SVRGN Weekly Digest #45 💫
HNY and so on, major reads for your stockings
Prelude
This is somewhat belated, but we have very good reasons for it. We’ve decided to forego any holiday calm and instead accelerate into the new year by doing some amazing f*cking investments we couldn’t be more excited about.
After the first year of SVRGN Weekly Digests (the first one was Jan 26 2024), my conclusion is that I enjoy writing these short form reflections, and some of you apparently enjoy reading them. It can be a pain to do, hence why they are sometimes published Sunday at midnight, but honestly I always look forward to looking at all the cool people, companies and ideas we explored this week.
If you’d like us to continue publishing, please share it with your cave-man friends who need some culture and refinement in their lives, this may positively reinforce our behavioral loops of external gratification as we see the subscriber numbers go up.
During the holidays we’re hoping you receive something very special: time to read. There’s so much interesting reading to do, and too little time, so choose carefully. What can be better than putting on your noise-canceling headphones, to tune out family, dogs, Kalle Anka (if you’re in Sweden), and getting deep into some details about NVIDIAs Blackwell architecture, Popper’s 3 worlds, or microbial TNT detectors?
All I want for Christmas is enlightenment.
🤼 People
Christian Reber - Founder at Interface capital
Christian founded and exited Wunderlist, Superlist, Pitch.com and now Interface capital. With Interface, he’s doing some very interesting early stage investments. We also share bullishness on Europe, and bearishness on Berlin.
💼 Portfolio jobs board
This weeks selection of opportunities from the portfolio:
🚀 Companies
Defense Media company - Covering New Defense in Europe
Status: Operating
Source: EDTH
Founders: Experienced journalist and ex-military
Why it’s cool:
We need knowledgable coverage of defense tech news in Europe. Not necessarily a traditional VC case, but certainly important and a white space that should be filled. If you’re interested in learning more, dm me!
💡 Ideas & Science
Seminal econ paper from the 1970’s. It’s one of those ideas/results that are very intuitive when you heard it. First heard by befriended VC Elad of Lunar when discussing a market structures in venture.
We are forever chasing ideas and mental models that help us widen our thinking. Popper describes a metaphysical framework for an evolutionary emergence of worlds. World 1 is the physical world. World 2 is the consciousness, experience and thoughts of (presumably mostly) humans. World 3 contains knowledge and ideas such as scientific theory, art and thought objects.
Neural operators are a key concept in how AI can model reality. It’s a deep learning architecture for learning mappings between infinite function spaces. They seem particularly well-suited to simulating and solving partial differential equations. One trained model can be discretized differently, allowing for flexible resolutions.
In short, probably there are headwinds to scaling by “just” applying more compute. Big surprise. To state that model scaling is over is likely another in the long line of “erroneous predictions about AI”. Inference scaling (à la o1 and DeepSeek R1) might provide some respite from lack of quality data, but probably won’t solve every problem. Finally, current capabilities are most likely far from fully exploited, which mostly comes down to the application layer.
A very well-performing open-source model is published by a Chinese AI company. What’s especially interesting is the cumulative training cost is something like $6M for a 600B+ param model. Compare this to GPT-4 which cost >$100M to train. It’s an explicit mixture-of-experts model, with a few interesting architecture tricks like training in FP8 and switching to BF16 for a few optimization runs. They also drop the banger “NVIDIA should make better chips” (I’m paraphrasing) in the section about fine-grained quantization.
Autonomous microbial sensors enables long-term detection of TNT explosive in natural soil
Genetic circuits in microbes that work in a natural environment. They tune and characterize the sensitivity of the microbe towards TNT explosive. But this could be used for general bio-safety by identifying dangerous pathogens, improving fertilizer usage or generally providing geospatial data streams at scale.
Alpha is defined by the “water you swim in” and your peers. It’s no longer a competitive advantage to have access to Bloomberg, a trove of quant PhDs. All investment strategies are information front-running. When Buffet set up his amazing fund structure where he committed to paying his LPs the difference between the annual return goal and actual return, he was largely competing against mom-and-pop investors. Today, markets are mostly sophisticated and e.g., Credit Suisse published explicit strategies for scalping retail (can’t locate the reference at the moment, but it was a gleeful paper).
[Europe under attack] EstLink cable disruption
I love the Finnish way of dealing with Russian state-backed terrorists. Just board the ship that obviously just dragged their anchor throw critical infrastructure and arrest the bastards. At the media interview, respond only mono-syllabically. Mostly “No”. Legends…
For some optimism, Gwern writes about some significant and some less significant improvements to life through technology and science since the 1990s. To further think about how good we have it, I recommend the opening chapters of “How I learned to understand the world” by physician/statistician/public educator Hans Rosling where the national economic benefit and health improvements of having washing machines is highlighted.
OpenAI is free from its allegiance with Microsoft once AGI is reached, which is defined by having $100B in profit. Makes sense that there’s a hard number attached to it. Both parties probably know that’s not necessarily it, but are happy with the incentives.
A very brief overview of one of the most exciting geographical developments of the next 15 years (until we get to Mars). The Arctic is thawing and suddenly unlocking new “riches” in the form of mineral and gas exploration, and trade routes being accessible. If climate change is gonna happen, can we at least try to profit from it?
The excellent Semianalysis has published a number of in-depth analyses of the Blackwell architecture, and now they also went into the supply chain. I highly recommend the training benchmark they posted earlier. Looking forward to NVIDIA and others pushing Moore’s law further in 2025.
End of 2024
In 2025, I’m looking forward to more great reading, exciting investments, a potentially huge governance dynamics with DOGE, various European elections, Russia/Ukraine/Georgia. I also think we will see further acceptance of defense tech investments in Europe, with several new funds and FoFs set up. We’ll see first major practical results from “AI for Science”, further consolidation in AI startups, and large state initiatives for training and applying AI to defense/national security. Stablecoin fintechs will continue to be closer to traditional fintech, with crypto as a category growing more infrastructure focused.
I’m not going to make any more predictions, but I’m certainly very excited about upcoming major milestones in our portfolio (e.g., Radical Aero, Fabric, Lodestar Space, Anytype, Iron to name a few).
If you made it this far, thank you! This first year of SVRGN Weekly Digests has been a great experience. If you have suggestions for improvement, want to submit content, chat about what we put out, please reach out!