Artificial intelligence: the next tech bubble?
- Nvidia sees increased demand for GPUs for large language models
- Potential to disrupt Google’s search market
Technology only begins to be noticed when a clear consumer use emerges. Earlier this year, DeepMind’s AlphaFold predicted all known protein structures. Before AlphaFold, scientists could predict 190,000; now they can predict more than 200mn. To people in the field this was remarkable, but without any immediate consumer applications it drifted through the public consciousness without much attention.
However, once ChatGPT was released by OpenAI, it was clear the impact it could have on people’s daily lives. ChatGPT is a language artificial intelligence (AI) program that chats with users via text. It’s based on OpenAI’s GPT-3 program released earlier this year and answers questions, in addition to debugging code and writing poetry (allegedly). The program was released to the public at the beginning of December and accumulated more than 1 million users within a week. It took Twitter and Facebook 24 months and 10 months respectively to reach the same number of users.
ChatGPT is based on a type of AI called a language model. It is trained using a large amount of information from the Internet and effectively guesses what its output should be based on all the previous data it has processed. It then irons out mistakes by having conversations with real people.
One big issue often discussed in AI is ‘alignment’: making sure the AI program aligns with human interests. In the case of ChatGPT, that means not telling people how to make bombs, or giving biased or racist answers to questions — a high-end, interactive filter. OpenAI made ChatGPT free to the public because it wants to generate user volume to test alignment.
GPUs in thebasket
Once ChatGTP is improved, it will have to become a paid program because it is expensive to run. His great language model is continued Microsoft’s (US: MSFT) Azure cloud servers that use Nvidia’s (US: NVDA ) A100 graphics processing units (GPUs). In 2019, Microsoft invested $1 billion (£810 million) in OpenAI and as part of this investment agreed to develop its products exclusively using Microsoft’s cloud servers. But rather than charging OpenAI to use its servers, Microsoft gave it credits for free use, according to a report in The information.
Each of these A100 cards costs about $3 per hour to use. Tom Goldstein, associate professor of AI at the University of Maryland, estimates that each word generated by ChatGPT will cost $0.0003. It doesn’t seem like much, but it naturally wears off over time. Each answer would cost about one cent at this price and that doesn’t take into account the hours of computing required to initially train the models.
The size of these language models creates significant hardware limitations. At the time of writing, the ChatGPT website says it is “at capacity”. If these large language models were to become ubiquitous in society, there would have to be a large investment on the hardware side. The demand for GPUs is likely to increase even further.
Nvidia, the leading GPU designer, saw its data center revenue grow 31 percent annually in its most recent quarter, and data centers now make up 65 percent of its revenue. Nvidia includes hyperscalers Amazon Web Services and Microsoft Azure as its customers, but growth is now increasingly being driven by smaller companies looking for consumer applications for large language models. “We have so many startups asking for installations of our GPUs so they can do large language model training,” Nvidia co-founder Jensen Huang said in a recent results call.
To improve its language model, OpenAI can try to improve the algorithm, but this is also a case of using more data and more parameters. Jared Kaplan of OpenAI published a paper in 2020 suggesting that there were “scaling laws”. The more data fed into the network, the better the network preformed, but doing so requires more GPUs.
Nvidia is about to release a new GPU code called Hopper, which has a higher capacity to run large language models and will be produced by TSMC (TW:2330). “One of the goals with Hopper was to lower the cost of large language model training to make it more applicable, and we’ve done that,” said Ian Buck, vice president of Nvidia.
Training these deep learning models from scratch, as OpenAI did with GPT-3, is extremely expensive. Sam Altman, CEO of OpenAI, believes that rather than building from scratch, startups will build on language models that already exist. “There will be a lot of value created in this middle tier,” he said at the Greylock Partners investor conference.
Jaspar, an AI copywriting company built with GPT-3, just received a $1.5 billion valuation and there is potential for AI to remove costs from companies currently seeing profit margins squeezed by rising wage inflation. Telecommunications company BT (BT), which has been plagued by worker strikes and downgraded its future cash flow forecasts, recently released a chatbot to automate customer service.
An area predicted to be disrupted is online search. Alphabet (US: GOOGL) dominates this market with more than 90 percent of global searches conducted with its website. However, ChatGPT’s capability means that competitors will see it as an opportunity to reshape the way consumers interact with information on the Internet.
Google itself has huge AI capacity and is ahead of OpenAI in terms of research capability, according to Alberto Romero analyst at Cambrian AI. However, because Google effectively makes all its money from searches, it will be reluctant to disrupt the model. “Google faces a case of ‘the innovator’s dilemma’: the company cannot jeopardize its core business with risky innovations just because others may eventually dethrone it,” Romero wrote.
Bing is Microsoft’s search engine and does not face the “innovator’s dilemma” because it makes up a small part of Microsoft’s revenue. Given its huge stake in OpenAI, Microsoft seems like an obvious competitor for Google’s dominance. But it’s just as likely that Google’s competition could come from a startup that has yet to be created.
ChatGPT is still flawed, sometimes producing authoritative but factually incorrect answers. In the past year, investors have learned, with both crypto and the metaverse, that cycles that are too high can lead to misallocation of capital and inflated valuations. In a blog post, Gary Marcus, a professor of psychology at New York University, describes ChatGPT as an expert imitator, but one that “never fully masters abstract relationships”.
However, the recent release was just step one in an iterative process to improve the program. GPT-4 is already in production and if the world follows Altman’s vision of the future, AI will greatly increase productivity for businesses: “The cost of intelligence and energy will quickly tend to zero, making two of the most important inputs towards the cost of everything”.
Large sudden reductions in the labor force often preceded accelerated technological development. In the 15th century, Johannes Gutenberg’s printing press was adopted after the Black Death tore through Europe. A few centuries later, the Factory Act of 1833, which prevented children from working, was followed by the widespread use of steam power in production. History doesn’t always repeat itself, but ChatGPT arrived at a suspiciously convenient time.