxAQ? Dr. Leichenauer says though the company is still an early-stage startup, 2025 will be a year of scaling up to do more: “We’re on a very aggressive growth path. Growing and scaling and getting out there to more customers, getting [our product] out there to more customers, is certainly a goal,” he said.
Along those lines, the company announced a partnership today with Google Cloud to integrate and optimize its LQM platform on Google Cloud, offering procurement and deployment of SandboxAQ’s solution through Google Cloud Marketplace.
Looking to the future of SandboxAQ’s technology, Leichenauer sees a new horizon for quantitative AI with the development of agentic systems capable of working with greater independence, a concept the tech world is eagerly exploring.
“Like everybody else, we’re very excited about agentic versions of these things. Having quantitative AI agents, or LQM agents, that are executing a lot of these workflows semi-autonomously is a very exciting next step.”
Related
This post was originally published on here
