(Re)Share | #27 - the AI Safety issue
AI Sentience | Exec Orders | Nuclear Fusion | CRISPR Drugs | Space Babies
I’d like to start this issue off by thanking the OpenAI board for a fantastic season 4 of Succession. Literally nothing else was discussed this weekend in my tech circles and justifiably so. There’s been a lot of armchair commentary already so I’ll spare you from my hot take (although Matt Levine’s is the best, per usual). But I very much look forward to the movie.
Prideful Plug
Two former colleagues that I am beside myself impressed with:
Safety cone of influence - EF co-founder and CEO Matt Clifford successfully led the UK AI Safety Summit, one of the most ambitious events ever organized in the genre. The scale and scope of this thing was astounding and represented the 100 most who’s who of artificial intelligence research, commercial application and policy - Rishi Sunak (obvs), Elon Musk, Kamala Harris, Mustafa Suleyman. Of course I was totally asked to go...
Dear Sugar - EU Talent magnate and good friend of mine, Zoe Hewitt, unveiled a shiny new career advice column in Sifted. I’m super excited for her and I think this is a fantastic way to scale Zoe’s insights and expertise that many VCs and founders have been lucky enough to see first hand.
Stuff Worth Sharing
Who am (A)I? - A light dive into the complex and hotly debated world of consciousness and how we might detect it within an artificial host (e.g., AI). We’ve covered the potential tests that could be run to assess sentience in past issues, but this article makes an interesting new argument:
Conscious processes like holding something in short-term memory are pretty limited—we can only pay attention to a couple of things at a time and often struggle to do simple tasks like remembering a phone number long enough to call it. It’s not immediately obvious what an AI would gain from consciousness.
Then again, we could be walking down a path of robotic enslavement, seems like a not good thing. (really long discussion with Paul Christiano on that).
White hat house - The Biden administration unveiled an executive order on AI safety, which unleashed all sorts of fury across the interwebs. Doomers cried it was toothless, Technoptimists wailed over an ounce friction towards progress, the whole thing feels like the gun debate. My two cents: the EO safety standards seems reasonable if not vague. Details around the threshold of national security are naturally unclear and to be expected. The most debated aspect relates to the compute levels that would require red team performance disclosure. Based on my current understanding these levels are quite reasonable (though correct me if I’m wrong). In reality this will end up creating a cat and mouse dynamic, similar to what we already see with chip controls. Beyond the safety dimensions it’s mostly aspiration signaling and / or punts to Congress.
Don’t touch my stuff - The Coalition for Content Provenance and Authentication (C2PA) has been developing an interesting breadcrumb watermark approach to image authenticity. By creating a tamper proof ledger for every image a viewer can now see the full history of the sale, transfer, edit, etc. Last month the org unveiled their “icon of transparency” - a Michelin star guide for AI safety that blends of technical sophistication and behavioral conditioning.
Finding NeMo - NVIDIA built a custom LLM that can assist one of the most complex engineering efforts: designing semiconductors. The initial killer apps are maintaining known bug descriptions and GPU architecture Q&A. But it’s just a matter of time before NeMo has a hand in designing its progeny.
Asian fusion - Japan currently holds the title of world’s largest fusion reactor. The JT-60SA is a four-story-high behemoth that harnesses massive magnetic fields from superconducting coils to contain a blazingly hot cloud of plasma to force hydrogen nuclei to fuse and release energy. Recent test runs have proven the basic design and functionality, providing great confidence in a fusion future.
Above the fold - Deepmind’s seemingly unstoppable development on protein folding is a frequent topic here so it should surprise no one that they did it again. The latest iteration of AlphaFold demonstrates significant improvement over multiple biomolecule classes - ligands, proteins, nucleic acids (DNA / RNA), and PTMs. The most near term application of this unlock is in new drug design, specifically through enhanced ligand / protein binding - a field I find extremely interesting. But the post goes on to frame AlphaFold’s new potential in gene editing, molecular machines (cell factories) and more.
248 mile high club - Can we have babies in space? That is the founding question of biotech startup SpaceBorn United. The company has designed a novel apparatus which can fertilize rodent embryos in space using simulated gravity. The embryos are then cryogenically frozen, returned back to Earth and implanted into a rodent mother. If this results in the birth of healthy pups the technology could eventually extend to human embryos. Just in time for space tourism to get really, really uncomfortable.
My chemical romance - A comprehensive analysis by Microsoft Research on the capabilities and limits of LLMs on scientific discovery. With a specific focus on GPT4, the paper dives into 6 disciplines - drug discovery, biology, computational chemistry, materials design, and partial differential equations (PDE), each with their own subfields. Far too much to go into here but I particularly enjoyed the material science section (pg. 126) whose predictions (strengths: composition creation & synthesis planning / weaknesses: synthesis route) broadly match the current thinking at Orbital Materials.
Rock, paper, genetic scissors - The world has its first CRISPR/Cas9 approved therapy and marks perhaps the biggest development in the technology’s ~25 year history. Casgevy, is for patients with sickle cell disease and a related blood disorder called beta thalassemia, both of which are inherited. In a study of 42 beta thalassemia patients, 39 did not need a blood transfusion for at least a year after treatment and the remaining three had more than a 70% reduction in transfusions.
Rays your hands up - the rollercoaster tale of Sunrun, the largest installer of residential solar panels in the US.
Portfolio Flex
Wayve got a much deserved shoutout from the UK Prime Minister.
Kheiron Medical featured in Nature Medicine.
compared to double [human] reading, implementing the AI-assisted additional-reader process could achieve 0.7–1.6 additional cancer detection per 1,000 cases, with 0.16–0.30% additional recalls, 0–0.23% unnecessary recalls and a 0.1–1.9% increase in positive predictive value
A happy Thanksgiving to all my US readers and to those that are making the long trek North to Slush next week - see you there!