(Re)Share #55 - Abracadaver!
Synthetic data | Robotics | Molecular design | LLM data training | Biotech scaling
It comes with a heavy heart that I must inform you, dear reader, that this will be the last free issue of (Re)Share. I’ve gone on record to promise that I would never put up a paywall for this newsletter but that was before my retirement accounts cratered due to a completely asinine shell game tariff policy. $5 trillion in lost value but who can put a price on the return of sneaker production?
Shameless Plug
I've started a research project to map out the current state of university spinout agreements and TTO norms across top universities in the US and EU. There have been some fantastic community efforts on this topic in recent years - Spinout.fyi, Gov.UK review - but I'm honestly shocked by the norms and behaviors that I see through my day-to-day investing work.
I’m not sure where this will end up. At a minimum I’ll be sharing my findings but based on the already strong response I might launch a more enduring survey / resource for aspiring founders. I’d love to speak with founders that are comfortable sharing their experience or TTOs leadership to understand their priorities. Please reach out if anyone comes to mind.
Stuff Worth Sharing
Ride or die - Fly portfolio company Wayve has introduced the second iteration of their generative AI world model, GAIA‑2. We discussed the predecessor model back in (Re)Share #20, but this new version brings significant improvements in control, realism, and adaptability. GAIA‑2 is engineered to enhance autonomous driving systems by simulating diverse and complex driving scenarios. While the sheer fidelity of the generated content is impressive on its own, what’s truly mind‑bending is the enhanced control over dimensional variability and the ability to generate edge cases on demand. The post does a great job of highlighting the model’s capabilities in dynamically controlling key variables like time of day, weather conditions, and road signage. Beyond environmental factors, GAIA‑2 can simulate core driving actions—both common maneuvers like braking and U‑turns, as well as rare situations such as encountering an oncoming vehicle in the wrong lane. I’ve only experienced Wayve vehicles in the fairly mundane London weather, but it’s reassuring to know that even if the city were overrun by organized wild dogs, my AI chauffeur is ready for any scenario.
Rule with an iron fist - Researchers have developed a biomimetic prosthetic hand that integrates neuromorphic tactile sensing to enhance grasping precision and compliance. This prosthetic emulates the human hand’s hybrid structure, combining soft and rigid components to provide both strength and flexibility. The incorporation of neuromorphic tactile sensors allows the prosthetic to perceive touch—and even pain—facilitating more natural and responsive interactions with objects. This advancement holds immediate promise for improving the functionality of prosthetic devices, especially as we rapidly approach the need for high-functionality automation amid escalating geopolitical challenges. This topic may or may not be related to a still‑in‑stealth investment we’re making in the coming days.
It’s what’s inside that counts - Anthropic has once again pushed the envelope on understanding the internal mechanisms of large language models. The challenge of black‑box decision opacity is widely cited in our field, and this marks the second major release Anthropic has shared—see (Re)Share #38. Their breakthrough comes from the emerging practice of circuit tracing, a technique that maps out clusters of neurons that consistently activate together in response to certain concepts and traces the activation pathways through the model’s parameters. One intriguing example discussed in the release comes from their work on Claude 3.5 Haiku, where the model solved basic math problems in some very odd ways. Although my friends with kids have tried to explain new math to me and that seems equally stupid so who am I to judge?
Make proteins great again - NYC‑based VantAI has unveiled Neo‑1, a foundation model that integrates atomic‑level structure prediction with molecular design. This is a seriously remarkable advancement—one that took me three full reads to truly appreciate. The model enables fine‑grained structural prompting of individual proteins, molecules, and their interactions. Neo‑1 functions as a unified system of simulation, structure prediction, and design that facilitates truly programmable proteins. If this proves effective, it represents the holy grail of biological research and could transform the field even more drastically than the revolution we’ve seen in software engineering. The piece provides detailed examples of how the tightly coupled steering engine can optimize a variety of protein binding targets and molecular glue strategies. Fortuitously, VantAI highlighted a CDK12 sequence—a target I’ve come to know well through Pear Bio. In short, we could be on the precipice of a transformational era in biology.
Maid a killing - Norwegian robotics startup 1X will begin testing its humanoid robot, Neo Gamma, in a few hundred homes by late 2025. So my condolences in advance to the few hundred casualties who will fall victim to their metallic hand. Designed for basic household tasks like vacuuming and plant watering, Neo Gamma isn’t fully autonomous yet—it relies on remote human teleoperators for oversight. Data from these home trials will help train 1X’s in-house AI, refining the robot’s behavior over time.
Abracadaver! - With advancements in bioengineering rapidly evolving, a small but growing movement is seriously exploring the concept of “bodyoids”—lab‑grown human bodies that are engineered to be devoid of sentience. These entities would be cultivated from pluripotent stem cells within artificial wombs, with precise genetic modifications ensuring the absence of consciousness and pain perception. The potential applications are vast. They could provide a reliable source of organs for transplantation, reduce our reliance on animal testing, and serve as near‑perfect platforms for drug development. It’s also extremely ethically fraught and seems like the first step of 8 steps that ends in a Matrix-like dystopia. Not particularly technical but still it is a fascinating read and excellent dinner party fodder.
Rock bottom line - A long but solid read delves into the evolution of drug discovery technologies and the resulting commoditization within the biotech industry. Early biologics pioneers like Amgen and Regeneron once dominated the field, but as competition intensified, their approaches became commonplace—a classic case of “your margin is my opportunity” across industries. Author Elliot Hershberg argues that the same trend is now taking hold with in silico modeling. As simulation technologies grow more robust, we should expect a significant collapse in drug development costs, ultimately realizing the dream of “techbio”: low‑cost, high‑convexity asset investing. However, the theory may be undermined by a historical reliance on Chinese CROs—a hallmark of the era of globalization—that now appears to be coming to an abrupt end. With the US and much of the West becoming increasingly unapologetic about nationalist policies, this shift will almost certainly stunt the commoditization trend in the long term.
Serves you copyright - Multiple lawsuits against OpenAI are now being consolidated into a single case in the Southern District of New York. While these aren’t the only legal actions underway, they could mark the first formal ruling on the hotly debated practice of training large language models on copyrighted works and the extension of the “fair use” doctrine. As a card‑carrying techno‑optimist and a staunch believer in property rights, I find the fair use argument to be absolute nonsense. I also spent some time in NYC last week and heard some rumors that a proposed restitution could include paying each author in the training corpus $50K per work—per country or language version. Considering that my wife’s work was included in that corpus, I might be a bit biased. Perhaps it’s not the best time to announce an insane round.
I’ve vaccine better days - A recent study in Nature examined the potential impact of the herpes zoster (shingles) vaccine on dementia incidence. Conducted in Wales, the researchers found that receiving the shingles vaccine was associated with a 20% reduction in the risk of developing dementia over a seven‑year follow‑up period—with the effect being more pronounced in women than in men. This finding adds to the growing body of evidence linking viral infections, particularly those caused by herpes viruses, to an increased risk of dementia through mechanisms like neural inflammation, a known contributor to cognitive decline. I have a history of dementia in my family and am a man and last year somehow got shingles, so FML is what I’m trying to say.