🔮 Superduperconductivity; Stack underflow; AI benchmarks; Gravity holes ++ #433
Your insider guide to AI and exponential technologies
Hi, I’m Azeem Azhar. As a global expert on exponential technologies, I advise governments, some of the world’s largest firms, and investors on how to make sense of our exponential future. Every Sunday, I share my view on developments that I think you should know about in this newsletter.
Latest posts
If you’re not a subscriber, here’s what you missed recently:
🚀 Thanks to our sponsor NexGen Cloud, NVIDIA Elite Partner.
Sunday chart: A new era?
Sukbae Lee, Ji-Hoon Kim and Young-Wan Kwon: names that will either be in future textbooks or tarnished by scandal. Last Saturday they published a paper on arXiv claiming to have discovered the world's first room-temperature ambient pressure superconductor, LK-99, a compound of lead, copper, phosphate and oxygen. A discovery that, if confirmed, would herald “a new era for humanity”.
When I was in high school I was lucky enough to play with early high-temperature superconductors, in this case a YBCO ceramic, chilled to a balmy -190ºC or so by the liquid nitrogen I was pouring. It was magical to see the little disc levitate. In the decades since we’ve been able to develop superconductors that work at warmer temperatures (although not room temperature) but they also insist on extremely high pressures. If superconductivity becomes possible beyond these constraints, it will open up potential applications ranging from quantum computers to magnetically levitating trains. For a great explanation of why superconductivity matters, Andrew Core’s tweets are worth reading.
However, the researchers’ claims will need to be independently verified before these possibilities can become a reality.1 Superconductor research has been rife with controversies. And the Internet has come alive with physicists and others pouring over and critiquing the paper, prediction markets firing up (rather pessimistically), and researchers of different hues trying to replicate the results. Andrew McCalip is live-tweeting his attempts to recreate LK-99.
Regardless of whether LK-99 is a hoax or revolution, it has reminded me of the power of Twitter (or the Internet) at its best. A rushed pre-print paper straight to Arxiv. Pseudonymous Internet sleuths digging into the background story, researchers collaborating to dissect the paper and thousands, millions of us following along with the replication experiments. We’ll know soon enough.
The interest and excitement around LK-99 show that there is an appetite for real news about meaningful scientific breakthroughs, for passionate discussion that isn’t about culture wars or influencers. It’s a reminder of the power of collective intelligence, in which a web of scientists across different disciplines can pick up a discovery and analyse it from multiple angles, far faster than they ever could in the past.
🚀 Today’s edition is supported by NexGen Cloud
Supercomputers: A new asset class? Experience a first-of-its-kind limited opportunity with Elite NVIDIA Partner, NexGen Cloud. With NexGen Cloud, you can now own supercomputers, purpose-built for servicing AI - a move that brings the once exclusive domain of Big Tech into your hands. Your supercomputer works around the clock to generate monthly uncapped ROI, targeting 150% profit over five years.
What made me decide for this investment opportunity? A channel for earning regular passive income in a unique industry, with significant growth potential. — Simon, Fintech Engineer, UK
Key (AI) reads
The unravelling of Stack Overflow. Over the past year, Stack Overflow, the go-to Q&A site for coders, has lost 40% of its traffic, possibly to ChatGPT and some specific LLMs for coders. Equipped with extensive knowledge of multiple programming languages and delivering instant feedback, ChatGPT offers undeniable convenience. Yet I question its superiority. While it excels at common coding challenges, ChatGPT may face difficulties with advanced and novel problems — which is exactly where Stack Overflow’s collective intelligence has shown its greatest proficiency. This proficiency may not last, though, if coders continue to desert it. (This raises the question of where generative models will get their training data without turning into autophagic ouroboroi, learning from lower and lower quality versions of themselves, and ultimately becoming useless. An interesting paper describes this process and the importance of injecting fresh non-synthetic data into such systems.)
The Turing test has been conquered, what’s next? Now that AI models can pass the Turing test with ease, researchers need a new way of evaluating them (see here for our Chartpack on the topic). Although LLMs excel in various reasoning benchmarks, they struggle with abstract reasoning tasks like the visual logic puzzles in the ConceptARC benchmark. The performance disparity between LLMs and humans is notable; GPT-4, for instance, scores only 33% on the ConceptARC benchmark compared to humans’ average of 91%. These results continue the age-old debate: are these models truly intelligent, or are they merely adept at spotting patterns? One thing is clear, their ‘intelligence’ fundamentally differs from that of humans. We will need benchmarks that provide a balanced view of these models’ strengths and weaknesses.
