🔮 The harsh numbers of the energy transition; autonomous coders; sovereign AI; consciousness & AGI ++ #465
Hi, I’m Azeem Azhar. In this week’s edition, we explore
The challenges and opportunities in the global energy transition.
The potential impact of AI on software engineering and the emergence of autonomous AI systems.
The growing importance of sovereign AI strategies for nations.
🚀 Today’s edition is supported by OctoAI, the fastest generative AI inference API.
Sunday chart: Can David slay Goliath?
Vaclav Smil’s guest letter in Electravision, a J.P. Morgan report on electrification, paints a sobering picture of the global energy transition. Despite a quarter-century of counter-efforts, the world has increased its dependence on fossil fuels, with global emissions continuing to rise (shown in the attached chart). Smil highlights just how complex the energy transition is, requiring a systemic shift involving multiple interconnected components and several expensive-to-abate sectors. Global steel production, which chunters out 7% of global emissions, currently amounts to 1.9 billion tonnes. Of this, green steel represents less than 0.02%. Exponentialists will recognise it would take 12-13 years of doubling green steel production to replace traditional processes. At least we know how steep, and high, the climb is.
However, we should not succumb to fatalism in the face of these challenges. The economics are rapidly shifting in favour of technology-driven approaches rather than brute-force, industrial-era ones — for example, solar panels (that generate electricity!) are now cheaper than plastic faux-wood fence panels. The learning rates of established technologies (like solar, batteries and wind) still have several years to run — and prices will continue to drop. Newer technologies, like electrolysers and clean industrial processes, have many more iterations of learning-based deflation ahead of them. And even in the bureaucratic West, we’ve successfully scaled hard technologies in less than a decade. Let’s look at the rampant expansion of horizontal wells in the US: in 1995, fewer than 6% of oil and gas drilling rigs in operation were horizontal. By 2014, this number jumped to 68% — and the total rig inventory doubled.
Of course, hope isn’t a great strategy. So, we must consider diverse approaches, including a better scientific understanding of higher-risk options like solar geoengineering. Now it’s more important than ever to engage in transparent discussions about adaptation strategies.
See also: to get a grounding of where the oil industry is at, we recommend EV reader Vijay Vaitheeswaran’s special report in The Economist. Elsewhere, a summary of recent milestones in commercial nuclear fusion.
🚀 Today’s edition is supported by OctoAI.
In the fast-paced world of AI, keeping up with the latest LLM releases is crucial for builders to deliver cutting-edge experiences and minimise costs. With OctoAI’s new LLM Time Capsule, you can witness the exponential growth of over 50,000 text-gen in just months, the rise of open-source alternatives to GPT-3.5 Turbo, and the massive impact of model releases on the industry.
Exponential View readers get $10 free credit when you sign up for OctoAI’s GenAI Inference API to try fast and affordable versions of these models for free.
Key reads
Devin ex machina. A year ago, we pointed out Paul Kedrosky’s argument that large language models could significantly impact software production by reducing the need for software engineers. While this hasn’t yet fully materialised, AI has been widely implemented as a coding co-pilot — albeit one requiring human intervention and producing more than the occasional error. And this week, AI lab Cognition demoed Devin, the world’s first ‘fully autonomous software engineer’. It’s worth viewing this with a critical eye until it’s publicly stress tested, but from its benchmark results, it does appear impressive, resolving 13.86% of issues unassisted on the SWE-bench benchmark1, a significant improvement from the previous record of 1.96%. Although a 14% success rate may seem low, it is reminiscent of AlexNet’s breakthrough in computer vision. In the 2012 ImageNet Challenge, AlexNet achieved an error rate of 16.4%, significantly outperforming the second-best entry’s 26.2% error rate. This milestone marked the beginning of rapid progress in the field of computer vision. Will we see the same for Devin?
You can’t spell Westphalia without A & I. Sovereign AI is becoming an important — and complex — question. Does it mean owning your semiconductor production? Or operating your own AI clouds? Or developing your own national foundational models? NVIDIA’s Jensen Huang argues that national strategies should extend to both the compute infrastructure and culturally relevant models. I tend to agree and know that many countries are turning to open-source LLMs to develop their own nationally sensitive models. Underlooked is South Korea and its efforts with HyperCLOVA X. And we can’t understate the sprint by Gulf monarchies, like UAE and Saudi Arabia, to acquire thousands of NVIDIA GPUs as part of their sovereign computing infrastructure. The UK has earmarked a meagre £100 million for a national AI supercomputing cluster. Perhaps the plan for Western nations is to rely on well-governed private sector efforts such as France’s iliad Group’s planned AI cloud. We are seeing a belated push by countries to reassert their sovereignty in the digital domain, something that few countries — bar China, North Korea and, more recently, Russia — have truly attempted to do2. I expect national AI plans to materialise in the shape of more national foundation models and infrastructure, but whether those will be economically competitive is somewhat unclear.
