🔮 Coordinating the energy transition; AI’s hiring bias; new molecules; urban recycling, teachers, and music taste ++ #473
Hi, I’m Azeem Azhar. In this week’s edition, we explore the coordination challenges of the energy transition.
And in the rest of today’s issue:
Need to know: LLM’s hiring bias
GPT-3.5 demonstrated stereotypical biases when evaluating resumes for various occupations.Today in data: AI-drug acceleration
AI-discovered drug molecules show promising 80-90% success rates in Phase I clinical trials, substantially higher than the historical industry average of 40-65%.Opinion: What the early internet boom can tell us about AI?
The AI revolution is following a similar trajectory to the early days of the internet, with infrastructure winners emerging first. However, the technology’s impact on incumbents and the emergence of new AI-native entrants remains to be seen.
💥 Today’s edition is brought to you by Sana, your new AI assistant for work.
Sunday chart: Time to bury fossil finance
The climate news this week was a mixed bag. Brazil’s devastating floods and scientists’ warnings of a 2.5°C warmer planet paint a grim picture of our potential future. Yet, amid the gloom, a bright spot is that renewables now generate 30% of our electricity globally, with solar leading the charge for the 19th year running. China’s solar manufacturing has even created an “overcapacity”.
This apparent paradox – an excess of solar panels in a world desperate for clean energy – is a result of a coordination failure in the energy transition. The failure to mobilise global finances to trigger demand for energy transition products is at the heart of this issue. Renewable energy projects require high upfront capital and low operating expenditure, necessitating innovative financing solutions. Asia’s excess savings could help overcome the upfront financial barrier to renewables for poorer nations. Encouragingly, Saudi Arabia and China are already exploring ways to support Africa’s energy transition. The key is for finances to bridge the gap between upfront costs and global savings. Countries are exploring options such as wealth taxes, shipping levies, corporate minimum taxes, and savvy funding structures to channel private capital into renewables.
Globally, the money for the energy transition is available. In 2022, fossil fuel subsidies reached $7 trillion worldwide – a staggering sum that has risen from $5 trillion in 2020 due to government support amid surging energy prices. Removing these subsidies and raising fuel prices to efficient levels could reduce global CO2 emissions by 43% below baseline levels in 2030, aligning with the Paris Agreement. Moreover, subsidy reform could generate $4.4 trillion (3.6% of global GDP) in 2030. Although politically challenging, a well-designed plan with clear objectives, stakeholder engagement, gradual price increases, targeted support for low-income households, and depoliticised energy pricing, can help countries successfully redirect these funds.
The money saved from ending fossil fuel subsidies is likely to be more than enough for the energy transition. Estimates for the annual investment needed by 2030 range from $4 trillion (IEA) to $9 trillion (Climate Policy Institute), but we think the actual figure could be lower due to rapidly falling prices and learning rates.1
Progress is underway, but governments must ensure their industrial policy pledges don’t become mere rhetorical smokescreens for protectionism. Global cooperation and coordination is needed to solve climate change – not geopolitical finger-pointing.
🚀 Today’s edition is supported by Sana.
Sana AI is a knowledge assistant that helps you work faster and smarter.
You can use it for everything from analysing documents and drafting reports to finding information and automating repetitive tasks.
Integrated with your apps, capable of understanding meetings and completing tasks in other tools, Sana AI won’t just change the way you access knowledge. It’ll change the way you work.
Key reads
TikTok’s on the clock. The TikTok divestment debate revolves around the question of who should control the media ecosystem. Before the internet, media outlets operated within national boundaries, closely regulated by clear ownership and content standards. However, the internet allowed the media to transcend borders. The US barely batted an eyelid, given that its firms dominated the landscape. And China, long recognising the risk of foreign-owned media platforms, erected the great firewall nearly 30 years ago.
With TikTok being a Chinese-owned company, the US faces a dilemma: is it prudent to allow a foreign, illiberal country to have control over a significant portion of the US media ecosystem? Particularly given the Chinese Communist Party’s potential influence over TikTok. The platform’s lawsuit against the US government, which argues it can’t divest from China due to CCP opposition, highlights the party’s sway. Bytedance’s Chinese leadership exerts increasing influence on TikTok’s operations; often, Chinese managers need to approve key decisions. A whistleblower also recently claimed to have sent vast amounts of US user data back to Beijing. While US entities have invested in TikTok, this doesn’t negate CCP influence. The Chinese big tech crackdown from 2020-22 is a testament to this. There is also an even simpler rationale: reciprocity. TikTok can operate in the US in a way Google or Facebook are not allowed to in China. As Chinese philosopher Confucius said, “Never impose on others what you would not choose for yourself.”
See also: Xi Jinping’s plan to save China’s economy through science. It seems he is well aware of the exponential age – an excellent read.
LLMs vs journalism. The media industry is grappling with the implications of generative AI, as some choose to litigate while others embrace partnerships. In a recent development, eight local newspapers filed a lawsuit against OpenAI, even as several news organisations, including the Financial Times, entered into content agreements. France’s regulatory authorities have taken a decisive step, imposing a €250 million fine on Google for training its AI models on copyrighted material in March. While copyright holders may get something out of this, the more pressing question is how media creators and journalists will be compensated in an era where information is accessible with zero clicks. Competing licensing regimes, such as a fixed sum per retrieval from a retrieval-augmented generation (RAG) system, could emerge as potential solutions, similar to book or streaming royalties. The outcome of these developments is crucial, as journalism, already challenged by the rise of digital platforms, now faces further revenue erosion from LLMs. As model developers like OpenAI selectively partner with outlets through “preferred publisher programs”, government intervention may be necessary to ensure a level playing field to protect journalism.
