Investment Manager’s Report: Summary

Market Review

As discussed in our last Annual Report, we believe 2022 is best understood as the year ‘risk was repriced’ as central banks moved forcefully to rein in the economy, defend their credibility and prevent inflation expectations becoming unanchored. Proving anything but ‘transitory’, inflation continued to surprise to the upside taking global risk-free rates with it. In the US, consumer price inflation (CPI) averaged 8.0% during the calendar year, while the +9.1% reading in June was the largest year-on-year (y/y) monthly gain since 1981. The inflation shock was hardly unique to the US, with soaring energy and food prices, labour markets with more jobs than available workers and the release of pent-up demand combining to create the most inflationary backdrop globally for 40 years. For the full year, global inflation averaged 8.8% compared to pre‑pandemic levels of around 3.5%.

As a result of this persistent inflation, 2022 was also a year of unprecedented interest rate rises, after an oddly slow start by central banks. In the US, the Federal Reserve (Fed; the US central bank) embarked on the steepest set of rate hikes in 40 years as rates were raised by 450 basis points (bps), including four 75bps hikes, in addition to the resumption of quantitative tightening (QT) whereby the Fed reduces its monetary reserves to ‘tighten’ its balance sheet. Futures markets at the start of 2022 had priced in expectations for Fed Funds (the key benchmark rate targeted by the Fed) to be at c1% by June 2023; by year end, this figure had risen to c5%. In Europe, the decade‑long experiment with negative interest rates ended as the European Central Bank (ECB) raised rates by 250bps despite a high likelihood of recession. Most other major markets experienced tightening in excess of 200bps.

Sharply higher risk-free rates weighed heavily on asset prices, not least bonds which experienced their worst calendar year returns since at least the 1970s, the Bloomberg US Aggregate Float-Adjusted Index losing 13.1%. This theme was painfully echoed in equity markets – the longer the duration, the worse the return. Ten-year US Treasuries suffered their worst annual performance since 1788 while record government bond losses were recorded in Japan, Europe, and the UK with drawdowns of 16.2%, 22% and c32% respectively. Having stood at $10trn in January 2022, the global stock of negativeyielding bonds had fallen to essentially zero by calendar year end.

Higher sovereign yields weighed heavily on global equities, which also had to contend with elevated recession risk and negative earnings revisions. During the calendar year, 2yr-10yr Treasury yields fell to their most negative spread (where 2-year yields are higher than 10-year yields) in more than 40 years. Aggregate earnings estimates for companies in the S&P 500 Index in 2023 fell from $245 to around $230, while 2024 forecasts fell to c$250, essentially losing a year of growth. As measured by the MSCI All-Country World Index (ACWI), global equities fell by -18.4%, in dollar terms, their worst showing since 2008. The S&P 500 Index (-19.4%) also posted its biggest fall since 2008 and its seventh worst year since 1926. The unusual correlation between bond and equity markets, courtesy of inflation, meant that 2022 will probably be remembered for being the first year that both the S&P 500 (equities) and 10-year US Treasuries (bonds) each registered losses of more than 10% on a total return basis. It was also the worst year for combined total returns of stocks and bonds since 1982.

A bad year for US equities proved a calamity for growth stocks which suffered their worst year compared to value stocks since 2000. Helped by energy’s record year (+59%) versus the broader market, the Morningstar US Value Index fell just c1% while the Morningstar US Growth Index plunged by c37%.

Equities started strongly in 2023 as extreme pessimism and bearish positioning were challenged by disinflationary data, weaker energy prices and sharply lower real rates, as well as a better than feared Q4 company earnings season and a momentum / short squeeze. European equities and 60/40 portfolios recorded their best start to a year since at least 1987, while the tech-heavy NASDAQ Composite Index enjoyed its strongest year-to-date performance since 2001.

However, sentiment turned more negative in February as a slew of strong economic data for January challenged the excitement that the interest rate tightening cycle was largely complete. Investment grade global bond markets gave back their year-to-date gains, while corresponding equity market weakness has seen US indices either approach or break 50-day moving averages as positioning and sentiment tailwinds came to an end and stocks began to fall on bad news or weak earnings reports.

