Automated Research
Abstract
Leveraging computational intelligence to hyper-accelerate scientific discovery
Description
No field has more raw transformation potential through AI than the hard sciences. The scale of complexity within protein development, chemical synthesis, and novel modality creation was previously thought to be almost impossibly intricate but will soon be fully captured within a single model. There’s a certain romance in the idea of the toiling researcher, painstakingly seeking truth through relentless trial and error, but it’s increasingly plausible that such manual efforts will soon be considered so wildly analog that they’re almost artisan. The sheer size of the scientific literature available today and the exponential growth rate of publications has created a corpus that is objectively impossible for any human mind to process in a lifetime, but extremely tractable for LLM
This is not a novel idea. Years of breakthroughs Alphafold (1-3), GNoME, PaperQA, BioNeMo, and the countless number of lesser known startups have continued to push the frontier across a spectrum of fields.
What is novel today, however, is AI’s ability to reason and analyze its own work. The ability to generate hypotheses and auto-test for viability, at least at a preliminary level, presents a true inflection point that answers why now is different. The combination of massively expanding computational infrastructure and a series of reasoning-based algorithmic advancements (test-time training, mixture of experts, self-play, chain-of-thought) makes the why now (IMO) quite clear. I hold no specific view on superior architecture, and frankly, I would question any rigid dogma here given the pace of change.
By freeing ourselves from our biological limits and cognitive clock speeds, there’s no telling what comes next. I’m personally quite interested in industrial applications (aero/space composites, agricultural coatings, synthetic Earth metals) and therapeutic enablers (peptides, LNPs / AAVs, etc.).
Working Thoughts
To critique my own work, this risks being a VC narrative remix of the previously hot field of AI drug discovery. There was (still is) a generation of startup founders who pitched a similarly sounding premise of expanding the search space. The modality, target, or disease may change, but everything I’ve written above could have plausibly been said by any number of those companies. Where so many failed is in navigating the value creation/capture tradeoff. There is a graveyard full of founders and techbio investors who learned the hard way that pharmas don’t give a shit about a novel asset without efficacy.
There is something intuitively attractive about investing against a narrative that is inherently immune to the embedded fears of joblessness and human replacement. In so many verticals, the core value proposition is productivity and efficiency (i.e. cost reduction). With the scientific domains, however, the unlock is knowledge generation and [medicine / material / chemical] acceleration. There’s also the reality that you can’t easily train up new human experts, so throughput is really what is on offer.
Related Reading
Relevant Companies
Orbital (portfolio)