The Future of AI is Human?
All Things Venture #105
$0 to $500m, Real Quick - Earlier this week, Brendan Foody, the CEO & Co-Founder of Mercor had some really interesting Twitter posts. For starters, he shared that Mercor had scaled from $0 to $500m in run rate revenue in 17 months, which is well, really impressive. But what I thought was more interesting was his article on the economy and AI’s impact on jobs. You can find it here. Brendan’s central point is that, just like in prior technology cycles, a new class of jobs will be created. He thinks that, “the AI revolution will create a new class of workers tasked with guiding machines and democratizing access to their abilities.” Guiding machines and democratizing access to their abilities is in short, a technocentric euphemism for data labeling and evals. Data labeling and evals is an absolutely monster business, where even though it’s a commoditized service, foundation model providers can’t get enough them. Scale AI, Surge AI, Mercor, Micro1, Labelbox, Snorkel AI, Handshake - all of these platforms offer the same thing: access to various fields of domain experts to help train models. Billions of dollars have been spent on these services, and my guess is hundreds of thousands of people have been recruited. Surge AI recently shared that in 2024, they did $1.2 billion in revenue. Scale AI, at the time they were acquired by Meta was reported to be at $870 million in revenue.
While the revenue figures seem to be what catches everyone’s attention, and why there’s similarly been a craze to replicate the model but for hardware/robots/the built environment, what I personally think is so interesting are the parallels between data labeling and content creators. Both career paths are second order effects of a technological shift, both careers rely on tech platforms that focus on aggregation, and both are powered by an incredibly wide, individually driven interest graph meaning there’s similarly a data labeler and youtuber for egyptology (niche and specific) as there is for 5th grade math (broad and general).
The main critique that people seem to have against the data labeling business is that it may come up against a wall. There’s a belief in some tech circles that the foundation models will become so good they’ll enter this recursive self learning loop where they’ll have fully cracked the code on human creativity and discovery, which in theory should render the need for labeled data obsolete. Social media gave rise to the content creator. LLMs gave rise to the data labeler. One super reductive way of looking at things is that content creation is a job that largely requires humans to review other humans, and data labeling is a job that largely requires humans to review AI.
Personally I’m on the side of thinking that data labeling will be a durable business. I don’t see a world where human ingenuity and discovery falls off a cliff, and as long as that continues to occur there will be a need for that information to be digitized, labeled, and fed into an LLM. Things could obviously change with some step function breakthrough in the models, but while LLMs still kind of suck at general business practices like writing a good cold-email I think that we’re probably pretty far off from humans being out of the improvement loop. Only time will tell.
$100m more for AI Services - Another week goes by, and another massive funding round for an AI services business is announced. This week’s winner is Invisible Technologies which was founded by Francis Pedraza back in 2015. Invisible states that “[our] mission is to scale, enhance, and automate labor, to combine human and artificial intelligence so that enterprises can achieve what was once impossible.”
Invisible has a few products underneath the hood, one of which is data labeling, but they also market an ability to create AI native applications for their customers. Invisible is very similar to a company I featured in last weeks article, Brain Co. just going after different customer segments. Invisible seems to be more US centric and focused on F500 size customers, whereas Brain Co. is focused on large multi-nationals and government entities. In either case, both businesses seem to point to the fact that AI has a last mile problem in the enterprise. Until this last mile problem gets solved, don’t be shocked when you see more of these types of funding round get announced. Services firms and systems integrators are often maligned by VCs because they lack truly recurring revenue and as a result trade on lower multiples, but they seem to be the necessary stepping stone to actually getting AI adopted in today’s environment.
Hope you guys enjoyed the post! Per usual, I’m open to feedback and would love to hear from you. Drop a note in the comments and don’t be afraid to share with others!