Oct 2, 2024

Microsoft's Awkward AI

There's a Copilot for that...

In a vacuum, the new consumer version of Copilot that Microsoft rolled out yesterday sounds compelling enough. It's framed less as an answer engine or math machine or search replacement and more as a companion service. Sort of an AI friend. It's actually quite similar to what the startup Inflection, which Microsoft ~~hired~~ hackquired, was trying to do with Pi. But that had startup scale. Which is to say, no scale. This has Microsoft scale.

But we don't live in a vacuum and neither does Microsoft. They live in an AI environment that is a cut above cutthroat. This new Copilot isn't walking into a pristine green field of opportunity, it's walking into a minefield of conflicts and interests.

First and foremost, this Copilot's main competitor is ChatGPT. Less specifically, but more because ChatGPT is everyone's main competitor in the space right now. It's the leading chatbot. The jury remains out in terms of how many different types of chatbots people are going to use, but it's ultimately likely to be closer to one than to many. As such, there are too many of them in the market already. With more coming each day. That makes some sense as it's still the wild west days of all this. But it won't be forever. One day, the music will stop, and a lot more hackquisitions will happen. See also: what happened to Inflection's Pi. And Character.ai. And Adept. And so on...

Again, Microsoft has the scale and the capital for Copilot to avoid that particular fate. But that doesn't mean they'll find success. Remember when Satya Nadella told us that with AI, Bing was going to make Google dance? How's that working out for you? Being clever? This version of Copilot is now take two or three or four (depending on which initiatives you want to count) at making Microsoft relevant in consumer with AI.

Why will this time be different? Because this time the AI has a new UI?

Microsoft, of course, will say this time is different because of that "companion" positioning. Maybe! But again, I'm just not sure how much that will ultimately matter to consumers versus having one AI to rule them all. And right now that's ChatGPT. That may very well change. But will Microsoft be the one to change it? If so, that's incredibly awkward.

If you know one thing about OpenAI, you likely know that Microsoft is by far their biggest backer. To date, it has been a hugely beneficial relationship for both sides. OpenAI got the resources needed to scale faster than any other AI startup – enough to compete with even Google – and Microsoft got a credible AI story which both boosted their stock but also allowed them to keep such cutting edge operations at arm's length. After all, they don't own OpenAI the way Google owns DeepMind. They only own the rights to 49% of the eventual profits (capped at some number) of the company. It was a masterstroke of an investment.

But that too is now evolving. If and when OpenAI becomes a for-profit entity, things are going to get decidedly more complicated. Actual equity does that to a company. It will be fascinating to see how Microsoft and OpenAI navigate that. But Microsoft is clearly going to end up owning a massive stake in OpenAI, for real this time. And by the time that happens, this new Copilot, being captained by a former co-founder of both the aforementioned DeepMind and Inflection, may be well on its way to moving off of OpenAI's technology. Awkward.

But perhaps that's less awkward than the situation right now, where this new Copilot is being powered, at least it part (but likely mostly), by OpenAI. Mustafa Suleyman is on the record noting that the "thinking" element of Copilot is powered by OpenAI's 'o1' (aka 'Strawberry') model. But they're apparently augmenting it. Presumably the rest of the service is doing the same with OpenAI's '4o' model. In other words, OpenAI is powering Microsoft's attempt to compete with ChatGPT.

Obviously they wouldn't frame it this way. Let a thousand chatbots bloom and all that. But again, this is likely to be closer to zero-sum than "there's a chat for that".

I do think the "companion" approach will potentially be more interesting in the end-state of all this. Back in June after WWDC, I noted how Apple had likely started the process of moving the industry away from an AI arms race and more towards building AI that people will actually use. At the time, Microsoft was in the middle of their 'Recall' debacle. And it felt like they could and should learn a thing or two from Apple's approach. This new Copilot is a step in that general direction.

And yet it's awkward for other reasons now. At the most fundamental level, doesn't it feel like a "companion" should be a bit more anthropomorphized than "Copilot"? Siri, Alexa, and even Cortana had this element, but they were actually dumb as rocks to talk to. Now ChatGPT, Gemini, and this new Copilot are far more conversational but it all feels backwards! I won't make a 'Clippy' joke here, but maybe I should! A companion should have a name...

At the same time, the 'Copilot' name is even more confusing given that it seemingly means about a dozen different things for Microsoft. It's the AI layer of their enterprise suite of tools. And search assistant. And a new AI PC initiative. The branding itself is fine, but it has too many contexts! And this companion one is yet another.

Also awkward: the fact that these companions make most sense in a mobile setting. And while Microsoft has Copilot apps for this purpose, how is that going to play out as Siri and Gemini become more companion-like on iPhone and Android phones, respectively? This consumer Copilot is going to have to be that much better than everything else to beat the built-in services of those smartphones.

They'll have an advantage on Windows machines, one assumes, but also how does this new service overlap with the, say, finally fixed 'Recall' Windows functionality? Or any of the other OS-level AI functionality? Will you be able to tell this consumer Copilot on your Windows machine to do something in Office? Or will that require talking to the enterprise Copilot bot? Will they talk amongst themselves?

"Hey enterprise Copilot, our idiot user is asking me, consumer Copilot, about fixing that pivot table in Excel again. Can you help them out? Sorry to ask, again!"

What if those services are all running on different models? It's all more than a bit muddled at the moment. And likely to get even more so!