With the advancement of AI technologies and tools, could their applications transform applied science? By utilizing large language models (LLMs) and sophisticated algorithms, these applications have the potential to drive scientific breakthroughs, streamline routine processes, and continuously propose new hypotheses.

From ancient stone tools to architectural marvels, scientific breakthroughs, and industrial advancements across civilizations, human progress has always been driven by its tools. Today, the rise of artificial intelligence marks the beginning of a new era, set to revolutionize industries and expand capabilities.

Despite remarkable technological advancements, the pace of progress in applied science seems to be facing challenges in keeping up with the increasing complexity of global issues, requiring intensified efforts and achievements. This paradox is attributed to the growing complexity of scientific problems, which require many interdisciplinary approaches and sophisticated tools. Many academic and research institutions enforce strict standards that can slow the pace of progress. Meanwhile, applied science demands exceptional skills, capabilities, and even behavioral traits that are often found only among a small elite in each field. Additionally, there is a growing gap between the skills once necessary for scientific advancements and those required today. While some individuals develop these skills through personal initiative, others risk losing them entirely due to lack of practice and evolving demands.

The dimensions of this paradox are evident in the spread of the "publish or perish" culture in academic circles, which turns researchers into workers on a frantic production line that makes publishing and recording research a goal, reducing the importance of scientific impact and stifling creativity.

Deep Science Ventures has emerged to lead a pioneering endeavour aimed at changing this reality using large language models (LLMs), as it saw great potential in them, especially in light of the widespread recognition they have recently received.

More than a year ago, the company began experimenting with an AI tool known as “Elman,” which is designed to improve various tasks within the company. Its algorithms can analyse detailed and specific requests and orders to the extreme and implement them to simplify operations. For example, it can find the required basic information from within millions of papers in less than half a minute, or identify the root causes of problems, or conduct instant searches.

Through successive experiments, the company's team discovered that this method could reduce the time required to develop a concept from months to weeks. This, of course, depends on the user's proficiency in the skill of asking the right questions in a precise formulation and at the appropriate times, which is related to the skills and knowledge related to the field of work itself.

Of course, this does not mean training a model on scientific papers and then directing it to become a genius. What the team aspires to is to elevate everyone who is active in the research field, so that technology is transformed from a tool to a work companion who participates with the researcher in thinking and reasoning, just as personal AI assistant applications help designers, software developers, and others.

The company also launched what it called a "The Venture Science Doctorate," which targets future scientists and aims to develop basic skills at an early stage of the innovation process.

The next step will be to combine computational research with human-like reasoning, and this requires effective methods in areas lacking inference data, as large language models do not infer in the true sense, but rather draw a line from similar inference data in their training data.

These methods should also be able to deal with the complexities of real-world problems. This is what the company is working to achieve in its high-quality applied scientific reasoning model, where data collected over seven years is used so that it can convert 90% of the concept into real results.

This idea has been presented in other experiments, such as the application of AlphaTensor from Google DeepMind, which produced new algorithms, or Google's FunSearch program, which developed the experiment by integrating a computational research methodology to rewrite programs that solve math problems.

At Stanford University, for example, a team of developers was able to create a fully AI-powered electronic researcher, which performed well in dealing with known problems, but any slight change in wording or hypothesis was enough to make it fail. This highlights the challenge of the sensitivity of large language models to context and wording, which is a big gap, especially when compared to human creativity, which involves "constrained" research through a very wide combinatorial space, meaning that its simulation requires much more than just basic language models. The company believes that approaching this challenge may be by taking advantage of broader computing concepts, perhaps expanding the horizons of research.

Today, Deep Science's focus is on scaling its innovations to generate long-term strategies, address more complex problems, increase its user base, and create a seamless user experience that encourages safe multidisciplinary interactions.

The company wants its software to perform directed, optimized, and applicable functions, and to enrich its data set with high-quality scientific thinking sources. At the same time, it aims to develop a phased process, methodology, and skills library for multi-step reasoning. It, like other pioneers in this direction, believes that this endeavour deserves attention and effort, and that it will ultimately lead to a change in the global innovation landscape.

References:

• https://acceleratingappliedscience.substack.com/p/the-path-to-applied-science-co-pilots?r=mxe8&utm_campaign=post&utm_medium=web
• http://elman.ai/
• https://web.stanford.edu/~chadj/IdeaPF.pdf
• https://deepscienceventures.com/venture-science-doctorate
• https://arstechnica.com/information-technology/2022/10/deepmind-breaks-50-year-math-record-using-ai-new-record-falls-a-week-later/
• https://deepscienceventures.com/venture-science-doctorate