Recently, OpenAI has been releasing products that don’t seem to advance the goal of creating AGI or Superintelligent AI. The social video product Sora is one such product. Though at least it involves a new AI model that might teach them lessons about developing capabilities. Another is the AI-enabled browser Atlas.
Why is OpenAI spending the manpower, compute, and management focus to create and support these products instead of using those resources to race toward more powerful AI?
I am most interested in public sources from OpenAI or answers from those who have direct knowledge. If you have a different speculation from mine, share that too.
Speculation on possible reasons:
Investors want OpenAI to have more revenue or more products. I could imagine that not all investors fully believe the vision of reaching AGI, or suspect that reaching superintelligence will destroy their investment, as capitalism is fully disrupted.
These products enable collecting data that they can use to train AIs. (I’m not clear how that would be the case.)
They are “dogfooding” their coding assistant AI and need projects to use it on, so they have made these products. (But even if you’ve developed it, that doesn’t mean it is worth the costs of releasing and supporting it.)
Their employees want to develop cool products and see the fruits of their work. Creating and releasing these products keeps the employees happy, so they will continue to be maximally productive in advancing AI.
OpenAI is bottlenecked on AI researchers. They cannot develop more powerful AI any faster than they are, so they can afford to hire developers and others to create products. Those products encourage the buildout of computing capabilities that will be available when the AI researchers need them.
AI development is slowing or hitting bottlenecks, so they want to capture the market share of products that can be developed with current levels of AI.
As much as I want it to be true that progress is stalling, I think Sora and Atlas and the rest are mostly signs that OpenAI is trying to become an everything app via both vertical and horizontal integration. These are just a few of the building blocks they would need, and the business case for targeting an everything app given their valuation seems strong (in that the only way the valuation is defensible is if it seems like they will eat multiple existing business sectors or create new business sectors as big as many existing ones).
why train them: you need strong world sim to get full agi
why release them: make money directly, increase doubt in media, addict users
Plus be well placed to control the narrative. I’d guess that’s the main one, along with be impressive in a way which makes fundraising easier.
also “crowdsource the search for bugs in the implicit physics engine of the video generator”
The cynical answer is that Sora and Atlas are likely to be profitable, and there is no mechanism left by which OpenAI can choose to not do a profitable thing.
A superintelligence would likely interact with a large part of the world via a browser. Building a browser that works well with their AI, seems to me like it helps developing an AI that can do task in the real world.
It also provides a lot of training data of AI acting as an agent that can be valuable for building superintelligence.
I’m not sure they really have a strong singular goal of AGI anymore, even if they say they do. The company seems to pivot more and more toward ordinary big-tech style business models, and investor expectations. “AGI” is one story they can tell to push their ordinary financial agenda forward. It also allows them to hedge themselves and survive as an ordinary big-tech, in case the whole AGI thing never really eventuates within a few decades.
The simplest answer is progress is stalling. They could have gone for the engagement optimization angle since 2023, but there were promising alternatives then. By 2025, these all failed. Pretraining returns stalled and reasoning proved too inefficient to scale.
I agree with Hank Green that it sure seems like it’s so they can start selling ads like a traditional social media company, and furthermore that that sort of behavior doesn’t feel like what one would expect from a company that thought they were building an AGI.
https://x.com/btibor91/status/1994714152636690834