(Re)Share | #28 - A material impact
Multimodal models | Material science | EU AI Act | Nuclear reactors | TTOs
Welcome to the final (Re)Share) of 2023! It’s been a remarkably dynamic year for the tech industry and while I find it impossible to keep up I thoroughly enjoy the process of trying via this newsletter. Thank you to my readers who have shared some kind words of encouragement and to all of you that have shared this effort with others. This issue is slightly longer than normal in case you need a bit of extra reading to avoid your in-laws over the holidays. Please enjoy and I’ll see you all 2024!
Prideful Plug
All that glitters is gold - A very, very big congratulations to my good friend Hussein Kanji for his debut on the European Midas List. Hussein was the very first VC I got to know when I first landed in London and has been a constant source of support and guidance over the years. He embodies independent, often contrarian thinking and has consistently received rave reviews from the founders that he backs.
Stuff Worth Sharing
Digital twin - Google raised the stakes in the LLM race with the unveiling of Gemini, their multimodal AI model. This took the interwebs by storm so I doubt this is the first you’re seeing it. The intro video is magical, though it turned out that this was far more edited than initially described, muting some of the wonder. Because Gemini has been built with multimodal at its foundation it’s able to do some pretty remarkable stuff across text, images, audio, video and code. Per the press release, in head to head text-based question testing Gemini scores 90% and human experts score approximately 89% and GPT-4 scores 86% On the multimodal questions, Gemini scores 59%, while GPT-4 scores 57%. My personal favorite demonstration was unlocking scientific literature although I had a holy shit reaction to this omelette scene.
A warning sine - There are a lot of takes on what exactly went down with the OpenAI boardroom drama, but it seems that the emergency, break glass move was over AGI fears. Q* (pronounced Q-Star) a powerful new model built internally, was able to correctly overcome elementary level math problems, long thought to be a major milestone in the path to robotic overlords.
Regulate to the party - The EU passed its far reaching AI Act, the first of a kind formal regulation to control AI systems. This is the GDPR sequel that’s been years in the making and cements the bloc’s role as the digital world police. Typical Twitter reactions ensued from the established camps and of course both war cries lack the required nuance. Some of the Act’s requirements are (IMO) quite reasonable, like watermarking AI-generated content and limiting government overreach (e.g. biometric profiling) but quite a bit is naive and toothless. The most ridiculous of which is a overly vague self assessment apparatus for transparency and “impact”, which is only going to result in a lot of lawyer fees. The most interesting wrinkle is that France is pushing back on setting compute limits with Macron refusing to fall behind.
Do not pass Go - YC hosted a conversation with FTC Chair Lina Khan and it’s a must listen for founders and investors alike. Khan, the youngest ever leader for the agency, gained notoriety for her fiery disdain for Amazon and her macro view that the US’ approach to assessing monopoly power was outdated. The Tl;dr being that consumer pricing alone is an ineffective measure of market health and Many in the tech community rail against Khan’s expansive regulatory stance over the past few years although her perspective does resonate remarkably close to the more recent concerns over AI centralization.
Lessons in Chemistry - Scientists got an early Christmas gift with ORDerly, an open source tool that’s designed to better prepare datasets and benchmarks for chemical reaction research. The Python package is designed for customizing and reproducibly preparing reaction data by cleansing US patent records in line with the Open Reaction Database schema. Big implications for medicines and materials.
When it rains it pores - More ml-powered material discovery; this time a hyper efficient carbon-material that could drastically improve supercapictors. The Oak Ridge National Laboratory (ORNL) designed a machine learning model to develop a "dream material" for energy delivery, which resulted in an extremely porous oxygen-doped carbon that showed a 4x improvement in storage capacity. The enhanced performance is driven by a unique design featuring two pore sizes, which provides huge surface areas for electrochemical reactions - namely energy storage and electrolyte transport. Full paper here.
His dark materials - Three times in one newsletter!? Deepmind once again graces the digital pages of (Re)Share, this time for material science. GNoME is their latest graph neural network (GNN) model that is purpose built to discover and test inorganic crystals, a core component of computer chips, solar panels, batteries and more. GNoME discovered 2.2 million new crystals in total and 380,000 stable materials, which equals nearly 800 years of our current knowledge base. For frame of reference, The Materials Project has so far discovered ~20,000 crystal structures to date using computational methods.
