What's going to happen next?

Highlights from Manifest, the SF Bay Area 'Woodstock for Futurists'

A big thank you to the reader who reached out to offer me comp tickets to ‘Manifest’, a three-day conference on forecasting, emerging technology, and the shape of the 21st Century. Imagine 600 hypernerds, including many of the world’s most successful ‘superforecasters’, gathering in a fabulous venue under a warm Californian sun to talk about science and prediction markets, play poker, scout for investment opportunities, and party (hard).

Many attendees were at the very cutting edge of their respective fields and were deeply involved in the latest biomedical and AI research, global policy, and economics. One session I attended even included a researcher being roped in on Zoom to discuss the implications of an advancement published two days earlier.

In short: Manifest was the place where glimpses of the future could be most clearly seen. Below are my notes on what I saw:

Manifest group photo by Misha Gurevich

Demographic Collapse
A worldwide fertility collapse is already well underway. Many are aware that China’s famous ‘one-child policy’ led to a dramatic reduction in birthrates, and is responsible for a rapidly aging population. Fewer are aware that most of the world is having fewer children than China did during the height of this policy. In countries throughout Europe, the Americas, most of Asia including the likes of India and Thailand, fertility is already well below replacement and continuing to fall fast. Post-Covid, young people are entering relationships far less often, marriages are way down, and with most people choosing to have much smaller families than in days gone by, expect each generation to be 30-50% smaller than the previous.

The implications of these trends are wide-reaching. Unlike Japan, which has been able to compensate for its aging population by importing young people¹, the same will not be possible in 30 years when every country is undergoing the same problem. Economies will falter. Voting demographics will shift further and further towards the elderly, who, as we’re seeing in the UK and elsewhere, are incentivised to vote to allocate ever-increasing proportions of tax revenue from a shrinking base of tax-payers. Even with advances in robotics, expect inter-state competition for young people to intensify, and wages for service and care workers to rise as demand begins to outstrip supply. The big unknown here is whether technological progress in robotics and AI can make the fundamental progress required to offset these looming shortages without being banned or otherwise regulated into oblivion. In any case, expect these trends to have severe social impacts on individuals and communities as they continue to play out.

Taiwan
Manifest was generally sanguine about the pre-2030 risk of an invasion of Taiwan. The Manifold prediction market puts the risk at 29%, but various shows of hands throughout the event put Manifest attendees closer to ~5%. One key point of evidence was the tendency of China to continue to invest in seabourne trade of key commodities like oil and gas, rather than expand overland links to Russia and Iran. Seabourne trade is likely to be cut off in the case of a major war, so the fact that China is not hedging against this possibility suggests they’re betting on continued peace. Notably, several prominent individuals with close ties to the seniormost levels of the Chinese government confided in me that they considered the risk of a major conflict to be remote, unless Chinese leadership was put into the position of having to respond to the broad public perception of a major US provocation.

In my view, what most discussion of war risk tends to miss is the tiny Kinmen Islands, just a few miles off the Mainland Coast of China. Kinmen is heavily integrated with the Mainland, even reliant on continued drinking water supply. Chinese officials tend to see the islands as strategically irrelevant to everybody but China, only of concern to Taiwanese national pride. Securing Kinmen would allow President Xi something of a face-saving credible fulfilment of steps towards near-term reunification. Kinmen would likely be taken using the ‘grey zone’ tactics they’ve been practicing in the South China Sea, particularly against the Phillippines, which successfully delivered Scarborough Shoal into defacto Chinese sovereignty. The case of Kinmen however is significantly complicated by the presence of US special forces on the island. Conflict is inherently chaotic, and if an incident resulted in the deaths of these individuals, I don’t believe anyone could answer definitively how the American public would respond, or whether significant unintended escalation could be avoided (especially if an unintended firefight resulted in deaths of military personnel on both sides).

Unfortunately it’s hard to see anything but a continuation of low-level I can only hope cooler heads prevail, and that a diplomatic solution can be reached. I do not see how anything risking escalation into a full-scale war can be in anyone’s interest.

Genome sequencing
The first mapping of the human genome took an international consortium of researchers and billions of dollars. Today, having your entire genome sequenced costs around $200USD. We’re still only scratching the surface of which phenotypes correspond to which genes, their interaction with different drugs, cancer risk markers, etc. Sequencing your genome will therefore yield an asset the value of which will only grow throughout your life, teaching you which drugs to avoid, which cancers to get early screening for, and so on. Highly recommended.