The collapse of Signature Bank and then Silicon Valley Bank (SVB) in March provided the most significant casualties of aggressive Fed tightening. In order to prevent contagion, the US Treasury, Federal Reserve and Federal Deposit Insurance Corporation (FDIC) announced that all deposits of SVB and Signature Bank would be insured, solving the immediate risk to deposit holders, and helping to stem rapid withdrawals which totalled $42bn in just four hours at peak. However, concerns remained that these bank failures were emblematic of wider issues in the banking sector, prompting extreme bond volatility and a ‘flight to safety’ with US 2-year yields falling by 130bps in just eight trading days. Credit Suisse fell soon afterwards, when actions by the Swiss central bank failed to stem client outflows and counterparty de-risking. UBS Group agreed to buy the 166-year-old lender for 3bn Swiss francs (40% of its market value) in a historic government-brokered deal aimed at containing the crisis.

Artificial Intelligence

While the macroeconomic backdrop remains highly uncertain, Chief Information Officer (CIO) spending priorities still align well with many of our key themes such as digital transformation (software), cloud and cybersecurity. The portfolio also has several additional core themes including connectivity/5G, digital advertising/ e‑commerce and EV/energy transition as well as secondary/ emerging themes such as fintech/ payments. However – as the theme of this year’s Annual Report attests – 2023 belongs to Artificial Intelligence (AI). We have been excited about the potential of AI for many years, highlighting the remarkable progress the technology has made in narrow fields. This was led by Google’s DeepMind acquisition which achieved ‘superhuman’ ability in games such as Go (2016) and Chess (2017) before solving one of the grand challenges in biology during 2021 when AlphaFold was able to predict 3D models of protein structures described at the time as “the most important achievement in AI ever”. 

That lasted until ChatGPT used a transformer model trained on 175Tb of text to generate human-like responses to seemingly any question. Able to take on different personas, write poems or programming code, even offer opinions, ChatGPT is already the first AI to “viably compete with humans”. This is likely to prove a pivotal moment for AI with Microsoft’s $10bn investment in ChatGPT maker OpenAI best understood as one of the ‘opening shots’ in an AI war that has just commenced. We have long argued that the semiconductor industry looks well positioned, with McKinsey arguing this sector might capture as much as 40-50% of the value associated with AI. This view was seemingly supported following recent record-breaking July quarter guidance from chipmaker Nvidia that was more than 50% ahead of consensus driven by AI-related strength. On the earnings call, CEO Jensen Huang spoke to a $1trn opportunity over ten years to replace CPU-based infrastructure with more efficient, accelerated computing based around GPU architectures as generative AI becomes the “primary workload of most of the world’s data centres”. Nvidia stock rose 24% on the day, despite having already gained 109% on a year-to-date basis prior to the report.

Of course, there are myriad risks associated with AI, many of which are beyond the scope of this report. However, the fact that ChatGPT makes mistakes (socalled ‘hallucinations’) is not one of them; most disruptive technologies begin as ‘good enough’ and trading accuracy for speed worked wonders for the telegraph, Encyclopaedia Britannica, and the biro. Moral and legal questions posed by AI are more difficult to dismiss, especially those regarding bias and the potential for it to “industrialise plagiarism”. While eventual regulation of AI seems inevitable, the industry would likely welcome the introduction of legislative guardrails. However, this will not be straightforward; rather than a restrictive set of regulations applied suddenly, we believe regulation may follow a ‘governance by accident’ approach that has underpinned the development of the airline industry; if aviation is any guide, it is possible that by reducing risk, regulation actually accelerates the adoption of AI, rather than stymies its progress.

As such, the focus on regulation – so soon after the advent of generative AI – might say more about investor fatigue around ‘technology disruption’ than it does about the risk regulation poses to the development of this nascent industry. This is understandable, following a period that has witnessed more than its fair share of investment hyperbole, much of which was catalysed by the pandemic. In contrast with blockchain and the metaverse – early stage technologies in search of a problem – artificial intelligence might be “the most profound technology humanity is working on”. From a historical perspective, generative AI could prove another key moment in human history when codification and dissemination of knowledge is accelerated. In the ancient world, these included the development of writing systems (such as cuneiform and hieroglyphics) around 3500-3000 BCE, as well as advanced mathematics and philosophy in Ancient Greece from the eight century BCE onwards. Libraries, historical record-keeping, and translation of ancient texts were other key developments in the codification and preservation of knowledge, aided by breakthroughs that enabled information to be stored (e.g., papyrus, paper), retrieved (e.g., cataloguing systems, encyclopaedia) and distributed (e.g., libraries, printing press). Advances in science, technology and communication during the Modern Era have “led to the codification of knowledge on an unprecedented scale” epitomised by the Internet which has facilitated knowledge sharing and democratised access to information in a manner that has changed the world.

