Ideas of the week
LFP: Itās as easy as 1-2-3. Lithium battery pack prices in China have fallen 37% in the past year to $75/kWh, driven by manufacturer margin compression, overcapacity, lower material costs1 and, of course, technological advances. This represents a 93% decline since 2010, when prices were $1,183/kWh. Research provider Bloomberg New Energy Finance (BNEF) expects these conditions to keep prices low for years to come. Iād go one step further and say they will continue to fall, and fall fast. The falling price of batteries will accelerate the uptake of electric vehicles and stationary energy storage. This creates a virtuous circle ā increased adoption leads to more production experience, which leads to further cost reductions and technological improvements. The cycle is likely to continue, with future technologies such as solid-state and sodium-ion batteries promising further advances. If current trends continue, battery prices could reach $23/kWh by 2030. Even if the price decline slows to half the recent rate, prices in 2030 would still be around 50% of todayās levels. For a glimpse of what this future could look like, see our essay on the dawn of the distributed energy age.
See also: The UKās largest retail energy provider, Octopus Energy (a startup and tech unicorn), has partnered with Chinaās BYD to bring vehicle-to-grid storage to its customers.
Flexible safeguards. In his most recent think piece, Yoshua Bengio, a pioneer in deep learning, reasons through the arguments against taking AI safety seriously, calling for increased investment in security and responsible development. While we disagree on some claims ā including the human inability to control advanced AI, or the inevitability of rapid progress ā we endorse his key argument: AI governance requires political and social solutions, not just technical ones. We support Bengioās call for principles-based legislation and international agreements, an approach in line with UK regulatory philosophy and one that is well suited to the current uncertainty of AI development. Unlike the rules-based systems favoured in the US, China and much of the EU, this governance model focuses on outcomes rather than specific mechanisms. It offers flexibility and adaptability to unforeseen scenarios, and encourages innovation while promoting accountability. Crucially, it avoids the pitfall of simply āfollowing the rulesā without considering the wider implications. For more on Bengioās ideas, see my conversation with him from earlier this year.
See also:
- A 10-year retrospective on philosopher Nick Bostromās āSuperintelligenceā.
- Hackers accessed OpenAIās internal comms forum in a security breach last year.
Nanometer Nanny State. A recent study examining 58 industrial policies across 17 countries reveals the size of government support for the semiconductor industry. Subsidies have been enormous, some $330 billion in the period of 2010 to 2022, or about 7% of global industry revenues. The pattern of this support varies. The US tends to focus on supporting research and development, rather than production subsidies. And, in the early years of the industry, the government was the buyer of first resort. In 1965, the US military bought 72% of all chips.2 More recent interventions, in Korea, India and China tend to focus on production subsidies, something the US has also emphasised recently with the 2022 Chips Act. Notably, outside the US, no domestic semiconductor industry emerged without substantial foreign technology facilitated by government policies. The industryās need for large investments, rapid technological advances and complex supply chains make it ideal for government support. The irony in all this is, of course, the ahistorical attitude towards the government held by much of Silicon Valley ā which only exists because of substantial government intervention. The reality is that Silicon Valley exists not because of the government alone or entrepreneurs alone, but both. What economist Mariana Mazzucato calls the entrepreneurial state or venture capitalist Bill Janeway refers to as the three-player game. Competitive realpolitik will trump intellectual dogma over the coming yearsā¦ expect more industrial nannying policies.
Share
Data
The time price of a roundtrip New York-London flight has decreased by 90.8% since 1970, from 140 hours of work to just 12.9 hours in 2024.
Political science professors Leonardo Baccini and Thomas Sattler found that in regions with a high proportion of low-skilled workers, implementing austerity measures of average intensity led to a 4.6% increase in populism, while regions experiencing minimal or no austerity saw only a 0.2% increase.
1PointFive, a carbon capture, utilisation and sequestration (CCUS) company, has secured a landmark deal with Microsoft to sell 500,000 metric tons of direct air capture-enabled carbon dioxide removal credits over six years, marking the largest such purchase to date.
US EV sales grew 11.3% year-over-year in Q2 2024 ā good for the market, but not for pioneer Tesla whose market share fell below 50% for the first time, dropping to 49.7% from 59.3% a year earlier.
99 hydrogen fuel cell cars were sold in the US in Q2 ā a 91% year-on-year decline.
OpenAIās next computing cluster will have 100,000 Nvidia GB200s (list price $70k a pop, notional sticker price for chips $7 billion).
Short morsels to appear smart at dinner parties
š The global economyās continued reliance on undersea cables makes it vulnerable, but space systems like Space X can boost resilience.
š» Silicon Valley mogul and Andreessen Horowitz co-founder Marc Andreesen issued a $50k āgrantā to an AI agent.
š Maths plays a key role in helping US top swimmers win Olympic gold.
š® A new study finds that taking fish supplements may increase the risk of heart problems in healthy people.
š„² The academic case for calling chatbot hallucinations ābullshitā.
š§ How to raise the worldās IQ.
In focus
Hereās a preview of my Saturday column. I dive into the $100 billion question ā can AI scaling laws hold? And what could the next five years bring?
š§ AIās $100bn question: The scaling ceiling
Azeem Azhar
Ā·
Jul 13
By 2027, a single artificial intelligence model could cost $100 billion to train ā a figure bigger than the GDP of two-thirds of the worldās countries. This staggering projection illustrates the breakneck speed at which AI development is scaling up, both in size and cost.
Read full story
1
Lithium iron phosphate (LiFePO4) batteries, commonly known as LFP batteries, utilize a cathode material that is more cost-effective compared to traditional lithium-ion chemistries such as lithium cobalt oxide (LiCoO2) or lithium nickel manganese cobalt oxide (NMC).
2
Interestingly, the first complete microprocessor is often considered the Intel 4004. But there is a strong case it was the F-14ās Central Air Data Computer. This compelling video makes the case. (Also, I have put together a 90-minute set as a homage to the F-14.)