USDT / Tether Coin
Tether is a crypto ‘stablecoin’, with the value of each token pegged to the value of the US dollar. Stablecoins are therefore unexposed to the usual crypto volatility, and are useful for transferring money overseas where an expensive and lengthy wire transfer process might be one’s only other option. Tether enjoys a massive first-mover advantage and is the highest-traded cryptocurrency by volume. Knowledgeable Manifest attendees outlined how illegal activities by Tether are visible on the blockchain, and highly recommended exchanging any Tether assets you have to more reputable and regulated stablecoins such as USDC instead.

Genetics
The dream of our parents, ancestors, is for their children to be better versions of themselves. Stronger, faster, smarter, healthier, able to surpass us in our passions. For almost all of history, our ability to influence the genetic makeup of our children has essentially been limited to our choice of partner. Otherwise, genetics are determined by the multi-generational influence of natural selection - which only loosely aligns with our values of a long, happy, active life².

In recent decades, widespread ante-natal testing has allowed us to screen for trisomies like Down syndrome, but is limited in its ability to do more than put parents in an awful position. IFV treatments now routinely use genetic testing and embryo selection to avoid known genetic diseases, but the efficacy of this scales extremely poorly (with the sqrt(ln(tested embryos) - perhaps the worst scaling law I’ve ever heard). Recent advances however suggest that an iterative approach to genetic recombination may soon be possible, enabling parents to take a much more active role in determining which genes to pass along to their children.

Given the lack of progress in cryonics, and the seeming impossibility of faster-than-light travel, I predict if ‘human’ interstellar travel is achieved, it will first be via uncrewed ships packed with our collective genetic material, the colonists incubated in artificial wombs using the techniques pioneered above. In any case, the future is going to be weird. Hopefully our descendants fare well.

AI ‘Covid moment’
In January 2020, Covid-19 first started to hit the press - a virulent new flu coming out of China. A reader asked me about it at the time. Briefly reflecting on the base rates for a new virus becoming a serious global concern sufficient to trouble us in New Zealand (1918-19 Spanish Flu), I suggested it was unlikely to be a problem for us.

I didn’t think much of it again until mid-February, when global financial markets started getting the jitters on the back of increasingly concerning news out of places like Italy, which had begun to register their first deaths, and some worrying exponential trends. On the 27th of Feb, I read Manifest-adjacent @yashkaf’s ‘Seeing the Smoke’, which was the kick I needed to galvanise action. I couldn’t yet see the full shape of things to come, but it was clear this was going to be a big deal, and we could expect considerable turmoil. I sold my shares and began to prepare as best I could, warning our civil defence team, began training staff in remote working, and purchased headphones and other equipment so that Council could continue operating during the lockdowns that were surely coming. I was relatively junior at the time, and had a hard time of it. Many people just laughed when I brought it up - even as late as mid-March, with major stock market indexes down 20% or more, and severe lockdowns beginning to sweep through Europe. Most were aware of covid by this point, but given how unprecedented it was, struggled to see how game changing it would be until, finally in late-March, covid was suddenly the only game in town.

It’s a good model for how major social changes propagate through society. First (in December), it’s of academic interest only. Then, with spread, some minor news stories. People close to the field begin to take serious interest. A breakthrough of some kind occurs (e.g. an outbreak in Italy), demonstrating to those paying attention that prior/mundane models are obsolete. Then, there’s a cascading effect as markets and generalists catch on and slowly spread awareness, until finally the impact is too large to ignore, and society as a whole flips from generally ignoring the thing to paying too much attention to it.

This is the current model of AI awareness. Decades ago, machine learning was of niche/academic interest only. Sometime later, I remember watching Alpha Go beat Lee Sedol in 2016, (and the considerably more complex feat of the OpenAI 5 beating OG in 2019), and the hairs on the back of my neck sticking up, realising that this technology would have far-reaching implications. GPT-2, an early ‘large language model’ (LLM) was released in 2019, and behaved as a sort of child-like schizophrenic poet. Fast forward to 2022, and GPT-3 was released as ‘ChatGPT’, causing a major stir as people enjoyed a high-school level competency chat bot how to make lasanga and LinkedIn-tier ‘10 tips’ lists. The latest models, like GPT-4, or Anthropic’s Claude, now routinely outperform university undergraduates across most academic fields, and can interact via text/video/audio, generate images, call code, and translate between all major languages.

