(Re)Share #57 - Starlink-crossed lovers
Embodied AI | NanoSat wars | Agent payments | AI Physicians | BCI | Bioengineering
Greetings after a multi-week hiatus—not by design, but by necessity. Since we last spoke, I’ve had a company retreat in SF, a funeral in Chicago, and a fantastically full week in London. I was particularly lucky to be invited to the inaugural ARIA Summit, a wonderful manifestation of unbridled technical ambition in the UK. This is something long overdue and genuinely refreshing to see coming from a government body. Those four weeks may have been sleepy for (Re)Share, but we’re back with a super-sized issue, so let’s get to it.
Stuff Worth Sharing
Know when to hold ‘em, know when to fold ‘em - In the world of robotics AI, the big man on campus is Physical Intelligence, and late last month they dropped π 0.5, their new vision-language-action (VLA) model that's purpose-built to generalize to unfamiliar environments. The model's capability was demonstrated through a range of everyday domestic tasks at impressive levels of mastery, but the real brag comes from the home environment being completely novel. Unlike models trained only on structured, lab-based tasks, π 0.5 was trained on a wide mix of visual, textual, and action data, allowing it to grasp both what a task is and how to carry it out in varied contexts. This is made possible because of Pi's work on high level / low level fusion within the same model, which was first introduced earlier this year. I find it surprising how many startups I see that believe executive function / planning alone is a sufficient value proposition. It’s not, grow up. In benchmarks π 0.5 achieved up to 94% success on out-of-distribution tasks, although you'll see in the embedded recordings that speed remains a limiting factor. More interesting is the demonstrated performance when datasets were excluded from experiment runs. Limiting the presence of web annotation cross embodiment data showed a 25%+ drop in success rate. Scaling laws are the infallible truth of AI, even in the atoms realm. Full paper here.
Hey nude, don’t make it bad - Earlier this month President Trump signed the TAKE IT DOWN Act into law, marking a significant federal effort to combat the spread of non-consensual intimate images and deepfakes. The legislation criminalizes the knowing publication or threat to publish such images without consent. It also mandates that online platforms remove reported content within 48 hours and prevent its reappearance. While this is a purely regulatory move, the clarity of the political signal and retaliatory measure should create significant behavior change with model developers and compliance systems. I’m betting that this also creates the path for new technical teams to build novel solutions and I would love to speak if that sounds like you.
Starlink-crossed lovers - Last month, Amazon’s Project Kuiper launched its first 27 operational satellites into low Earth orbit, marking the official start of its ambition to deliver global broadband internet. It’s also the first—and so far only—viable competitor to Starlink’s dominance. Project Kuiper plans to deploy over 3,200 satellites to form its constellation, roughly half the size of SpaceX’s network. As a frequent United flyer, I’m a big fan of satellite internet and have no shade to throw at Starlink. But I’m an even bigger fan of corporate rivalry, and Musk vs. Bezos is an HBO series just waiting to happen.
All work and no pay - As longtime readers know, I’m particularly interested in the payment rails and control systems needed to enable an agentic future. That investment thesis just got harder. Mastercard has unveiled Agent Pay, an agentic payments technology aimed at revolutionizing commerce in the age of AI. The goal is to integrate AI-driven agents directly into the payment ecosystem, enabling more intelligent, secure, and personalized transactions across everything Mastercard touches. The platform is built around Agentic Tokens, which enable programmable payments—akin to Web3 smart contracts, though Mastercard is careful to avoid that comparison. This is a big deal for both consumer and enterprise commerce. In theory, any standard transaction flow could be automated with Mastercard infrastructure underpinning it—at least for low-friction payment levels. I still believe there’s space for an agentic-native solution tailored to micropayment transactions, which would be crushed under the weight of traditional rails like ACH and SWIFT. But peace of mind is a hard thing to disrupt.
I’m fallout of love - I’ve been terrible about reading physical books this year, so I’m giving audiobooks an honest shot—and to kick things off, I picked Nuclear War by Annie Jacobson. This was a terrible idea. Not because the book wasn’t good—in fact, it was excellent—but because it scared the shit out of me and left me with a deep sense of powerlessness. The book presents a hypothetical “scenario” in which North Korea launches an unprovoked nuclear attack on the U.S., triggering a rapid and catastrophic sequence of military responses. Spoiler alert: the world ends in just over an hour. The real terror comes from Jacobson’s meticulous research into detection systems, deterrence protocols, military communication infrastructure, and the Cold War–era game theory that still underpins modern nuclear strategy. Fuck Catcher in the Rye—this should be mandatory reading in schools.
