(Re)Share #59 - Tatooineage dirtbag
Nuclear proliferation | Autonomous Science | Self-driving cars | Robotic dexterity
Greetings from 35,000 feet - somewhere over Canada, I think. I’m returning from a jam-packed but extremely productive two weeks in Europe. The trip included the standards: board meetings, founder meetups, and some face-to-face time with the Fly team. But I was also able to celebrate my birthday with Steph in my favorite city, London. The celebration was slightly muted by the realization that this is the final year of my thirties and the specter of middle age rapidly approaches. But I digress. Loads to cover this issue, with a particularly big month for the portfolio—so let’s get to it.
Stuff Worth Sharing
Mime the gap - If you’re like me, at some point you’ve questioned why mimes are a thing. Even by historical standards of entertainment, they seem dumb and pointless. After reading this paper, I still feel that way—but marginally less so. A team out of MIT and Imperial set out to overcome the inherent dependency on verbal and language cues that most models have for interpreting human behavior. It’s said that 85%+ of our communication is non-verbal, but because LLMs are primarily trained on language-centric data, a strong dependency persists. The MIMEQA paper introduces a new benchmark for evaluating nonverbal social intelligence in AI systems. The authors collected 8 hours of mime videos from YouTube, curated 101 clips, and created 806 annotated question-answer pairs testing three levels of reasoning: grounding imagined objects, scene-level social interactions, and global social understanding. When tasked with interpreting subtle, context-dependent cues, even SoTA models scored below 32%, while humans clocked in at 86% or higher. Fine-tuning helped a bit with reasoning, but grounding imagined objects remains a major hurdle. The paper hopes to inspire future research in this area but I hope it doesn’t, because, and I hope this is now clear, I really dislike mimes.
I’m fallout of love - Earlier this month, the World Bank formally lifted its decade-old ban on financing nuclear energy projects in developing countries. This policy shift is a massive deal for capital markets and reflects the rapidly changing stance on nuclear as a viable option in the global energy mix. Nuclear fusion and modular reactors are semi-frequent topics on (Re)Share, but that’s the tech gossip and VC hype corner of the market. This is the World Bank. Under the new policy, support will now extend to prolonging the life of existing reactors, upgrading power grids, and accelerating the development of small modular reactors (SMRs). Some estimates project that electricity demand in developing regions will more than double by 2035, and without nuclear, that demand will almost certainly be met by fossil fuels. One of the most notable developments is the shift in stance from Germany, which has long been staunchly anti-nuclear. As a Partner at a Germany-based fund, I’m particularly interested to see whether this trickles down into LP guidance and capital allocation norms.
Tatooineage dirtbag - Researchers have discovered that even the arid, bone-dry air of Death Valley holds enough moisture to extract potable water using atmospheric harvesting technology. A small panel, cooled just enough, was able to draw in water vapor and produce about one glass of water over the course of a day. Fog harvesting with mesh nets has existed for decades—something I was completely unaware of—but it’s struggled to scale. This prototype, a tabletop-style sorption panel, can capture moisture even in ultra-dry air by cooling surfaces below the dew point. Currently, it can extract around 5.8 liters/day in moderately dry conditions, but the findings suggest that even the world’s driest regions sit atop vast reservoirs of airborne moisture. With further development, devices like this could provide decentralized drinking-water sources where none existed before.
Bi Bi Bi - I’ve been actively exploring Autonomous Science as a working investment hypothesis. One of my recent favorites comes from a Stanford team that released Biomni, a general-purpose biomedical AI agent that combines the reasoning power of LLMs with RAG and code execution to autonomously perform complex research tasks across 25 biomedical subfields. Biomni is dynamically composable and can plan and execute diverse tasks like gene prioritization, drug repurposing, rare disease diagnosis, and microbiome analysis without task-specific tuning. The system is composed of two parts: Biomni-E1, a literature environment built from 150 tools, 59 databases, and 10,000+ scientific papers; and Biomni-A1, a reasoning agent that identifies relevant tools and generates executable code plans to conduct experiments. In short, it behaves like a scientist would. Upon receiving a user prompt, the agent identifies the most relevant resources within E1 and generates a logical flow expressed as executable code. On LAB-Bench benchmarks, it matched human experts on database queries and exceeded them on DNA/protein sequence reasoning.