Nvidia, a graphics card design company whose chips turned out to be fundamental to AI, is poised to overtake Apple as the world’s most valuable company. Even the market price of power companies are being bid up as data centre electricity demand skyrockets. Financial markets are betting big on AI. The capabilities of today are already transformative³, and even if fundamental advances plateau not far beyond here (and I think it’d be foolish to think they will⁴), the results will soon be too enormous for policymakers, the media, and everyday folks to ignore. A ‘covid moment’, where AI shifts from being of humourous or peripheral concern, to the only topic in town, is coming fast.

Electricity Demand
AI has historically been a tiny fraction of overall data centre power requirements (primarily storage and traditional computing). Data centres have been a major growth industry for years, but their power requirements on-net have been relatively low since their growth has partly come from the shutdown of small-scale inefficient business servers, chugging away in some corner of the building with inefficient climate control, etc. Thus, despite enormous data centre growth over the past decade or so, the electricity requirements have easily managed - US electricity demand has remained ~level since 2009, despite a growing population.

AI training (developing the model) and inference (using the model, e.g. running ChatGPT) demand is currently growing at around 280% per year⁵. Should this trend continue, electricity demand will soon accelerate rapidly. Goldman Sachs recently estimated that by 2030, additional data centre-driven demand in the US will be in the order of 400TWh, around 10% of current US demand - requiring an additional ~45GW of power generation. However, their modelling assumes AI demand will only grow by a sedate 20% per year - essentially a bet that no further advances in AI will be made, and growth will come from the slow integration of current technology into the global economy.

With large economic incentives, I think we can probably manage 10% growth over the next 6 years. However, assuming AI progress continues at anything like the current rate, demand will soon outstrip supply. Our electrical grids are almost completely unprepared for this, especially given that consenting and construction of new power plants can take decades and our AI covid moment has yet to arrive. Stable jurisdictions with good internet connectivity and a streamlined permitting process for new power plants are likely to benefit handsomely by enabling data centre construction and selling that capacity to their peers.

On a somewhat personal note, the conference left me highly optimistic about the future. My sense is that concerns about AI are abating, and shifting more towards mundane geopolitical concerns about a peer competition with China that is unlikely to lead to major conflict as long as people can keep their cool. I believe that to the extent that people do their duty, the future will be as good as any of us can hope for. History cannot be changed, but it has led us all here.

1. The number of foreign workers in Japan has almost quadrupled since 2008, but even this is insufficient to compensate for generations of low birthrates. Japan’s population is already falling, down 4m from its 2007 peak of 128m, and is expected to halve by 2100. The impact of aging demographics is already evident. The ‘National Museum of Emerging Science and Innovation’ covers the achievements of decades past, lacking young curators excited about sharing the cutting edge. Tokyo’s tokenistic playgrounds sit empty. Entire towns are without children, dioramas of dolls carefully arranged in old classrooms to remind the aging population of a bygone era.

2. In a world increasingly devoid of strong intergenerational familial ties there is a strong selection pressure towards people who live healthy productive lives, then drop dead upon retirement so their children can inherit their wealth and purchase housing to have a family of their own. This is so, even if those people are otherwise valued members of their community who help run the local playgroup and reforestation project. Mother Nature is ‘red in tooth and claw’, and has zero consideration for values like human decency or stable societies, except as far as these things improve the genetic fitness of a given lineage. Thus, I consider the moral case for measures taken to improve human health, longevity, and capabilities, to be extremely straightforward.

3. For example, call centres are already obsolete. Multi-model GPT-4 level models, when properly trained and used, have better performance at lower cost, and are indistinguishable from a real person for human callers.

4. If nothing else, ever cheaper computing power will allow advances in frontier models that were previously uneconomical. Robotics is undergoing a renaissance at the moment (you may have seen Boston Dynamics/Tesla/various Chinese firm’s latest humanoid robots) and even the integration of well-established AI models and techniques with these will lead to economic transformation.

5. From Nvidia's quarterly earnings report ‘data centre’ revenue segment. Nvidia supplies the overwhelming majority of chips used in AI training and inference.