It’s time to Facebook facts - I’ve been a little meh on the past few months of Dwarkesh’s interviews but this conversation with Zuckerberg really brought things back. The Meta CEO, sporting stupid AF RayBan collab glasses, covers the new business models of AI, embedding cultural ethics in a company, the unlock of distillation in models and, of course, the changing political winds in Silicon Valley. The interview is 75 minutes long but if you’re tight on time I would highlight the discussion around DeepSeek & China (00:35:10) and Monetizing AGI (00:54:15) - two topics that Meta / Llama have a uniquely refined perspective on. I’ve been thinking a lot about the new systems required for advertising in an LLM / agent native world and I found Zuckerberg’s recollection of early FB ads and the switch to mobile to be quite instructive for how I should frame this new paradigm shift.
Socialized medicine - Google released some new research on advanced diagnostic processing through conversational, patient-facing interfaces. While LLMs have proven to be remarkably capable in loads of medical applications, the nature of their probabilistic vector mapping creates a certain deviation in how models run diagnosis, at least for less definitive symptomatic conditions. Reasoning, even with some of the more recent MoE deployments, is still sub-par. The new model, Articulate Medical Intelligence Explorer or “AMIE,” leverages the inference-time chain-of-reasoning method of creating “patient states” as model outputs in between nodes or within temporal memory. This allows AMIE to engage with a patient in a conversational manner and update the patient state as new information is gathered. AMIE was developed through a self-play reinforcement method but was ultimately tested across 105 evaluation scenarios with 25 patient actors. Evaluation of diagnosis by 18 specialists revealed superior performance of AMIE relative to GPs in the majority of cases—with and without multimodal data included.
Coming to grips with it - Amazon introduced Vulcan, a warehouse robot equipped with tactile sensitivity. Vulcan integrates tactile sensors and AI to enable human-like dexterity and can handle ~75% of the items in Amazon’s fulfillment centers. It has already processed over 500,000 orders and can work up to 20-hour shifts in a seemingly endless demand environment. As I’ve discussed in my own robotics thesis, throughput is the key criterion—not human performance parity. The latter offers a short-term, false sense of security. Fly recently made an unannounced investment in a fine-dexterity, touch-sensitive robotic manipulation company, so I for one say: bring it on, Bezos.
Twinner take all - A highly entertaining talk by Jim Fan, Director of AI at NVIDIA. Fan describes the cutting edge of robotics data scaling through the blend of advanced simulation and digital twin transfer learning. NVIDIA seems to be all in on everything right now, but robotics is clearly at the forefront of their strategic thinking. Fan introduces his theory of “Digital Cousins,” a generative physics engine underpinned by diffusion models to create close-enough edge case data. Only 18 minutes and well worth it.
What’s on your mind? - Neuralink is no stranger to (Re)Share, and for good reason. The BCI leader just received Breakthrough Device Designation from the FDA for restoring communication in individuals with severe speech impairments. Say what you will about Elon, but this is a remarkably exciting development for anyone touched by neurodegenerative communication loss. The BCI device is designed to assist individuals affected by conditions such as ALS, stroke, spinal cord injury, cerebral palsy, multiple sclerosis, etc., by decoding neural signals on the brain surface itself (technically a few mm deep). For those interested, this is a great watch on Nueralink’s first patient one year later.
The conditions are dire - In the last two issues of the newsletter, I’ve included stories about the incredible work of Colossal Biosciences, and I’m becoming a bit of a fanboy. I promise this will be the last one for a while—but also, wolves. Time penned a deep dive on the biotech’s recently announced birth of three genetically engineered wolf pups—Romulus, Remus, and Khaleesi—described as the first “de-extincted” dire wolves in over 10,000 years. Scientists analyzed ancient DNA from dire wolf remains and used CRISPR to modify 14 genes in gray wolf cells to express 20 traits associated with dire wolves. I know it’s controversial, but I mean… come on. Dire wolves! My House Stark pride is dialed to eleven.
Manganese cum laude - GM unveiled a new lithium manganese-rich (LMR) battery technology that can increase EV energy density by 33% without raising production costs. LMR chemistry reduces reliance on expensive and scarce cobalt and nickel by using abundant manganese, which comprises up to 70% of the cells. Historically, LMR batteries have faced issues with shorter lifespans and voltage fade. However, GM claims that its new LMR tech matches the lifespan of current high-nickel batteries.
Portfolio Flex
Cakewalk announced their $7.5 million Seed round led by my good friends at Kindred.
Fly’s latest investment, Zendo, came out of stealth and set their sights on overhauling data center energy procurement.
Agemo shipped a fancy new website.
Transcelestial dropped a bad ass branding video.