SAN ANSELMO, CALIFORNIA - JANUARY 27: In this photo illustration, the DeepSeek app is displayed on ... [+]
Technology happens quickly. This time two years ago, few people had heard about ChatGPT, and few investors knew what Nvidia did. On Jan. 6, 2023, I wrote a note entitled ‘Talos’ where I remarked …
’A recent development here is the arrival of ChatGPT an interactive ‘intelligent’ bot that has been developed by OpenAI (set up seven years ago to build socially constructive AI, and recently valued at USD 30bn). ChatGPT is catching on quickly, not least because students have found that it can write half decent essays. I recently tested it out, asking for a response to the question ‘Is globalization over?’- the result is below, and in my humble opinion is a good rendition of the kind of response that a ‘two handed economist’ might give (‘on one hand…on the other’). I think I can just about do better, and if there is any lesson to draw it is for human writers to be more opinionated, quirky or style driven in how they write. I am not out of a job just yet.
I am still writing, and don’t use ChatGPT to do so. In January 2023, my presumption was that because ChatGPT had already been in place for some time, investors knew about it and had factored it into share prices (the theory of efficient markets in finance says that all publicly available information is quickly reflected in market prices). This was not the case, and with hindsight, I realize that students of economic and finance of the 1990’s spent far too much time imbibing the ‘efficient markets hypothesis’, which was a big thing at the time and the big guns of academic finance raged about it into the 2000’s. I suspect that more people today believe markets are rigged than efficient, with the rigging done by vampire squid type funds and central banks, to name a few culprits.
The performance of AI centric stocks (notably Nvidia) shows that the early promise of AI was underestimated, perhaps to the extent to which it is overestimated today. Indeed, the concentration of the top 10 companies in the U.S. stock market (they make up 40% of the market capitalization) is as high as it only has been in 1930. In that context, the apparent supplanting of ChatGPT by DeepSeek in the Apple app store is a reality check.
It is also an efficient markets check. I had seen the initial model test results from DeepSeek at the start of this year (Azeem Azhar’s blog had mentioned it back in November, and also as far back as December 2023). To that end the technology world has been well aware of DeepSeek for months, so it is a surprise, and another slap in the face for efficient markets hypothesis (I am still a little obsessed about this).
Whatever about efficient markets, the advent of DeepSeek is a reminder of how the cycle of innovation in economies works, and will have some commentators rushing for their Schumpeter. The Austrian economist coined the term “creative destruction,” but his views on capitalism are worth re-reading for he believed that it would undermine itself (different cohorts or corporates would effectively block the system) and collapse. His work on business cycles is also well overdue a comeback.
We don’t quite know whether DeepSeek (who claim to have produced superior model results compared to ChatGPT and Claude AI for a fraction of the cost) have been helped by the Chinese state, whether any espionage is involved nor the extent to which they have piggybacked on work already done by U.S.-based teams. Nor is it too important that the bandwidth of the model is restricted in China and that it is unlikely to be widely used in the West.
The important development is that the cost of production threshold for large AI models has been set much lower. Other cheaper models are on the way, Bytedance's Doubao-1.5 or Moonshot's Kimi k1.5 are two emerging examples.
To that end, the first phase in the AI boom now comes to an end. Much like other key technology infrastructure breakthroughs – railway (railway companies made up 50% of the stock market in 1900), automobiles (France dominated the auto market from 1903 to 1929, at times manufacturing half of the world’s cars) to the internet (anyone remember AOL or Netscape”) – the initial innovation garners huge amounts of capital, is expected to change the world, which it does but often not in the ways investors expect. The legacy is usually a new web of infrastructure.
In the next phase of AI, Investors will now not likely pay up for model developers but will focus on unique datasets, applications of AI (health) and the AI industrial supply chain — energy for AI for example (see last week’s note ‘Humphrey’).
The view of the stock market, which quickly rebounded, is that the consumer is the winner — that DeepSeek opens up the prospect of cheaper, better AI for common use, and I think the investor conversation will turn to how this impacts consumption patterns and the way people work (from police officers to teachers). The fact that the stock market took the positive view of the DeepSeek news suggests that the AI bubble in stocks is alive and well.
Elsewhere in AI, model development is accelerating in a sinister way. In a paper titled “Frontier AI systems have surpassed the self-replicating red line,” scientists at Fudan University highlight how AI can build replicates of itself, and when this process runs into obstacles, demonstrate a survival instinct (such as rebooting hardware to fix errors). It strikes me that this is the real AI development we need to pay attention to.