See also: The benchmark WebArena was released this week to test LLM’s ability to navigate the web autonomously.
Game of Codes: Who writes the AI rules? Anthropic, Google, Microsoft, and OpenAI have formed the Frontier Model Forum, an industry body focused on the safe and responsible development of frontier AI models. Their move is part of a wider push by the AI industry to self-govern and assert its commitment to safe AI development. Given their close involvement with these technologies, the Forum members’ ability - and suitability - to regulate them remains to be seen. In any case, they are getting government regulation whether they want it or not. So far, they’ve committed to the safeguards recommended by the White House like independent testing and data sharing, but their full compliance is still uncertain. The commitment of OpenAI in particular seems inconsistent with its opaque working practices (e.g., not releasing GPT4’s architecture). Is the Forum a promising step towards more responsible AI development or just a case of regulatory capture? As EV reader, Carissa Véliz points out in this trenchant critique “no industry can regulate itself.”
See also: A coalition of a half-dozen open-source AI stakeholders (including Hugging Face & GitHub) is calling on EU policymakers to protect open-source innovation as they finalise the EU’s AI act, which so far looks to cripple open-source models (according to the coalition, lawmakers may disagree).
Data
The number of full days working from home in the US has flattened out at around 30% in 2023.
Japan’s population drops by nearly 800k in a year, around a 0.5% reduction.
Around a quarter of European companies progress from Seed to Series A funding within three years. This rate, stable from 2012 to 2018, persists despite increased overall funding.
The excess death rate among Republican voters was 43% higher than the rate among Democratic voters after vaccines were made available in the US.
South Africa increased solar installations from 1GW to 4.4GW in 14 months. South Africans seem to have been spurred by the constant load shedding and unreliability of their energy sector.
Clinicians express a preference for Med-PaLM chest X-rays over those produced by radiologists in up to 40.5% of cases.
Short morsels to appear smart at dinner parties
🔗 The strength of weak ties, 50 years on. Granovetter’s paper is one of the most important in history.
🌊 A ‘gravity hole’ in the Indian Ocean explained.
💾 A deep-dive on the old faithful IBM mainframe.
🔎 OpenAI shuts down its AI detection tool.
💥 How the war in Ukraine is revolutionising drone warfare.
🙏 Who would have thought? Robot preachers get fewer donations.
⭐ Twinkling of giant stars reveals how their innards churn in first-ever simulations.
🧵 Why we’re struggling to ditch Twitter for Threads.
End note
My buddy Barney Pell reminded me of just how good Richard Dawkins’ The Blind Watchmaker is. I hadn’t read it for over thirty years and picked it up this week. It’s a really light, digestible and thought-provoking read if you want to remind yourself how complex machines can arise without super clever designers. As a chaser, it works really well with another book I’m reading, David Deutsch’s The Beginning of Infinity.
Have a great week!
Azeem
What you’re up to — community updates
Husayn Kassai is building an AI coach delivering ‘just-in-time’ learning.
Dave Goldblatt, the GP of Vibe Capital, is announcing Fund III, a $10m Fund that invests in the AI, Web3, and Deep Tech sectors at the Pre-Seed and Seed stage. He is actively looking for aligned LPs to invest. If interested, reach out to Dave at dave@vibecap.co.
Daniel Stanley is holding an online event on ‘Using AI to Democratise the Economy’ on 2 August with speakers from the Institute for the Future of Work, and the Ada Lovelace Institute.
Abhishek Gupta is launching a journal club for augmented collective intelligence that brings together researchers and practitioners in the field of AI and complexity sciences across industry and academia.
Share your updates with EV readers by telling us what you’re up to here.
Only last week a Nature paper on superconductors was retracted after being found to include falsified data.
@azeem - key caveat for the Seed to Series A data: it appears to be for Europe *only*
Likely in the US or other geographies quite different
I watched a video last night of a chinese guy who went to a subburban building, and into the basement of the claimed superconductor lab, they wouldn't let him in, but they gave him a card, and were having water delivered. I'll see if I can find it.