The stables have turned - The Senate passed the GENIUS Act, the first major U.S. bill regulating stablecoins. The bulk of the bill is pretty dry financial regulation - maintain full reserves in liquid assets, third-party audits, and monthly public disclosures - but that’s exactly the kind of thing that unlocks serious capital. USD-backed stablecoins have grown ~65% YTD, and some estimates predict an 8x increase to over $2 trillion in the next few years. Circle’s glorious post-IPO performance has been the belle of the ball, but financial giants, both young and old, are rapidly jumping into the space to stake their claim. Not such a bad time to have a rapidly scaling seed-stage company in one’s portfolio.
Embody building - I’d normally put this under Portfolio Flex, but Wayve CEO Alex Kendall’s presentation at NVIDIA’s GTC is truly a must-watch. The talk lays out Wayve’s vision for transforming autonomous driving through end-to-end embodied AI and dives into the series of decisions they’ve made to unlock it. Unlike traditional AV systems that rely heavily on HD maps or rigid perception-and-planning stacks, Wayve is building an adaptable, sensor-agnostic AI driver designed to generalize across vehicles and urban environments. Because Wayve utilizes a modular AI stack and a “safety-by-design” philosophy, they’re able to dynamically interrogate the model in real time. Kendall spends most of the presentation walking through how they’ve engineered transferability across geographies, driving behaviors, and regulatory frameworks, all built into the system’s core. That capability expansion is underpinned by Wayve’s PRISM-1 and GAIA-1/2 generative simulation models, which synthetically generate safety-critical and out-of-distribution driving scenarios. With Wayve’s recently announced partnership with Uber (see below), it won’t be long before more people get to experience what we at Fly have had the pleasure of watching develop over the past several years.
Material evidence - A-Lab is a self-driving materials lab that autonomously discovers and synthesizes new inorganic compounds. The paper is a few years old, but it remains a compelling example of verticalized scientific discovery designed to overcome the sim-to-real gap. Over a 17-day run, A-Lab computationally selected and evaluated targets, ultimately producing 41 novel compounds. The platform used active learning grounded in thermodynamics to refine synthesis steps and intelligently interpret failures—essentially a dynamic, data-driven approach to trial and error. There’s a lot of excitement around AI’s potential to accelerate or even reshape scientific discovery, but so many companies I see stop at the discovery phase. As the techbio boom-and-bust has clearly demonstrated, that’s rarely enough for durable value capture. Especially once the venture hype cycle cools.
Dexterity’s Laboratory - Robotics startup Generalist unveiled a series of fine-dexterity capabilities through a bi-manual gripper rig. I’ve heard from my own portfolio that the team behind Generalist is legit, and there were high expectations to see what all that brainpower could produce. From my perspective, they did not disappoint. What really stands out is the specificity of motor control on display. Singulation, grabbing a single item from a group, is a notoriously tough task, but here we see screws, zip ties, and Legos being manipulated with a smoothness and natural precision that’s pretty remarkable. Capabilities like these unlock a world of potential applications that would’ve been impossible just a year ago.
Quandumb ass - An interesting read that explores the growing claim that advancements in artificial intelligence are encroaching on domains once considered the exclusive terrain of quantum computing. Physics, chemistry, and materials science have all seen rapid momentum from AI—and unlike quantum computing, they don’t suffer from issues like hardware fragility or scale limitations. Optimized algorithms and tailored datasets have demonstrated a remarkable ability to simulate quantum-like behaviors, particularly in drug discovery and materials design. Quantum maxis still argue that high-precision tasks will remain out of reach for classical AI, but I’m seriously doubtful of that.
Too molecule for school - Speaking of mastering quantum effects, a second Fly portfolio company, Orbital, presented at NVIDIA’s GTC this year to showcase just how far they’ve pushed the envelope on novel material discovery and synthesis. This ~30-minute talk dives into much of what the Orbital team has been hard at work on over the past few years. The crux of the presentation focuses on the massive scale advantage computational methods have over traditional discovery approaches—and the technical choices they’ve made to make it all run on (surprise, surprise) NVIDIA GPUs. Just a few years ago, state-of-the-art material simulation could handle a few hundred atoms. Orbital has already built a system capable of half a million atom prediction. They’ve used this capability to develop proprietary carbon capture and cooling absorbents—primarily for data centers—but the same system architecture could easily expand into advanced batteries, aerospace materials, or even medical applications. For the more technically inclined, you may also appreciate Jonny’s commentary on how Orbital’s compute structure diverges from LLMs and how they’ve leveraged graph neural nets to handle the sparse computational demands of modeling molecular bonds.
Portfolio Flex
Wayve announced a massive partnership with Uber to bring autonomous driving to its 150 million users worldwide.
Metaview announced a $35 million Series B led by GV.
Remberg announced a €15 million Series A led by Acton Capital and Oxx.
💯 for the N Sync reference.