Original spin - The UK released a comprehensive review of university spinouts and the results were shocking. The report is full of interesting research so I do recommend a read but if you’re short on time I’d direct you to page 33, which summarizes the range of equity stakes taken by universities. On average Tech Transfer Offices take a 24% ownership stake and if you threw up after reading that, rest assured that is the correct response. In my day job I see a good amount of academic spinouts and I’m pleased to say I’ve never seen such egregious levels, but clearly it takes place. I applaud the UK for undertaking this review and the government’s recommended actions are quite sensible, but we’re very behind where we should be.
Gone fission - Episode 4 of the Age of Miracles podcast focused on the manufacturing challenges of building nuclear and the novel solutions being proposed. Ops nerds will love this. Even if you’re in the anti-nuclear camp the ingenuity of production innovation is captivating. My personal favorite comes from Blue Energy who is ignoring reactor design or new fuel chemistry and instead rethinking the way we build reactors by repurposing shipyards and transporting pre-fab reactors to the site.
Walk the plankton - Hong Kong-based AIGreen discussed their efforts to engineer a scalable carbon capture solution through the miracle plant of algae. The team is approaching the challenge by designing a photo-bioreactor for air purification in what they call a "liquid plant". They aim for a 1,000-liter scale bioreactor powered solely by solar energy in the coming year. Fly has a related investment in Phycoworks, which takes a more computational approach to strain engineering itself. A fascinating area with more than enough carbon to go around.
Don’t throw Rosetta stones - As an American I enjoy my God given right to speak one and only one language. While my European colleagues may enjoy smug bilingual superiority now, those days are numbered. Meta released Seamless, a suite of AI language translation models that can preserve expression, tone, etc. in real time. Seamless can handle almost 100 languages with attention to elements like speech rate and pauses. Full blog here.
Extra credit? - The University of Berkeley released the results of a three-year study that investigated the legitimacy of carbon offset procurement. Tl;dr it’s a lot of bulls**t. For those that follow this space, the findings won’t be new - additionality concerns, over crediting, pricing variance - but the depth of the report is what stands out. Like most institutions, Berkeley seems to have taken a seemingly militant view that only REDD+ credits (i.e. trees) are worthy, which is probably why the report ends the way so many others have. As I’ve said before, bring on the engineered solutions!
The UnGrateful Dead - Anduril unveiled Roadrunner, an autonomous kamikaze drone. Roadrunners offer faster launch and take-off time, three times greater warhead payload capacity, 10 times effective range and three times greater maneuverability in G-forces over legacy drone systems. But the most notable feature is the drone’s ability to be reused and refueled, similar to have SpaceX has revolutionized orbital rockets. Demo video here.
To infinity and bey.. - Speaking of SpaceX, last month brought the second test flight of Starship. 33 Raptor engines of the Super Heavy booster successfully fired and brought the spacecraft to the company’s first “hot stage” separation, a first for a vehicle of this size. The booster didn’t make the return trip, but the hot stage milestone was the real win. The video linked above is nearly an hour long, so if you’re short on time I’d point you to minutes 3:00, 16:45, 27:30 and 38:45 + this tweet.
A token of appreciation - The UK took a big step forward in the journey to digitize real world assets. The FCA has outlined a series a efforts to sort out the safe and legal enablement of tokenization. Web3 may have fallen out of favor in the zeitgeist but the potential of blockchain to drastically improve financial efficiency still remains. I’m quite excited to see this development, though I’m highly biased by our investment in Palisade.
A chip on their shoulder - New York is the latest to throw its hat in the semiconductor ring. The state unveiled a plan to invest $1 billion in chip research, attract another $9 billion in industry funding and host a new national semiconductor technology center.
Portfolio Flex
The Knowledge - Wayve is back with another autonomous ride along, this time with guest Azeem Azhar of Exponential View. EV was one of my first tech newsletters and it remains a staple of my reading diet to this day. As for Wayve - the same remarkable engineering feat that we’ve grown to love. Super complex London roads are no match.
Booster shot - DeliveryHero released a blog outlining their modern ml stack and they highlight Evidently AI as their go-to solution for drift detection. Drift is a mission critical issue to address for most companies but especially the high volume nature of delivery. Friendly reminder to technical readers, Evidently is 100% open source!