When sketches jump almost straight to drive sales & manufactured output, Ai promises similar Science breakthroughs are here. Can they be ? Here's How.
Where current models of Ai systems ( and many based on text / language based processing: LLMs ) and where we have constantly seen great promise in a many design ( and sometimes ) scientific " discovery " like efforts, the idea of guided exploration ( which can sometimes be considered ) serendipity discovery or " ah ha " explorations ( and certainly in examples such as the robust development of Ai products such as Vizcom for the industrial design world, we see that the expectation to the same level of scientific jumps that could ( faster than light travel ) push us into a new world of scientific breakthroughs is where Ai systems are not the same for non qualitative outputs eg Physics, Biology, Materials, Chemistry domains.
With this in Large Language Models in Scientific Discovery an incredibly robust effort focusing on project-level performance, and where the default to easily measurable validity of output where Ai models can and must propose testable hypotheses, design simulations or experiments, and then via interpreting results, is where we see state of the art frontier LLMs being directly effected from pre-training data and objectives rather than from their distinctive architecture and implementation details but more specifically the ability to NOT to be able to make jumps that currently are in the " human " and " ah ha " vector of desirable output. This is better described as AGI / human thinking and where next predictable statistical Ai systems struggle with such an objective. Scientific discovery after all advances through iterative cycles of hypothesis proposal, testing, interpretation, and refinement and fully connected to accidentalism which some have defined as a serendipitous factor. We use the word synchronistic.
In creating environments where beyond per-question Ai accuracy synchronistic ah ha situations can occur frameworks that enable project-level assessment and specifically in context of the system devised to require models to formulate testable hypotheses, execute analyses or simulations, and interpret outcomes to approximate an end-to-end discovery workflow this is where a new form of Ai systems, Serendipity LLM driven optimizations Projects, and specifically characterized by abundant, well-structured open-source data and codified knowledge, such as protein design, transition metal complex (TMC) optimizations, organic molecule optimization, crystal design, and symbolic regression, just to name a few exhibit the most significant gains from LLM integration to form " ah ha machines " as we have defined them. Further examples of this are firms such as Ginkgo Bioworks and Atinary both having partnered with leading US based Ai Labs use robotic hardware to Achieved major cost and yield breakthroughs in biological design loops but where this is still not " jumps " that synchronistic Ai is chasing.
Where this takes thins a step further is in the creation of SDE ( Scientific Discovery and Experimentation ) frameworks pushing project-level evaluations through well-defined research scenarios, enabling direct analysis of error propagation from Q&A to downstream discovery and where this may sound similar to image based " investigation " / generative diffusion based Ai systems. It is here also where the capacity to discern optimization directions and facilitate serendipitous exploration appears more critical even though top-performing LLMs score highly on questions regarding retrosynthesis, reaction mechanisms, and forward reaction prediction ( statistically possible due to limited edge constraint based possible outcomes ) it is here where these systems also struggle to generate valid multi-step synthesis routes due to pre-training data creating less decisive in discovery projects than in static question-level evaluation.
Where is the design world where there are almost no boundaries other than manufacturability thus effectively orchestrate the loop of " design " scientific discovery, future Ai developments that prioritize balanced knowledge and learning capabilities across diverse scenarios over narrow specialization is desired and the goal of the widest possible way to see the unseen and where such divergence capabilities underscores that proficiency in standard scientific examinations does not guarantee mastery of the nuanced, context-dependent reasoning required for scientific discovery, yet certainly within in say, the capability of furniture and object design.
Project-level evaluation indicates that question-level
patterns only partially predict discovery performance and that a model’s capacity to drive a research project relies on factors more complex than a simple linear correlation with its Q&A accuracy. This implies that precise knowledge of structure-property relationships may be less critical than the ability to navigate a hypothesis space effectively
In the raw scientific domains and where training and evaluation paradigms must expand beyond textual accuracy ( LLM systems ) to prioritize executable actions and where invoking tools, debugging execution failures, and iteratively
refine protocols in response to noisy feedback this is well within the reach of next generation, domain specific systems. A fully integrated with all of humanity's knowledge systems ( scientific of other ) is still not within reach and where in using scientific discovery loops and LLMs as proposals over a hypothesis space (e.g., the space of all possible molecular structures, symbolic equations) not only can LLMs generate offspring based on parent hypotheses sampled from the pool ( eg Discovering the best spin configurations that minimize the Ising model energy presents significant challenges due to vast combinatorial configuration spaces ) these are isolated and domain specific applications. Applying this to say a way to increase wood bendability for unconventional furniture design and within the context of " what if " design sessions all the way down to the financial implications of the demand necessary to build the machines necessary for specific new designs to be realized are, like scientific systems, still some way off.
#iGNITIATE #Design #DesignThinking #DesignInnovation #IndustrialDesign #iGNITEconvergence #iGNITEprogram #DesignLeadership
