(Re)Share #51 - Nevermind the gap
Cheap LLMs | AI policy | Tabular intelligence | Robotic dexterity | Humanoid batteries
It’s been a few weeks since our last issue, but I’ve been heads down on a new investment that we’re making. Nothing I can share yet but there may be Easter eggs below for the sleuths among you. 2025 is kicking off in a big way and the range of deep tech news suggests as much. As my friend Finn so eloquently suggests, the time is now—so let’s get to it.
Stuff worth sharing
Rolling in the DeepSeek - Obviously starting here because it’s all anyone cares about. Last week, a Chinese AI lab, DeepSeek, released an open-source reasoning model, R1, which leverages a chain-of-thought architecture. This was the same team we covered in #50 with the release of their V3 model. This release, however, had a much, much bigger impact. Like the V3 model before it, R1 was developed on a paltry budget of just a few million dollars and the performance was pretty exceptional. There are conflicting reports, but the broad consensus is that R1 is near parity, if not superior, on some tests compared to o3, Claude 3.5, etc. Most notably this model was built by a hedge fund as a side project on a rounding-error budget, making the tech world freaking the f**k out. On Monday, the NASDAQ fell 3% and NVIDIA an astounding a 17% drop—the single biggest one-day loss in U.S. history! Is it an open-source miracle? A Sputnik-style national challenge? A macroeconomic psyop? I’m ignoring the saber-rattling red meat, although there are some very important questions as to the effectiveness of protectionist policies. More generally, DeepSeek proves that advanced intelligence is within reach of any founder. The barriers to entry are falling, and as a pre-seed guy, I’m all for it. Full paper here.
Watch this space - Last week, Bezos’ intergalactic baby, Blue Origin, threw its space helmet in the ring. The spacetech company successfully launched its New Glenn rocket and reached orbit for the very first time. The NG-1 mission objective was to ignite the 320-foot rocket and bring a prototype Blue Ring Pathfinder payload into space—both of which were successful. A secondary goal was to mirror the booster recovery landings that SpaceX has made famous. On that front, Blue Origin was not successful, but that barely registered considering the momentous achievement—especially given the dozens of setbacks the team has faced over the past several months. For the first time, we have a plausible competitor to the SpaceX launch monopoly, and who doesn’t love a friendly rivalry?
Nevermind the gap -The UK unveiled the AI Opportunities Action Plan, a comprehensive 50-part recommendation to position the nation as a global leader in AI development. The plan was led by VC and global AI czar Matt Clifford—full disclosure, I used to work for him at Entrepreneur First. Readers may remember Matt’s name from spearheading the 2023 AI Safety Summit, but gone are the days of overly-cautious growth. The plan covers a lot, but in short, it’s a laundry list of heavy investments aimed at catching up with the US and China. Some of the recommendations are pretty expected, like public compute capacity for researchers, accelerated visa pathways, and reduced regulations. Far more interesting is the introduction of AI Growth Zones (AIGZs) to spur localized investment in data center development. I think this is a fantastic idea to drive growth in parts of the UK that have been broadly left behind. One of the greatest challenges with the UK tech scene is that everything is based in London or Oxbridge. That talent clustering has its benefits, but it severely limits the distribution of opportunity—a reality we’re all too familiar with in the US as well. AIGZs also make economic sense, allowing regions with natural energy advantages to capitalize on them. The most obvious critique is that the plan relies heavily on state capacity and overlooks some of the more obvious capitalist drivers. If the UK—or any EU nation—really wants to attract the best people in the world to build, it all comes down to tax policy and exit pathways.
Put it on my tab - One area of my digital life where AI has been surprisingly absent is spreadsheets. As a card-carrying Type A, I am no stranger to the comforting call of structured tables. Despite spreadsheets being the world’s most accessible programming language, there has been relatively little innovation in intelligence applied to them. The nature of tabular data and the aesthetic components of table design make transformer application challenging, if not ineffective. Thankfully, the good people at Microsoft are doing God’s work. Last summer, they unveiled SpreadsheetLLM - an attempt to overcome token constraints and encoding gaps. This is achieved through three modules: structural-anchor-based compression, inverse index translation, and data-format-aware aggregation. Essentially, it allows an LLM to read the structure of a table and the relationships between rows and columns without ingesting the entire array. If you know me well, you won’t be surprised to hear that I geeked out over this paper—hard. At least I’m not these guys.
Sensory overlord - One of my more recent rabbit holes was robotic dexterity and giving our future mechanical overlords the ability to feel—in the tactile sense. Last fall, Meta FAIR published their research on enhancing touch perception in robotics. They developed a new open-source tactile sensor and corresponding software to improve robots’ ability to understand and interact with their environment through touch. This innovation aims to enable robots to perform tasks requiring fine motor skills and delicate handling, such as manipulating small objects or operating in unstructured settings. My deep dive made me realize just how incredibly impressive—and impossibly complicated—the human hand is. I could now spend hours talking about the unique kinematics we’ve developed and the electrical circuitry that enables us to understand force feedback, but I digress. Today, there are several obstacles to achieving true robotic dexterity, but I’m working on it (wink).
What’s all the colocomotion?! - Friend of Fly, Jenny Read, who runs ARIA’s Robotic Dexterity program, posted an opinionated take on the viability of humanoids on a technical level. This has been one of the more active debates in the community for a couple of years now, and her post provides a good summary of the diverging schools of thought. Specifically, I’m referring to the conversation that developed in the comments, which was wide-ranging. I fall on the same side as Jenny in the view that humanoids are likely over-engineered for the vast majority of needs. But of course, time will tell, and there are several billion dollars that would argue otherwise.
Large and recharge - Speaking of humanoids, The Information had an interesting article investigating the current state of power capacity in bipedal robotics and the limitations of mobile batteries. Most rely on lithium-ion, which, despite widespread use, offers limited operational time (only 1 to 2 hours) before requiring a recharge. There’s a load of alternative power research solutions—hybrid power systems, silicon-based anodes, etc.—but none yet achieve the right balance of energy efficiency, weight, and size. Despite all the time I’ve spent thinking about robotics, I never really considered this nuanced form of range anxiety. My partners at Fly are particularly interested in battery innovation, so I’m excited to see what entrepreneurs come up with. But for now, I do find it mildly entertaining that even mindless robotic drones need breaks.
Monkey see, monkey no longer do - OpenAI released a research preview of Operator, their latest product designed to automate the drudgery of desk work. The AI agent is capable of autonomously performing web-based tasks by interacting with a browser, just as a human would. Operator is a Computer-Using Agent (CUA) model, combining GPT-4o’s vision capabilities with advanced reasoning through reinforcement learning. This enables it to interpret screenshots and interact with GUIs. Like Anthropic’s Computer Use release, it controls and records screen engagement—so user beware.
As a matter of fact - Microsoft Research is back with another release in their materials science work. MatterGen is a generative AI tool designed to revolutionize materials discovery by generating novel compounds based on specified design requirements, such as chemical composition, mechanical strength, electronic properties, or magnetic characteristics. MatterGen employs a diffusion model tailored for 3D material structures, adjusting atomic positions and lattice configurations to create stability. The big achievement here is the experimental transfer, albeit within a class of materials that isn’t super commercially useful. Nevertheless, it’s open-sourced and pushes the envelope of harnessing AI for science.
Stuff worth sharing
The CEO of Transcelestial shared his path into deep tech entrepreneurship with CNBC.
The CEO of Orbital Materials, along with our friends at Compound, shared their view on the future of material science on this episode of Okay, Computer.