Datacurve raises $15 million to take on Scale AI

As AI corporations mature, the struggle for high-quality knowledge has grow to be some of the aggressive areas within the trade, launching corporations like Mercor, Surge, and, most prominently, Alexandr Wang’s Scale AI. However now that Wang has moved on to run AI at Meta, many funders see a gap — and are keen to fund corporations with compelling new methods for gathering coaching knowledge.

The Y Combinator graduate Datacurve is one such firm, specializing in high-quality knowledge for software program improvement. On Thursday, the corporate introduced a $15 million Collection A spherical, led by Mark Goldberg at Chemistry with participation from workers at DeepMind, Vercel, Anthropic, and OpenAI. The Collection A comes after a $2.7 million seed spherical, which drew funding from former Coinbase CTO Balaji Srinivasan.

Datacurve makes use of a “bounty hunter” system to draw expert software program engineers to finish the hardest-to-source datasets. The corporate pays for these contributions, distributing over $1 million in bounties thus far.

However co-founder Serena Ge (pictured above with co-founder Charley Lee) says the most important motivation isn’t monetary. For prime-value providers like software program improvement, the pay will at all times be far decrease for knowledge work than standard employment — so the corporate’s most essential edge is a optimistic person expertise.

“We deal with this as a shopper product, not an information labeling operation,” Ge mentioned. “We spend quite a lot of time desirous about: How can we optimize it in order that the individuals we wish have an interest and get onto our platform?”

That’s notably essential because the wants of post-training knowledge develop extra complicated. Whereas earlier fashions have been skilled on easy datasets, at present’s AI merchandise depend on complicated RL environments, which have to be constructed by particular and strategic knowledge assortment. Because the environments develop extra refined, the information necessities grow to be each extra intense for each amount and high quality — an element that would give high-quality knowledge assortment corporations like Datacurve an edge.

As an early-stage firm, Datacurve is targeted on software program engineering, however Ge says the mannequin may apply simply as simply to fields like finance, advertising, and even medication.

“What we’re doing proper now could be we’re creating an infrastructure for post-training knowledge assortment that draws and retains extremely competent individuals in their very own domains,” Ge says.