🔮 AI & creativity; exponential compute; peak GHGs, recycling concrete & marriage saves lives ++ #476

Hi, I’m Azeem Azhar. In this week’s edition, we explore AI’s capacity for divergent thinking and how this can help the scientific process. 

And in the rest of today’s issue:

1. Need to know: Modest gains

   Daron Acemoglu’s latest research looks at the next decade of AI’s macroeconomic impact.

2. Today in data: Peak emissions

   BloombergNEF estimates that GHG emissions this year could fall by 2.5%.

3. Opinion: The next five years of AI

   I spoke on this topic at the UN’s AI for Good Global Summit. Watch here

---

⚡️Today’s edition is supported by our sponsor Prolific, quality data for AI research that reflects the breadth of humanity.

---

Sunday chart: Thinking outside the black box

Back in February this year we published a two-part deep dive for members of Exponential View on the ways AI is changing science and accelerating discovery. Two new pieces of research this week made me reflect back on it (if you haven’t read it, I highly recommend you catch up this weekend). 

First, a new research paper offers an in-depth analysis of divergent creativity based on state-of-the-art LLMs and a dataset of 100,000 humans. Divergent thinking is one of the pillars of creativity and innovation – it allows us to solve problems, find new solutions and think outside the box. The research, co-authored by Yoshua Bengio (listen to my discussion with Yoshua here), finds that GPT-4 outperforms humans on verbal creativity tasks that use divergent thinking, while GeminiPro is on par with humans. Interestingly, one of the smaller open-source models with 13bn parameters outperformed GPT-4-turbo and GPT-3. Divergent thinking is pivotal in science, and we believe that AI’s divergent capabilities will make for a powerful partner in scientific discovery and innovation.

Second, a report by the Royal Society maps out evidence, challenges and opportunities for AI’s use in science. They identified key roles for AI in scientific processes:

• First, providing information through advanced simulation and data representation that cannot be obtained through experimentation.

• Second, expanding the scope of human imagination or creativity by unlocking surprises in data and modelling.

• Lastly, translating complex information and observations into new knowledge. 

Expand this to other industries, and we’ll be equipped to use the scientific method and divergent creativity with our AI co-pilots in any domain. 

---

Today’s edition is supported by Prolific. 

Leading AI researchers use Prolific to gather high-quality human data for AI models at scale.

Instantly access 150k+ vetted participants and domain specialists, ready to respond to your project's needs. Our self-serve platform makes it easy to gather text, voice, image, or video data from diverse and engaged participants – in hours.

Make your models more versatile and robust with data that reflects humanity.

---

Key reads

Macroeconomics of AI. Earlier in the week at the AI for Good Global Summit, I made the point that the impact of AI can be seen everywhere but in the productivity statistics. Now, economist Daron Acemoglu has come out with a sobering assessment of the macroeconomics of AI. Acemoglu estimates that total factor productivity (TFP) gains over the next decade from AI would be modest at 0.66% at the higher end. To put this into perspective, during the “Golden Age” of economic growth between 1950 and 1973, the average annual rate of TFP growth was 0.77% in the US, 0.73% in the UK, and 1.79% in France. The crux of AI improving the macroeconomic environment in the near future is in new kinds of services, products and value created. However, if new tasks and outputs are manipulative, addictive and misleading (“bad tasks”), we’ll see a net negative impact despite the potential increase in TFP and GDP. In the last year alone, AI-generated images have become nearly as common as text manipulation and misinformation. This is exactly what Acemoglu is warning against as “bad” or counterproductive use of AI.

• See also: A review of AI legal research tools. 

The frontier’s exponential. The reason economic forecasts of AI impact vary so widely is that the rate of growth and the resulting advancements in AI capabilities are difficult to predict accurately. An analysis of 300 machine learning systems shows that the amount of compute used in training is growing at four to five times per year – given the scaling laws, AI performance is likely to follow. What are the implications of this growth? Google’s former boss Eric Schmidt expects that we’ll have autonomous agentic systems working for us in the background within the next five years as we get across three capability ladders: an infinite context window, enhanced chain-of-thought reasoning and text-to-action programming. 

Clouded judgement. One of the great failures of the UK government in the last decade has been its misunderstanding of a key ally, the US, argues FT’s Janan Ganesh in this excellent essay. This cost us EU membership and jeopardised the UK’s critical trade alliances. But what if the UK’s not the only Western power guilty of this? What if a mix of strong fear, pride and ignorance is skewing decision-making in the West in relation to China? It’s worth asking this question no matter how uncomfortable. Singapore’s former minister of foreign affairs, George Yeo, addressed this question head on last week. From his point of view (23 years in government), the US is seeing the rivalry with China through the lens of American exceptionalism and its current approach to outcompete will backfire, weakening the US and its allies. It’s worth watching the interview.

See also: 

• Activist investors are not happy with Texas Instruments’ decision to bring semiconductor manufacturing back in house. 

• The US is delaying AI chip exports to the Middle East over China concerns. 

---

Data

BNEF is projecting a 2.5% fall in GHG emissions globally in 2024. 

India also added 10GW of solar in Q1 – that’s 65% of the UK’s entire solar capacity.

Ransomware attacks at hospitals lead to a 20.7% relative increase in in-hospital mortality. Elsewhere, some data suggests that the economic impact of major cyber incidents appears to be falling.

Prospective data centres in the US wait as long as seven years to get the supply of power they need.

47% of Americans say they have never heard of ChatGPT.

Legumes have a high return on investment in terms of protein per dollar spent: $0.50-1 for 20-30g of protein.

---

Short morsels to appear smart at dinner parties

🫣 Former board member at OpenAI Helen Toner says the board learnt about ChatGPT’s release on Twitter.

🥵 Temperatures in Delhi hit 52°C. 

🚨 Over 2,500 documents showing how Google stores data on content, links and user interactions have been leaked.

🤔 A new study suggests the survival benefit of marriage in patients with certain types of cancer is larger than the published survival benefit of chemotherapy.

🚧 Researchers find a new way to recycle concrete. This could be significant as concrete produces 7.5% of anthropogenic CO2 emissions.

🦠🚫 A new antibiotic that targets gram-negative bacteria in mice is effective against 130 drug-resistant bacteria, without harming the gut microbiome.

⚡ There may be a simple fix for renewable projects waiting to be plugged into the grid in the US: installing new wires on existing high-voltage lines. This alone could double the size of the grid.

---

End note

While in Geneva this week speaking at the AI for Good Summit, I had a chance to try on an AI-powered exoskeleton designed for everyday use.

It was absolutely incredible, and it made me instantly see the value of this technology for people with mobility issues, such as those over 75. One needs to balance the new mobility freedoms against muscle atrophy, but I can see it having its uses.

Until next week, Azeem.

---

What you’re up to – community updates

• Robert Garlick releases “What Machines Can’t Master”, a 108-page report on skills in the AI age to which I contributed.

• Alice Casiraghi evaluates the circularity of a food system. 

• Abhishek Gupta looks at the Copilot+ PC frontier. 

• Denis Piffaretti and his team published a textbook on building LLMs for production.

Share your updates with EV readers by telling us what you’re up to here.