Generative AI offers similar- if not greater - promise. Built using ‘foundation’ models which contain “expansive neural networks inspired by the billions of neurons connected in the human brain”, generative AI applications are able to process extremely large and varied sets of unstructured data and perform more than one task. This allows them to “augment human creativity, automate labour-intensive tasks and generate novel solutions to complex problems”. They can also understand natural language which means that generative AI could “change the anatomy of work” by automating activities that today account for as much as 60-70% of employees’ time. However, in contrast with historic patterns of technology automation, disruption is expected to be disproportionately felt by knowledge workers. While Goldman Sachs estimate that more than 300m jobs could be at risk, we remain optimistic that humans will graduate to higher value work just as 60% of workers today are employed in occupations that did not exist in 1940. Furthermore, McKinsey forecast that generative AI could deliver $2.6-4.4trn annually to global GDP driven by productivity gains that could be as high as 3.3% per annum when generative AI is combined with other technologies. This would be remarkable given current labour market tightness, ageing Western populations and below-average productivity growth achieved during the past twenty years.

Artificial intelligence also has the potential to become a transformative ‘general purpose technology’ (GPT) which -like electricity, steel, and the internet – may “reshape economies, drive innovation and create new opportunities”. If so, history suggests that bold, early predictions about AI may prove extremely conservative. Not just because humans struggle with non-linear change (an observation that has long informed our investment approach) but also because as yet unknown technology improvements subsequently transform the opportunity set. If early applications for steel were predictable (e.g., bridges, ships, rails), later and significantly larger market opportunities represented by skyscrapers, cars and home appliances could not be known in 1855 when Bessemer perfected his steelmaking process. The same was true for aviation when the jet engine (and other avionic developments) transformed the cost and safety profile of flight, resulting in passenger traffic growth compounding by more than 10% per year between 1950-1970 and helping travel and tourism become one of the world’s largest sectors. More recently, the confluence of internet, cloud and smartphone has presaged widespread disruption and exponential change well beyond late 1990s predictions that were only able to peer into a near and incomplete future that was yet to feature Google, AWS, and iPhones. Today, the app economy is worth c.$63trn, more than 60x times greater than the value of the handset market in 2007, the year that Apple introduced the iPhone.

The impact of generative AI is likely to be felt more rapidly than either the internet or the smartphone. In part, this reflects the role that both earlier pervasive technologies will play as AI-enablers with access to ChatGPT (and other natural language ‘chat’ interfaces) only requiring an internet connection and a smartphone. These low barriers to adoption have already supported an unprecedented rate with ChatGPT taking just 2.5 months to reach 100m users, as compared to Instagram which took 2.5 years (in itself extraordinary). Another major difference between AI and prior technology shifts is the astonishing speed of AI improvement. This is most evident when comparing the capability of two OpenAI large language models (LLMs) – GPT-4 (the latest version) and the earlier GPT-3.5 (ChatGPT) released approximately a year apart. While GPT-3.5 was trained on 175bn parameters (akin to internal variables the model learns during its training phase), the newer GPT-4 may have been trained on as many as 170trn. In addition, GPT-4 also has a much larger context window – 25,000 words vs. c.3,000 for its predecessor – which means it is able to retain far more information from earlier conversations. Aside from its “mastery of natural language”, GPT-4 “can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting”. In all of these tasks, model performance is “strikingly close to human-level performance”, evidenced by consistently high exam scores across a diverse range of disciplines (see chart).

The improvements in GPT-4 have been so remarkable that Microsoft recently posited in a whitepaper (‘Sparks of artificial general intelligence (“AGI”) that the LLM “could reasonably be viewed as an early version of AGI system”. The concept of AGI was popularised in the early 2000s to differentiate between ‘narrow AI’ being developed at the time and “broader notions of intelligence”. Until recently, AGI remained a popular science fiction topic and long-term aspirational goal within AI. That is until the range and depth of GPT-4’s capabilities “challenge(d) our understanding of learning and cognition” with the model said to “exhibit many traits of intelligence”. Naysayers argue that large language models do not ‘understand’ concepts and are merely adept at ‘improvising on the fly’. However, like Microsoft, we believe the question is moot. After all, one might ask “how much more there is to true understanding than ‘on-the-fly’ improvisation?”.