The AI summer — Benedict Evans

🌈 Abstract

The article discusses the rapid rise and adoption of ChatGPT, as well as the challenges and uncertainties surrounding the widespread deployment of large language models (LLMs) in both consumer and enterprise settings. It explores the parallels between the current AI hype cycle and previous technology adoption curves, highlighting the need for a more gradual and thoughtful approach to integrating these powerful AI systems into real-world applications.

🙋 Q&A

[01] The AI Summer

1. What are the key points made about the rapid adoption of ChatGPT?

• ChatGPT exploded into public consciousness in late 2022 and quickly gained 100 million users in just 2 months, much faster than the adoption of previous technologies like the iPhone or cloud computing.
• However, most people who tried ChatGPT have not been back, and there has been limited growth in active users over the past 9-12 months.
• The article suggests that the rapid adoption of ChatGPT may have been a "trap", as LLMs look like finished products but are still lacking in terms of real-world usefulness.

2. How does the article compare the adoption of ChatGPT to previous technology adoption curves?

• The article draws parallels to the Dotcom bubble, where many failed ideas eventually succeeded after a longer period of time for infrastructure and consumer/business behavior to catch up.
• It also compares ChatGPT's rapid adoption to the iPhone and cloud computing, which took much longer to reach widespread deployment and usage.
• The article suggests that LLMs may need to go through a similar "slow, boring hunt for product-market fit" before becoming truly useful and widely adopted.

3. What factors are contributing to the slow enterprise adoption of LLMs?

• The article cites enterprise IT sales cycles, which are longer than the time since ChatGPT was launched, as a key factor slowing enterprise adoption.
• It also notes that while there is a lot of experimentation and pilot projects, far fewer organizations are actually deploying LLMs in production use cases.
• The article suggests that the usefulness of LLMs varies greatly depending on the specific use case, with some areas like coding and marketing seeing more adoption than others like legal or HR.

[02] The Trap of LLMs

1. How does the article describe the "trap" of LLMs?

• The article suggests that LLMs "look like products and they look magic, but they aren't" - they appear to be finished, generalized products, but in reality still lack the necessary refinement and integration into real-world applications.
• It argues that LLMs have "skipped the slow painful process" of working out product-market fit and instead went straight to being touted as a solution for "everything", before actually meeting the needs of users.
• The article compares this to previous technology adoption curves, where new technologies had to go through a period of growth and learning before becoming truly useful.

2. What are the factors driving the rush to deploy LLMs?

• The article cites competitive pressure, stock market pressure, and a sense that LLMs represent the "next platform shift" that companies need to "grab with both hands".
• It also notes the "standing on the shoulders of giants" effect, where LLMs can leverage existing infrastructure and don't require waiting for new devices or networks to be built.
• The article suggests this has led to a "firehose of cash" being poured into LLM development and deployment, despite the technology still being mostly in the experimental stage.

3. How does the article view the current landscape of LLM startups?

• The article presents a "glass half-full, glass half-empty" perspective on the current LLM startup landscape.
• On one hand, the large number of startups could be seen as a collective bet that LLMs are a technology that needs to go through the conventional product-market fit process, rather than a finished product that can immediately replace existing software.
• On the other hand, the article acknowledges the possibility that LLMs may be able to "swallow most or all of existing software" and automate vast new classes of tasks, potentially making the current startup landscape a "bubble" that will inevitably burst.