Accelerating discovery. In a groundbreaking achievement, researchers have used AI to identify a novel therapeutic molecule and advance it to early human trials in just 18 months. This process typically takes five to eight years using traditional drug discovery methods. The study employed two distinct machine learning models to streamline the drug discovery process. First, the PandaOmics platform analysed vast datasets to identify TNIK as a promising new target for treating idiopathic pulmonary fibrosis. Next, the Chemistry42 platform used genAI to design a potent and selective TNIK inhibitor with desirable drug-like properties. The AI-designed compound demonstrated anti-fibrotic activity across the lungs, kidneys and skin in animal models and exhibited a favourable safety profile in two Phase I clinical trials. In a separate development, researchers combined AI and supercomputing to identify a promising new solid electrolyte for batteries from over 32 million candidates in just 80 hours. They then synthesised and tested the material, creating a working prototype within six months. As we explored in a recent essay, AI’s potential to catalyse scientific breakthroughs is now bearing fruit, ushering in a new era of accelerated discovery and innovation.
See also: AI has been used to analyse over 30,000 policy documents from 1,035 school websites, demonstrating its ability to process vast amounts of data swiftly and cost-effectively.
Newsreel
The EU has struck a political deal on AI, passing the AI Act. For my two cents on what we must get right in AI governance and regulation, see my Three quick steps towards beneficial AI and The complexity of AI governance.
Figure, in collaboration with OpenAI, has released a demo of a humanoid robot actioning human verbal commands.
Covariant is creating a robotic foundation model.
Craig Wright is not Satoshi Nakamoto, according to the UK High Court.
Midjourney has deployed a new feature that allows you to create character consistency between multiple image generations.
GPT-4.5 details may have accidentally been leaked.
Market data
Workers in Asia have a significantly more positive outlook on the impact of genAI tools on workplace productivity compared to their Western counterparts. According to a recent survey, 67% of Indian workers believe that AI has enhanced overall productivity in their workplace over the past year. In contrast, only 17% of American workers share this positive sentiment.
Porsche ‘overwhelmed’ by 10,000 orders (not including China, Taiwan and Japan) for an all-electric version of one of its iconic models, the Macan.
11% of the finalists for the Pulitzer Prize in journalism confessed to using AI in the process of researching or writing their submissions.
The compute to train language models for a given performance benchmark has been decreasing by half roughly every eight months since 2012.
Americans lost roughly $1.3 billion in 2023 to scammers pretending to be from the government or tech support, according to new FBI data.
Cerebras has launched the world’s fastest AI chip, claiming it can train Meta’s Llama 2 70B in a single day (vs. the month it took Meta).
Short morsels to appear smart at dinner parties
✨ A 25-year bet between Christof Koch and David Chalmers has ended: researchers haven’t figured out how the brain achieves consciousness.
🥸 Vintage: In 1970, Marvin Minsky predicted we’d have “a machine with the general intelligence of an average human being” by the 80s. And, an inside story of Klatu, a fake android that made a splash in German media in the late-70s.
🏥 An Apple Vision Pro was used in surgery to assist with a procedure.
💨 Radia’s colossal cargo plane WindRunner aims to make large-scale wind turbines more accessible for onshore energy projects.
🎾 An AI ball machine for racquet sports can simulate gameplay.
🖥️ Spreadsheets are all you need.
Latest posts
If you’re not a subscriber, here’s what you missed recently:
End note
My favourite factoid in this week’s wondermissive is that solar panels are now cheaper than faux-wood plastic fence panels.
It’s a reminder that technology is a gift that keeps giving. I recently bought a full-duplex gigabit 5-port switch and a pair of clear acrylic containers to hold my vitamin bottles. They both cost £15 (about $20). Yet one is a remarkable piece of technology, and the other is some ho-hum storage boxes.
The digital tchotchke is capable of switching 10Gbps. That switch could handle all the monthly traffic of the global internet as recently as 1996. Unlike the acrylic boxes, the switch is a technology that benefits from economies of learning. And it’s a bit like magic. Fifteen quid!
Have a great week,
A
What you’re up to — community updates
Dr Rumman Chowdhury has been selected as one of four scientists to serve as new US Science Envoys in 2024.
Anna Bateson, CEO of the Guardian, has been recognised as ADWEEK’s Media Executive of the Year.
Share your updates with EV readers by telling us what you’re up to here.
A benchmark that tests AI’s ability to solve issues on open-source GitHub projects.
A good academic backgrounder is Four Internets by Kieron O’Hara and Wendy Hall.
The main event for AI this week was not in the mainstream news - but should be covered and discussed: https://www.coindesk.com/business/2024/03/27/three-decentralized-platforms-to-merge-ai-tokens-create-ai-alliance/#