Why Generative AI Alone Can't Do R&D
Why Generative AI Alone Can't Do R&D
But it can commercialize R&D faster
Wildfire spreads slower than generative AI does right now.
Take just ChatGPT. Many of its users wonder: can we use it to make progress on deep technical or scientific questions? Can it do R&D? It can already answer many questions better than Google.
Except, ChatGPT models language itself. It was trained to predict and generate the next word in a sentence. It understands the nuances of language, with unknown emergent capabilities -- but it wasn't designed for deep scientific or engineering concepts, experimental methods, or more importantly, skepticism in general.
R&D is a systematic and iterative process of developing new technology. Entire academic journals focus on R&D methodology. Hypothesis-testing, data analysis, experimental design -- questioning the data, its interpretation, generating alternative hypotheses, integrating with conjectures and theories outside the immediate context -- all these are central to R&D. And all live outside ChatGPT's training data and intended function.
In short: what's missing? The ability to analyze an idea with skepticism.
Generative AI conflicts with skepticism
Because it is generative, its skills are almost the opposite of skepticism. The point is to generate output, not to determine the truth. Like all generative AI, it will confidently hallucinate false results in an articulate way. For its intended use cases, even with hallucination potential, it can still yield a significant amount of value. Large efforts are underway to improve accuracy and get the model to check facts as well as indicate its uncertainty.
R&D however requires both idea generation and idea destruction. AI efforts thus far have not achieved GPT-level quality for the latter.
So even though language embraces all of science and technology, the system was trained for language understanding and generation, not the actual underlying concepts. As with countless verticals, generative AI can serve as an ideation tool. As Dave Vennergrund, Vice President of AI and Data Insights at General Dynamics pointed out:
"No matter the tool, AI will always be an accelerator and a quality enhancer, but it will never be able to give us answers to every imaginable problem. There will always be a market for human ingenuity, and when it is combined with technology like AI and others, we really start to delve into the art of the possible."
Combine with the context-jumping ability of human brains. We can synthesize ideas from different contexts beyond the current capabilities of any machine learning model. However, even human brains initially struggle with blending two ideas from different contexts. Generative AI can get the gears going – especially when it comes to commercializing the results of R&D.
Generative AI can commercialize R&D faster
With the right dataset, generative AI can kickstart commercialization strategy. In my discussion with Jim Hornthal, CEO of LaunchPad Central – a company whose generative AI platform GLIDR now applies generative AI to its large dataset of customer discovery interviews – we applied it to a specific example. GLIDR holds a diverse dataset from many entrepreneurial teams, including many cohorts that went through the National Science Foundation's (NSF) I-Corps program, where scientists and engineers with a deep IP portfolio become entrepreneurs.
As an experiment, we first asked ChatGPT to generate a short description of Citrine Informatics, an AI-based platform for materials science product development. It came up with this:
"Citrine Informatics is a technology company that specializes in the development of artificial intelligence (AI) and machine learning (ML) solutions for materials and chemicals research, development, and manufacturing. The company's platform combines data-driven analytics with advanced scientific simulations to accelerate the discovery, design, and optimization of new materials and chemical processes."
Plugging this description into GLIDR's new generative AI tool, it produced an entire pre-filled Business Model Canvas – an industry-standard tool for commercialization. Potential customer segments, cost structure, revenue streams, market size, and more.
In normal cases, R&D-intensive efforts start commercialization from scratch. Potential customers especially take a lot of iteration and conversations to determine.
But now? With existing datasets on past commercialization efforts, generative AI can create viable templates for entrepreneurs and all other innovators. Greatly speeds up customer discovery and path to market fit.
While remarkable, the result actually makes intuitive sense. Because a conversation between two people is, to a significant extent, an interface between two language models running in two brains. Hence, it's natural to expect a large language model trained on masses of such conversations to hold significant predictive power over the tangible outcomes of conversations.
The frontier of AI applied to deep tech
Even though language models and generative AI can't perform complete R&D cycles yet, substantial economic value will flow from the adoption of AI in deep tech. Seeded with corporate, government and venture capital. As an early harbinger, Alpha Intelligence Capital already deployed $271M for deep tech AI.
For example, as with Citrine, AI-based platforms have sprung up to shorten design cycles for drug discovery, optimizing crop health for farmers and more. Generative AI built from large language models can be used alongside or even plug into such platforms -- and just as in many other applications, act as a cognitive copilot to generate diversity, find relevant context from within the platform, pump the intuition and get the intellectual gears moving.
Language is powerful. And generative AI deployed at scale on language now looms over us with its power. But generation is only one important part of several. To evolve novel solutions, it must partner with skepticism.