🔮 Q*; elite overproduction; battery breakthrough; bacterial memories ++ #450
An insider’s 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.
Latest posts
If you’re not a subscriber, here’s what you missed recently:
Sunday chart: AI capabilities
Just when you think the OpenAI drama is over, it bursts open again. Sam Altman is back as CEO, and now there’s talk that his firing was due to a secret project, Q*. According to Reuters:
Several staff researchers wrote a letter to the board of directors warning of a powerful artificial intelligence discovery that they said could threaten humanity.
Reportedly, this secret AI can do grade-school maths on problems it hasn’t seen before. This has previously been a challenge for LLMs like GPT-4. While ‘grade-school maths’ doesn’t seem like much, with exponential progress of the type we’ve witnessed, grade-school maths could quickly become PhD maths. (The Information has more details on what this breakthrough could be. Others speculate it is about a different approach to reinforcement learning that can operate “model-free”, in complex or highly evolving environments. Yann LeCun suggests it could be about planning. But equally, this could be no more than breathless hype, diverting attention from OpenAI’s recent troubles.)
Let’s imagine that this “grade school maths” thing is real. So far, LLMs have built capabilities quickly. A recent study by Anthropic, Cohere, and NYU researchers pitted humans, GPT-4 and GPT-3.5 against a range of difficult graduate science problems. Highly skilled non-experts, pursuing PhDs in unrelated disciplines, were given access to Google to help them. They got 34% of questions right, better than GPT-3.5 but worse than GPT-4. PhD students in their own domains, naturally, did the best, scoring 65%. So, today’s state-of-the-art token predictor is nowhere near expert (PhD-level) performance, but it is better than your smarter-than-average-bear with a search engine–and will likely get better.
Perhaps, Altman had been giving hints. One day before his ouster, he said:
Four times now in the history of OpenAI, the most recent time was just in the last couple weeks, I’ve gotten to be in the room, when we sort of push the veil of ignorance back and the frontier of discovery forward, and getting to do that is the professional honour of a lifetime
This week won’t be the end of the drama. To help contextualise things and what might come next, I spoke with Karen Hao, a contributing writer at The Atlantic, who has been researching the company closely. We had a great conversation.
Key reads
Elite excess. Talking about capability leaps… What happens when AI displaces all but the very best experts? Peter Turchin discusses the potentially destabilising impact of AI on society, focusing on the erosion of workers’ social power. He highlights the risks posed by an oversupply of highly educated individuals, especially lawyers, who may become radicalised due to AI-induced unemployment. If you’re familiar with his work, then you’ll recognise his central thesis: that the formation of a counter-elite (alongside rising inequality) drives political disintegration.
As we pointed out previously in our Chartpack on AI and white-collar work, degree-educated professionals are likely the most exposed to AI. There is a strong bottom-up demand for white-collar automation tools. Tome (presentation building), Replit (software development) and Runaway (video editing) are all seeing exponential growth.
AI is currently augmenting, not replacing, these jobs, so they’ll become less common but won’t disappear. Yet, as I discussed with Perplexity co-founder and CEO Aravind Srinivas last week, we don’t know how long this “co-pilot” era will last. In 1997, Deep Blue beat Gary Kasparov, and for the next eight years, Centaur Chess (humans with AI co-pilots) outperformed either humans or computers alone. By 2005, computers had beaten all comers.
Context ain’t king. Anthropic released Claude-2.1 with a whopping 200k context window. This translates to roughly 500 pages of text, and nearly double the context window of the GPT-4 Turbo model, which has 128k context windows. But it’s not all about length. Accuracy in retrieving information decreases the larger the context length, as highlighted by Greg Kamradt’s analysis. GPT-4 Turbo on the other hand, while smaller, seems to maintain more accuracy across context lengths. The lesson? Conciseness is still going to give you a better performance. See also, my discussion with Anthropic’s co-founder and CEO Dario Amodei.
See also:
While everyone was looking at the drama elsewhere, Meta quietly disbanded its Responsible AI team.