How do you test a superpower? Researchers investigated the potential biases of LLMs in hiring. They used GPT-3.5 to conduct two experiments on names with varied race and gender connotations for resume assessment and generation. The findings were concerning. GPT-3.5 showed stereotypical biases when scoring resumes across occupations and evaluation tasks. The generation study further revealed underlying biases, with women’s resumes featuring less experienced occupations and Asian and Hispanic resumes containing immigrant markers. Although OpenAI evaluated GPT-3’s propensity for bias, which revealed varying sentiment responses across ethnicities, the company is yet to adequately address the downstream effects of its LLM on societal applications. Similarly, Google’s AlphaFold 3, released this week, faces evaluation hurdles. EV member Bharath Ramsundar underscores the difficulty in predicting its performance on out-of-distribution data, a task made increasingly complex to evaluate by the likely vast size of the training dataset. Both cases highlight the lack of comprehensive evaluation – or at least a willingness to release models without thoroughly addressing bias and limitations.
Newsreel
The US has revoked Huawei’s licences to purchase semiconductors from American tech giants Qualcomm and Intel, further escalating trade tensions with the Chinese telecommunications company.
Prof. Fei-Fei Li is building a startup focused on “spatial intelligence” that aims to enable AI to reason about the 3D world.
Apple is developing its own AI chips for data centres, code-named Project ACDC, in a bid to gain an edge in the intensifying AI arms race. They have also unveiled their new M4 chips this week.
Microsoft deploys the first genAI model fully isolated from the internet, enabling US spy agencies to securely analyse top-secret data. They also appear to be hedging their investment on OpenAI, developing their own internal LLM, called MAI-1, led by Mustafa Suleyman.
Data
Long Covid is estimated to have reduced labour supply by 0.3-0.5% in the EU in 2022.
In 2019, fixed-tilt solar plants saw a 52% increase in the median power density of their solar panels compared to 2011.
Global defence spending, adjusted for inflation, rose by 6.8% from 2022 – the largest year-on-year increase since 2009.
AI-discovered drug molecules show promising 80-90% success rates in Phase I clinical trials, substantially higher than the historical industry average of 40-65%.
Microsoft research claims 75% of knowledge workers are using genAI today.
More than 20% of voters in swing states consider crypto a key issue in the US elections.
Short morsels to appear smart at dinner parties
📚 Experiments have shown that teachers cannot accurately distinguish between ChatGPT and student-produced text.
🏙️ Urban recycling: empty office towers of Manhattan are getting a new life as apartments.
🚗 Caregiving-related travel has negative effects on the wellbeing of female caregivers in car-centric cities.
🌍 Infrastructure networks are remaking geopolitics – good read in Foreign Affairs.
🪖 The war in Ukraine may be remembered as the one that kickstarted a new generation of drones.
🎼 Research suggests that our musical preferences tend to stagnate around age 30.
End note
I had a great time visiting Boston and New York this week. I spent time at Harvard Business School, where the topic was generative AI. It was remarkable how many HBS faculty members are leaning into genAI, with multiple robust research programmes trying to understand its impact on workforce, productivity, learning and development and more. It is more than remarkable and encouraging. These are powerful tools, like chainsaws, photocopiers, typewriters and conveyor belts. And, like powerful tools, managers need to know how to use them, how to adapt their team structures and how to ensure they are applied safely. Perhaps genAI tools are even more powerful than the humble fax machine. (I suspect they are.)
One thing that was clear — and compounded to me by a very senior Chinese AI scientist later in the week — is that LLMs, in various forms, are going to be a dominant part of enterprise architecture for the next several years.
From a scientific perspective, LLMs aren’t the last word in computer science, cybernetics, machine learning, or AI research. There will, of course, be new approaches that will introduce new capabilities for us to play around with in coming years.
🎧 Finally, I recently spoke with the super-smart
of . It was a fun conversation that moved beyond the discussion of the exponential age to tackle questions around the learning economy, the nature of increasingly complex knowledge and how the sociology of academia could be hindering interdisciplinary breakthroughs.Enjoy the discussion, link below.
A
What you’re up to – community updates
Alex Kendall’s company Wayve raised $1 billion, Europe’s largest AI deal to date.
John Thornhill wrote an opinion in the FT about the resurgence of America’s innovation engine.
Dr Jon Roberts’s company has completed a licensing partnership with OpenAI.
Simon Vannuccini and co-authors have published a study analysing the nexus between robot adoption and product innovation at firm level. They find that following robotisation, firms innovate less.
Rich Walker’s team has announced the launch of a new class of robot designed for machine learning applications, developed in collaboration with Google DeepMind.
Share your updates with EV readers by telling us what you’re up to here.
To be clear, we think there is a possibility that the dollar-sum for energy investment required for net zero will be lower than IEA estimates, but investment levels may remain high because we’ll invest in expanding the energy system.
Love the mention energy, how it relates to climate change.
Unfortunately, the fossil fuel subsidy has many dimensions and its roots run deep, making it difficult to make meaningful progress. From OPEC to oil drilling, gas stations, and trucks, many individuals depend on the #fossilfuel industry for their livelihoods. And as the saying goes - “if your livelihood depends on not understanding something, you won’t”
That's why instead of demonizing those who work in fossil fuel industry, we need to understand what’s their challenge. We need to set our differences aside and work together to find a path forward. We have no time to waste. I'm asking this community for help.
If you or someone you know is working towards creating economic opportunities for those in the fossil fuel industry, please connect with me. I'm eager to offer my problem-solving and communication skills to accelerate progress towards a green future.
Although I'm not an expert in sustainability, I'm committed to understanding the problem better. If you're a sustainability expert and would like to help me think through the problem, please reach out.