(Re)Share | #32 - Battle Roy-AI-le
Diffusion models | Text-to-video | Particle accelerators | Robot challenges | SAI
Today was a very fun week in the AI world. Not since Zoolander vs. Hansel has a there been such an entertaining walk-off. All of which has been painstakingly curated by yours truly for your reading pleasure. Let’s get to it!
Stuff Worth Sharing
It’s not the size that counts? - Google has been pushing Gemini releases like hotcakes and the latest model just dropped. Gemini 1.5 Pro is the newest mid-size multimodal model that’s optimized for scaling across a wide-range of tasks and is built off a Mixture-of-Experts (MoE) architecture (divided into smaller "expert” neural networks). This was newsworthy in particular because of the staggering context window of 1 million tokens, which roughly translates to 1 hour of video, 11 hours of audio, 30,000+ line codebases or 700,000 words. According to Jeff Dean they’ve actually pushed it to 10 million somehow. The demos in the link are crazy. An automatic ingestion of the entire Apollo 11 mission transcript is my particular favorite but I think the video search functionality is the biggest oh sh*t moment.
Movie magic - Not to be outdone, OpenAI answered the call with their release of Sora, a text-to-video diffusion model that is next next level. I don’t even have words for this one, it’s just magical. While still in private beta, the model shows extremely impressive image quality and overall length that actually achieve the most Turing complete photorealism that I’ve personally seen. There are obvious risks with something like this, but the creativity explosion it can usher in is undeniable. Also my condolences to Pixar.
Ozymandias 2.0 - Sam Altman: prolific angel investor, inspiring founder, AI visionary and most recently, Bond villain? Last week Altman made news as he traveled the globe to charm investors and governments into funding his latest effort to remake the global semiconductor industry all for the very reasonable and in no way spit-take inducing sum of seven trillion dollars. Of course I certainly don’t believe he has illicit motives, but raising that amount of money for the singular goal of ushering in our synthetic descendants does feel like it would fit right into a science fiction novel. While that amount of money may be absurd the likelihood of some significant deal being struck does seem strong with all the geopolitical pressure and economic saber rattling over the past 12 months.
What’s the matter? - Plans for CERN’s proposed new particle accelerator, the Future Circular Collider (FCC), were unveiled and it’s a doozy. At 91 kms long, 5.5 meters in diameter and a punchy $17 billion in cost, the FCC would 3x the Large Hadron Collider (LHC) and provide scientists an unparalleled ability to explore dark matter, bosons and other things I don’t understand. Firing matter at the speed of light is a bit of an engineering challenge so we wouldn’t see any particle smashing until ~2050 and the start its highest-energy collisions around 2070.
Powerlifting - I showed my wife this video of Boston Dynamics’ Atlas robot and her immediate response was, “well that is terrifying”. Though I consider myself a techno-optimist, if you did worry about terminator-like murderbots this is shockingly close. This demonstration shows the robot’s fine motor controls as it picks up and relocates some 30 pound car parts. I continue to be doubtful on the economic viability of humanoid robots (more complex than is necessary) but it’s still very very cool to watch.
Material evidence - A very interesting recent paper on a foundation model for atomistic chemistry. Simulation and synthesis of material science and molecular reaction has been discussed a few times here in the past but this paper was notable for the claim of its’ application aperture. Nearly every model that aims to discover / unlock / engineer at a molecular level is purpose built - e.g. a model for gas / liquid phase transition, for heat exchange, for zeolites, etc. This diverse team of researchers developed MACE-MP-0 model, a simulation engine built off the Materials Project that was able to maintain qualitative and at times quantitative accuracy diverse set problems in the physical sciences (e.g. transferability). The model has its limits, of course, and from the footnotes they’re not replacing purpose-built models anytime soon. But I may or may not be speaking with an author turned founder…
A hot startup - I often think that San Francisco is basically Gotham City. There are a variety of reasons for this but most notably because its citizens love to do crimes. On Sunday a crowd vandalized an autonomous Waymo car and things slightly escalated to include setting it on fire. I can appreciate that SF’s Superbowl loss may have stung a little, but collective arson? Very thankfully no one was in the car at the time.
Sunscreen spray - Stratospheric aerosol injection (SAI) is a bit of a catch-22. It remains extremely controversial so there’s very little research and because there’s little research it remains risky and controversial. This article from MIT Technology Review sheds a bit of light into how we might deploy SAI in a controlled manner to generate impact and risk data. It’s been a while since we spoke about SAI so the brief TL;DR - shooting sulfur oxide into the stratosphere to partially reflect sunlight and lower the atmospheric temperature. The article describes the logistical challenges of SAI, namely deploying at a sufficient height, which is believed to be ~20km above sea level, at least near the equator. That’s 2x the cruising altitude of most aircraft apart from spy planes so a simple aerial dump won’t work. There is an interesting geospatial theory which takes advantage of the slope of the atmosphere across latitudes. There’s a lot more to the novel approaches, but this summary is already too long.
Taken for granted - As part of my ongoing robotics research I had the pleasure of speaking with Jenny Read, head of ARIA’s in development program on robotics. For those unfamiliar, ARIA is the UK’s answer to DARPA; a non-dilutive government funding mechanism to catalyze radical technical innovation that can generate significant societal impact. ARIA’s projects include everything from programmable plants to deep compute but Jenny is specifically aiming to innovate in the field of robotic dexterity. If you’re a UK / EU founder with an aspiration to do some wild stuff in this area I highly encourage you to submit an application.