(Re)Share | #34 - Cosmic dust bunny
Space debris | Embodied intelligence | Gen AI games | Humanoid robotics
It’s been two weeks since my last issue and I know you’re all broken up about it. I partially blame a sprint for a new investment (!!!) and mostly blame a visit from my mother (she doesn’t read this). Fortunately it’s been a very interesting fortnight so lot’s to cover…
Stuff Worth Sharing
Cosmic dust bunny - Not the star studded awards ceremony. Of course I’m referring to everyone’s least favorite Sesame Street can gremlin. In the last issue I shared the latest mission from Astroscale with the long term goal of cleaning up space debris. For those that are less familiar with the increasing low Earth hoard, this overview explainer from Vox is a good entry point. Currently there’s an estimated 1 million uncontrolled orbiting fragments and every additional component of the debris cloud increases the likelihood of a chain reaction of collisions, known as Kessler syndrome. The video mainly focuses on the leading solution of laser-based sharpshooting. That may sound equal parts awesome and impractical but since this was sent to me by space laser founder, Hira Virdee, I think it checks out.
Let’s get physical - We’ve now entered the commentary phase of OpenAI’s incredible video generation model, Sora. One of my favorite overviews was this a16z podcast with Stefano Ermon, Professor of Computer Science at Stanford and one of the leading voices in diffusion models. The conversation covers a lot including the history of diffusion models, the relative merits and tradeoffs with other ml approaches and his view on the surprising power of scaling laws. The most thought provoking was the discussion surrounding the emergent capabilities of physics understanding. If you look at the ~60 second demonstration videos of Sora you’ll notice advanced cognitive controls like object permanence and temporal consistency, which were a) not programmed and b) not present in the shorter form Gen AI videos we’ve seen in past months.
Your attention please - Demis Hassabis was interviewed by Dwarkesh Patel and it’s a very good listen. The conversation is more wide ranging than the previous podcast but as leader of arguably the most important AI organization in the last decade his depth of thought does not disappoint. Here too the topic of emergent physics awareness is discussed as is Google’s general perspective on safety, alignment and security. Plus some interesting predictions on AGI timelines, robotics impacts and the usual mix of armchair strategy.
Back in the bottle - It sometimes makes me mad to think I was born when I was. “Competitive Gamer” is now a career path that is a) more fun and b) probably more lucrative than most of VC. But now kids can actually make their own games with nothing more than a drawing. Deepmind introduced Genie, a foundation world model trained from Internet videos that can generate an endless variety of playable worlds from synthetic images, photographs and even sketches. This is remarkable for a number of reasons. Beyond the sheer computational significance and magical effect the research provides, it’s a compelling step forward in simulation-generated world models, which many believe to the be only viable path for real world, higher risk autonomy (robotics, self-driving, etc.). One of the lead researchers is a friend of Fly and I’m lucky enough to be interviewing him this Friday so early readers feel free to shoot me some questions.
Get off my property rights! - In 1980 Congress passed the Bayh-Dole Act, a bipartisan law that enabled universities to fully own IP that they developed even if supported / funded by federal cash. Before that patents and other innovations that were government backed were licensed out on a non-exclusive basis, which as you can imagine did not go well. Since Bayh-Dole the university research system has created untold value across pharmaceuticals, medical devices, telecommunication innovation and more. The current administration is evaluating a modification to the law that would in effect grant the federal government manifest destiny rights to IP exclusivity. This all comes via the understandable mission to curb soaring drug prices, but this change could have massively negative effects on innovation, particularly with tech transfer offices (TTOs).
Life by Committee - Mike Dempsey wrote yet another excellent thought piece, this time an overarching perspective on multi-agent debate and collective intelligence. For the uninitiated, the basic premise is that creative and cognitive tasks (strategic decision making, resource allocation, etc.) don’t have an objective right / wrong which inhibit classical methods like self-play reinforcement learning. Instead the solution to a global maxima may be through a marketplace of ideas debated by agents themselves. The piece covers a brief history of architectural innovations (RL, GANs, MOEs) and a wide range of applications from social norming to stock trading to war games.
Astronomical transfer fees - SpaceX is scheduled to make their third and most ambitious attempt at a Starship launch. The hope is to push a number of boundaries for the vehicle including a re-ignition of a Raptor engine while in space and a propellant transfer demonstration during the coast phase. These capabilities are world’s firsts and are widely seen as critical to unlocking the lunar economy. Save for a last minute launch scrub, by the time you’re reading this the results should be known so allow me to share a preliminary congratulations / condoloences.
That’s crossing the line - Another Compound VC friend, Shelby Newsad, penned a short piece on the latest traction benchmarks for bio platforms. She argues that there’s been a massive recalibration in the post-ZIRP era and this is uniquely pronounced in bio due to the extreme asymmetry of outcomes. Bio follow on funding seems to be overwhelmingly migrating to an asset-first mindset, which reflects my own portfolio experience. If you’re a biotech / techbio founder this is a must read.
Comparing apples and lemons - A not so well kept secret project of Apple was their decade long attempt at a self-driving car. Earlier this year the company finally sent Project Titan to the junkyard (😏) after ~1,000 employees’ efforts and billions spent. The article is a fun, pseudo oral history of the project dating back to 2015. Can’t fault the company for the ambition but they probably should just buy Wayve…
Domo arigato - Robotics startup, Figure, announced a mind-bending $675 million round from the who’s who of tech companies. For a two year old company this is insanely impressive so out of the gates kudos to the team. Am I incredibly skeptical of humanoid robotics at large? Yes. Do I think this is a complete tech bro / science fiction fantasy hype bubble? Yes. Do I think this Figure demo video shares a shocking resemblance to the mobility of my extremely frail 74-year old father? Yes.
Bring in the early birds - A group of researchers engineered a novel malware method for the generative AI era. The Morris II worm is an adversarial self-replicating prompt that in demonstration showed that it could effectively trick email agents to f**k up their own system and future recipients. Helpful explainer video here.
Portfolio Flex
Do you hear me now? Good! - Transcelestial announced a supplier agreement with Wesco International to bring their laser-based communication resiliency to cell towers. They’re now working with three of the largest telcos in North America!
I got the part! - CloudNC was featured in the keynote of Autodesk University, an audience of 10,000, for there work on AI-drive toolpath design (see 19:15)