Work on problems constrained by software
Will generative AI re-emphasise the value of software over content?
An important concept in chemistry is that of a limiting reagent. When two reagents are mixed to trigger a chemical reaction, there will always be an excess of one of them. Add a drop of base into an acid solution, and you get a slight less acidic solution with a pinch of salt. The base acts as a limiting reagent, constraining the reaction from proceeding further. Add an excess of base and the reaction continues further, until the acid becomes the new limiting reagent.
The same principle applies to the growth of plants as well: they are constrained by water, sunlight, carbon dioxide, or a nutrient such as nitrogen, phosphorus, or potassium. Add fertiliser to a plant growing in poor soil and it grows taller, up to a point, when adding more fertiliser would no longer help.
If processes in nature are usually limited by a single limiting factor, is it also the case in human society and business? I think it is. Early human societies were constrained by lack of food: the discovery of agriculture solved the problem and kickstarted the rise of civilisation. Similarly, production of goods in pre-industrial societies was constrained by the number of skilled artisans, as well as the time each of them had available to hand-craft their goods. Invention of industrial machines like the Spinning Jenny removed this constraint, and ushered an era for increasingly cheap, mass-produced goods.
There as still many problems to be solved in society. Some of them are constrained by software.
Constrained by software
First computers didn’t do much. As expensive calculators, they were used for mathematical calculations, data processing, and cracking the Nazi encryption code Enigma. These problems were constrained by software — no human analyst could test enough permutations of the Enigma code to be able to crack the messages. As computers became more powerful, their utility increased as well. More and more problems turned out to be constrained by software.
In the 2000s, a tried and tested pattern for building a successful startup was to digitise an analog process or service. Booking a table at a restaurant is more efficient for everyone when done through the internet than over the phone [1]. Digital maps need not be re-printed every year, and can hold the whole planet at every resolution in your pocket! With software companies building profitable businesses in so many sectors of the economy, it was easy to think every problems in society are constrained by software.
Content is king?
Netflix became the largest streaming service in the world by solving video rentals through software. While subscription services seem ubiquitous today, a business model where a small regular payment gives you access to unlimited content is actually quite revolutionary, and only possible with software, a world free of the analog world’s constraints.
However, Netflix is struggling today. Traditional broadcasters and film studios like Disney and HBO are catching up. Streaming software is no longer a key competency, but a commodity: users no longer choose Netflix over Disney+ thanks to the features of Netflix’s video player. Streaming is now constrained by content, not software. [2]
Netflix is not alone. In 2023, it is no longer enough to launch a platform and expect users to flock into it. Whether in holiday booking, e-commerce, or social media, platforms are a commodity. What you need is content and community — a reason for people to start using your service. A new wave of post-Covid startups are embracing this trend. WeRoad is a travel platform that groups solo travellers into like-minded groups. WeRoad competes with community, avoiding the commodised market of travel booking websites dominated by megacorps like Booking.com (NASDAQ: BKNG), Tripadvisor (NASDAQ: TRIP) and Trip.com (NASDAQ: TCOM). In my opinion this insight is still undervalued in today’s market, as many people still view platforms as a new and innovative business model.
GPT — software strikes back!
I started writing a much longer blog post about the role of custom content in technology plaforms in autumn. It now has to be rewritten, as the role and importance of content has been revolutionised by generative AI.
Pre-ChatGPT content was valuable due to its scarcity. Building a crude website with a few key screens can be done in a weekend, while a back-catalog of interesting content may take decades to acquire. This was even clearer in specialised industries like law, healthcare, and education, where material used by professionals has to adhere to common standards and regulations and thus needs to be written by domain experts. This has allowed traditional publishers to maintain a lead over their digital rivals in educational and professional publications, while their mass-media counterparties have lost to user-generated content platforms.
Generative AI might upend all of this. GPT-4 recently passed the Uniform Bar Exam (and defeated me in Chess). Like Spinning Jenny, GPT-4 does not need years of experience to craft its articles, and can write more in one evening than I can in a lifetime [3]. GPT-4, and its main application ChatGPT, have taken the world by a storm, but the long-term effects of these tools are still uncertain. I predict that generative AI will transform many problems from being constrained by content back to being constrained by software.
This is already happening in language education. In February 12th I tweeted the following:
By carefully constructing a prompt, I was not only able to transform ChatGPT into a Chinese language tutor, but to also output its answer according to a specific, pre-defined, machine-parseable format! Add a website, app, or voice assistant on top of the output, and you have a powerful language-learning app that could give Duolingo a run for its money — buildable in a weekend, convertible to any language in the world, no content team required!
Duolingo is well aware of their disappearing moat. In March 14th, the company released Duolingo Max, an AI tutor built on GPT-4 similar to my prototype tutor. For a company built on content and courses designed by humans, this is a major pivot. Duolingo’s financial statements combine spend in content with the wider Research and Development category, so it’s not possible to direcly determine how much the company invests in content every quarter. However, in 2022 Duolingo spent $44.5 million in research and development, which suggests a content spend well in millions, or tens of millions, of dollars a year. In a world constrained by content, that investment was a formidable moat against aspiring startups. But how much is it worth in a world constrained by software?
Time will tell what the impact of generative AI will be. But as more problems transform back to being contrained by software, agile startups uncumbered by legacy processes will be able to dethrone legacy incumbents. If you are a decision maker in one of these market leaders, be warned! The rules are being rewritten as we speak.
At the same time, if you are thinking of founding a technology startup, make sure the problem you are solving is actually constrained by software. Many incumbents have terrible software, but none of it matters if software is not the problem. You can also learn from WeRoad by designing your product around a community from the ground up, thus raising above the competitors.
Footnotes:
- This is a generalisation: there are few cases where speaking directly with staff is a better option than booking online, such as when organising bookings for a large group of people, requesting a specific table for a special occasion, or if a diner requires special accommodations. It’s important to keep “edge cases” like these in mind when designing your approach to customer service!↩︎
- Ben Thompson’s Stratechery blog has a great in-depth analysis of Netflix: https://stratechery.com/2023/netflixs-new-chapter/↩︎
- For full disclosure: I never use ChatGPT or other generative AI services when writing blog posts. I don’t write for money, so I have no incentive to churn out multiple mediocre articles a day. I’d recommend all readers to follow small, hobbyist creators instead of full time content farmers who do have a real incentive to generate blog posts with GPT. I do occasionally use Dall-E and other tools to generate illustrations for my posts.↩︎
