OpenAI's steamroller, or the insatiable giant

Platform and ecosystem operators have always done splash damage to the people who build for them – but is OpenAI's bystander threat worse than anything that came before?

Last week, OpenAI announced a new feature relatively quietly, without the usual fanfare it assigns to product introductions or capability expansions: ChatGPT Record, a fairly innocuous new ability for its LLM-based chatbot that’s very limited in scope and availability at launch – but that seems symptomatic of a potentially significant issue for anyone building on top of OpenAI (or other model makers’) foundations.

ChatGPT Record (which I also talk about on my new weekly podcast Artifactor, with Greg Kumparak – check it out) is just what it sounds like: A recording function for working with ChatGPT. It’s Mac-only at launch, exclusively available via the desktop Mac app that OpenAI released last year. You also have to be on either ChatGPT Enterprise, Edu, Team or Pro, containing the bounding box of those who have access even further.

And at launch, in terms of functionality, it’s not doing all that much: It provides almost an async voice mode, where you record a voice note, memo or local meeting and it can then turn that into various outputs, including meeting summaries, to-do lists, brainstorm notes and more.

OpenAI really seems to bet slow-rolling this one, given how narrow its user group is at launch relative to the whole sum of ChatGPT’s user base. Plus, there’s no accompanying blog post and limited first-party hype, beyond a mention in an overall ChatGPT for Business video update from a couple of weeks ago, a help article and a LinkedIn update from a few days ago when it actually went live.

Increasingly, I think we’re going to see OpenAI play the tricky dance of trying to deploy useful and intuitive new products and extensions of its core features that work for both business and creative users – while also trying not to compete too directly (or at least, loudly) with the partners who build on top of its API and developer-facing tooling and infrastructure.

ChatGPT Record is a direct shot across the bow of startups and companies building in one of the buzziest and most active generative AI subcategories there is: meeting notetakers. They’re so prolific that talking about them and their omnipresence is replacing idle chitchat about the weather as the default small talk for when you’re waiting for everyone to join at the start of a call.

You can bet that the Granolas, Fireflies and Fathoms of the world sat up and took notice when Record was announced; likely, they were already on high alert before. ChatGPT’s potential to become a true productivity suite is reminiscent of when Gmail first arrived, and some startups focused on building things like a ‘calendar’ app specifically for Gmail users (check out the origin story of Justin.tv, which became Twitch, for more about how that ended up working out).

For those of you who follow the Apple blogosphere and news cycle, you’ll probably already be familiar with the term ‘Sherlocking’ – basically, the process of a platform provider expanding its core to include the features offered by a third-party building in its ecosystem. Astropad, which was itself Sherlocked, has a good practical primer for those looking to learn more. It’s not a new phenomenon in tech, nor is Apple the lone example (Shopify is another repeat offender), but in the case of OpenAI and the other foundational model makers who also have user-facing products, I think the speed and frequency of Sherlocking is going to increase rapidly.

Intuitively, it should already be clear that the current glut of similar products at the application layer in generative AI is unsustainable and bound for consolidation. Exits to foundational model companies will be part of that, as will mergers of startup competitors looking to survive and thrive against the big dogs. Meanwhile, the information gap works very much in favor of the platform providers: OpenAI, Anthropic, et al. know way more about what kind of products are working on their platform thanks to access to metrics like volume and frequency of API calls across customers than those building on top.

But I think that a surprising number of startups – including many with promising early traction – might just end up caught under the treads of the relentless metastasizing of the product machinery of OpenAI, Anthropic and others as they seek revenue and build on some of the fastest-growing direct user bases of any tech era to date.