Happy 2nd Birthday ChatGPT. What do you have to show for yourself?

ChatGPT is coming up on its second birthday. Few toddlers have gotten as much attention or such high expectations. Now the hype-cycle appears to have peaked with Goldman Sachs saying the huge investments in Generative AI may not be recouped. I think both the initial hype and the new pessimism are overdone. GenAI is rapidly moving up the path Clayton Christensen laid out for disruptive technologies in the Innovators Dilemma way back in 1977.

Christensen’s thesis is that in tech markets there are different types of customers whose needs range from low quality to the most demanding quality. Incumbents tend to compete for the most demanding uses since they bring the most margin. Disruptors tend to establish themselves at the low end of the market and move up as their quality increases. This has many advantages including that the incumbents don’t feel threatened until the disruptor has had a chance to establish a business. I call this eating a dinosaur from the bottom up. [I think it’s an original term, but if anyone knows someone who said it first, please let me know.]

GenAI is doing exactly what Christensen described. It’s making multiple markets at the bottom while getting dinged for not being ready for the top… yet.

1)?????? Compete with non-consumption Below the low-quality users are customers who would like to use, but even the lowest quality version of the dominant technology is still too expensive.

Human translation is expensive. Today AI is producing captions, transcripts, summaries, and translations of calls and meetings that would not have existed without it.? For deaf and non-native speakers these can be the entrée to fully participate at work and at school. Call center software is now analyzing transcripts, making “this call may be monitored for quality and training” a reality.

A sign of just how fast AI transcription is moving up the curve is that the Association of Court Reporters and Captioners is fighting against using AI in courtrooms, though some are using AI-assistance. ?

?

2)?????? Disrupt low quality uses

Writing: People are rightly worried about AI displacing journalism, literature and other important writing. However, there are mountains of very pedestrian writing for which GenAI is more than capable. Some of the biggest early users of GenAI are e-commerce sites that need product descriptions for thousands of items in multiple languages.?

Many of GenAI’s biggest early applications replace low quality search of proprietary knowledge stores with RAG (described here.) RAG applications can understand questions and return answers in plain English, or almost any other human language. A public example is H&R Block’s AI tool for answering tax questions. RAG applications are far from perfect, and the answers aren’t always correct, but they are enough better than key-word search to justify much higher cost. Any organization that spent money building a proprietary knowledge base should be looking at RAG to make it more valuable.

3)?????? Augment higher quality uses – Copilots and assistants

One of the reasons I remain bullish on GenAI despite the daunting economics is that it can get very big by augmenting rather than replacing a huge percent of the software in use today. This means that GenAI doesn’t need to fight the incumbents. The incumbents will add it to their offerings as copilots and assistants. Normally software makers would buy or build a new feature, but the economics of training and hosting LLM means most will need to pay to use GenAI as a service for a long time.

Here are some of the GenAI augment applications I’m most bullish about:

·?????? E-commerce: IKEA, Wal-Mart and other e-commerce sites are adding conversational front ends to their existing e-commerce sites. The AI can answer questions and make recommendations. Computing ROI is straight forward - a big bonus for driving adoption.

·?????? Co-pilots: We’re starting to see assistants and co-pilots being built into existing software to make users more efficient.

o?? GitHub Co-pilot is used by a staggering 70% of software developers, who report it makes them more efficient and their work more enjoyable, though there are no hard metrics to justify the cost. It is available as an add-in to VS Code, the leading development platform.

o?? Legal software has incorporated GenAI faster and more fully than any other profession I’m aware of. It seems surprising because the lawyers I know tend to be conservative and tech-phobic. However, lawyers are also the rare profession that bills by the 10th of an hour. Confirming a central tenet that AI is adopted first and fastest when you can compute the ROI.

4)?????? What’s next?

GenAI’s next use cases are well defined but still require some invention to operate at scale. There include:

·?????? Analysis of quantitative data: LLMs are large LANGUAGE models. They can appear to reason their way through math problems, but in a very clever experiment some Apple engineers showed that most of what appears to be LLMs doing math is just copying solutions they have seen before. (In the experiment they confused the LLM by changing names and quantities in common story problems.) ?

Once GenAI is able to do math reliably we will be able to ask it to query a database and return answers. Millions of reports that today must be custom coded will be possible by non-coders asking questions in plain English. We’ll still need data pros to set up access to sensitive data, but a massive amount of low value coding work will no longer be necessary.

·?????? Agents: Today Gen AI is largely confined to producing text, but there are huge opportunities if it can reliably take actions on our behalf. IKEA and other e-commerce sites are experimenting with conversational product search that queries inventory and can add items to a shopping cart. But from what I’m seeing it isn’t reliable enough for situations where the actions are less constrained. However, once the agents are reliable things will get very interesting… in both the good and bad ways.

Net, I don’t believe the incredible hype, but I also don’t buy the new ennui. Generative AI is a disruptive innovation of the first order. Like other new technologies in the beginning it is confined to applications where it is good enough, but over time the quality and reliability will improve. Happy Birthday ChatGPT.