A study on GPT-4 suggests it can’t understand new ideas and abstract concepts it hasn’t seen in its training.
A taxonomy to categorise hallucinations in LLMs, examining their causes, detection methods, and solutions.
Small model releases had a good week. Orca 2 at 13B parameters managed to outperform Llama 2 at 70B parameters across many evaluation tests. Intel Neural Chat 7B is the top-ranked 7b parameter model on the Open LLM leaderboard.
Salty batteries. Northvolt AB, a Swedish battery company, has revealed its first sodium-ion battery - “the first product ever completely free from critical raw materials,” according to Patrik Andreasson, Northvolt’s VP of strategy and sustainability. It holds over 160 watt-hours per kilogram of energy, slightly less than lithium-ion batteries’ 200-300 watt-hours per kilogram. If manufacturing successfully scales, this is a game changer for battery storage and reduces reliance on China which has a stranglehold on many critical battery minerals like cobalt and lithium.
Last November, I talked about how the pressure to reduce reliance on these critical minerals could spur new market developments:
The point is that there are technological choices we can make. And these choices can be spurred by economic incentives, political pressure, and strategic decisions. They can be catalysed by new technologies, or by simpler types of innovation. And they can fundamentally adjust our understanding of how markets might develop.
Annual demand for cobalt and lithium were both forecast to increase by 450% in 2050 but if the development of alternative battery chemistries continues, this might not be such a sure thing.
See also:
Hard, fast and achievable. BloombergNEF released a report on the prospects of tripling global renewables by 2030 (their reports are usually paywalled so worth a read).
Scientists have created a supercapacitor material that stores four times more energy than existing materials.
Globally, falling battery costs and their grid-stabilising role are challenging gas power economics, causing delays or cancellations in 67 gas plant projects in the first half of this year.
Market data
Goldman Sachs forecasts battery pack prices to drop to $70/kWh by 2030. This is 20% less than previous estimates and more than a 50% price drop compared to 2023.
In the first 10 months of 2023, 60.5% of Chile’s electricity came from renewable sources.
On the Sunday of the OpenAI crisis, there was a 27% increase in Eight Sleep app users getting under 5 hours of sleep in San Francisco.
Right now, e-bikes and scooters are cutting oil demand four times more than all the world’s electric vehicles.
Using AI to help read medical exams led to the finding of 13% more breast cancers. Of 30 extra cancers detected, 83% were invasive types that would’ve been missed without AI.
Across all income brackets, Americans say they need a 30-50% raise to feel happy/less stressed.
Short morsels to appear smart at dinner parties
🛫 The evolving social psychology of flying.
🇨🇳 Major social platforms in China now require influencers to reveal their real names to the public - a move sparking many to leave those platforms out of security concerns.
💻 Zoom fatigue is real, according to neuroscience.
🧫 Some bacteria use iron to store memories and pass them onto the next generation.
🔎 David Attenborough’s long-beaked echidna has been caught on camera for the first time.
🌊 Another piece of evidence that deep sea mining harms marine ecosystems.
🐣 MixedNames helps you find bilingual baby names.
🎶 Mysteries of rapper André 3000’s new flute album.
End note
According to Steve Jurvetson, Google now spends more on compute than on people. It’s a watershed moment.
Have a great week!
A
What you’re up to — community updates
Gianni Giacomelli co-authored a paper on using generative AI for innovation.
David Aronchick’s Expanso, a distributed data processing startup, raised a $7.5 million seed round.
Stephen Cave is director of the new Institute for Technology and Humanity at Cambridge University.
Nick Arini’s Generative Engineering - a startup with the goal of increasing the efficiency of physical engineering - just raised a US$4 million pre-seed round.
Chanuki Seresinhe expands the Beautiful Places AI Kickstarter campaign.
Joel Hellermark and the Sana team launched Sana AI, a model-agnostic platform that helps companies build specific knowledge graph and create custom assistants to augment teams and workflows.
Share your updates with EV readers by telling us what you’re up to here.