AI is not a feature
What Blockbuster can teach us about integrating AI. And why the sparkle emoji is not the answer.
In October ‘96, Blockbuster teased their upcoming website as “online at full speed with a hot and happening Entertainment Central”. What launched a month later was a website where you could see the top releases available in store. Less than a year later, Reed Hastings would found Netflix. It would take Blockbuster seven years to launch a competitor to Netflix’s web powered DVD rental. Online at full speed may have been an oversell.
There’s plenty of fodder on how misunderstood the internet initially was. It was unlikely to have a greater impact than the fax machine. Could a “couple of geeks” who “sketched out some software” really destroy Sears Roebuck? Does radio ring a bell? The root of these misunderstandings was viewing the internet as a “sustaining innovation”. Sustaining innovations, per Clayton Christensen, improve the performance of existing products. That’s compared to disruptive innovations, which rewrite the rulebook and unlock new capabilities. These new capabilities are typically undervalued at first and the products which incorporate them tend to be less expensive and worse overall. But over time the new capabilities – like not having to leave your house to order a DVD – become highly valued, disrupting the incumbents.
We’ve all learned a lot since the 90s – thanks Clay. Yet, there is a parallel to Blockbuster’s “entertainment central” spreading throughout SaaS today. It’s the sparkle emoji. You click it. You get AI. A nice sustaining innovation indeed.
AGI is already here (it’s just not very good)
Earlier this week, Des Traynor posted on X about how “Early movies were filmed in 1 static shot, cause that's what theatre plays were; it's all we knew… Later we'll learn old rules don't apply, and all the good stuff will flow…” Des was talking about the Apple Vision Pro but the same point applies to the current state of AI.
GPT-4 can write code, diagnose medical issues, summarize scientific articles, translate between languages, generate images – I could go on. This is AGI. The Deepmind team categorizes the current LLMs as “emerging AGI”:
We now have the ability to incorporate “unskilled human” intelligence throughout our applications. And “unskilled” is unfair. Current LLMs are akin to a smart undergrad who has read every book in the world but only has an attention span of 10 seconds. Is the best we can come up with really a sparkle emoji that generates more text? We already have ChatGPT.
A new substrate
The most interesting applications of AI so far are those that don’t treat AI as a standalone feature but rather see it as a new substrate for building features. We got this wrong in our first applications of AI. We added a new CMD+J action for generating queries and formulas:
This worked but it never quite felt right. It didn’t feel right that it was so separate, that AI was invoked in its own special way. Is this not just our command palette, Command+K? It turns out this was just Command+K and we merged the two in December. When we merged these features it started to feel like AI was an actual consideration of our product, not just bolted on:
So, where you put AI matters. However, “put text in a box to get more text” is still a pretty tepid use of this technology. It’s useful, yes, but remember we have a college student who’s read every book in the world. The truly interesting applications of AI are those that are deeply integrated into the product and that offer smarts without being asked.
One application I love is Superhuman’s AI summaries. Sitting below the subject line of every email is a one sentence summary of everything that’s happened in that thread. This is a tiny but perfectly executed feature. It’s not proud of being AI, it’s just part of the product:
Another example is Equals’ proactive error assistance. When you make a mistake in a query or formula, we’ll use GPT-4 to debug the error and popup a “comment” in your document. When this works, it feels truly magical. And of course, the canonical example might be Github Copilot. No prompting required, it just adds code when it thinks it can be helpful.
I hope these examples look timid in a few years, yet for all of them AI is deeply woven into the product experience. I think that will endure. AI will not be a separate experience, it will just be the product. AI is not a feature, it unlocks features. If you’re not building with AI, you probably should be. And if you are currently building with AI just make sure you’re not building a hot and happening entertainment center. Build a Netflix instead.
I don't agree with the way you classify current AI applications (lameish) as sustaining and future AI applications (new interface possibilities, amazing experiences) as disruptive.
Sustaining x disruptive innovation means, respectively, improving the product along currently valued product attributes for the most profitable customers x reframing value along "new" attribute vectors valued by "poorer" customers.
Sustaining: AI helps your product tackle even harder corner cases. Disruptive: make a simplified version of your product with much less features for a substrate of your customers that think your product is becoming too complicated (AI could help users set it up).
That said, my conclusion is that AI innovations, current and future, lame and amazing, can be both sustaining and disruptive, depending on how they are applied.