DatologyAI is building tech to automatically curate AI training datasets

Massive training datasets are the gateway to powerful AI models — but often, also those models’ downfall.

Biases emerge from prejudicial patterns concealed in large datasets, like pictures of mostly white CEOs in an image classification set. And big datasets can be messy, coming in formats incomprehensible to a model — formats containing a lot of noise and extraneous information.

In a recent Deloitte survey of companies adopting AI, 40% said data-related challenges — including thoroughly preparing and cleaning data — were among the top concerns hampering their AI initiatives. A separate poll of data scientists found that about 45% of scientists’ time is spent on data prep tasks, like “loading” and cleaning data.

Ari Morcos, who’s worked in the AI industry for nearly a decade, wants to abstract away many of the data prep processes around AI model training — and he’s founded a startup to do just that.

Morcos’ company, DatologyAI, builds tooling to automatically curate datasets like those used to train OpenAI’s ChatGPT, Google’s Gemini and other like GenAI models. The platform can identify which data is most important depending on a model’s application (e.g. writing emails), Morcos claims, in addition to ways the dataset can be augmented with additional data and how it should be batched, or divided into more manageable chunks, during model training.

“Models are what they eat — models are a reflection of the data on which they’re trained,” Morcos told TechCrunch in an email interview. “However, not all data are created equal, and some training data are vastly more useful than others. Training models on the right data in the right way can have a dramatic impact on the resulting model.”

Morcos, who has a PhD in neuroscience from Harvard, spent two years at DeepMind applying neurology-inspired techniques to understand and improve AI models and five years at Meta’s AI lab uncovering some of the basic mechanisms underlying models’ functions. Along with his co-founders Matthew Leavitt and Bogdan Gaza, a former engineering lead at Amazon and then Twitter, Morcos launched DatologyAI with the goal of streamlining all forms of AI dataset curation.

As Morcos points out, the makeup of a training dataset impacts nearly every characteristic of a model trained on it — from the model’s performance on tasks to its size and the depth of its domain knowledge. More efficient datasets can cut down on training time and yield a smaller model, saving on compute costs, while datasets that include an especially diverse range of samples can handle esoteric requests more adeptly (generally speaking).

With interest in GenAI — which has a reputation for being expensive — at an all-time high, AI implementation costs are at the forefront of execs’ minds.

Many businesses are opting to fine-tune existing models (including open source models) for their purposes or opt for managed vendor services via APIs. But some — for governance and compliance reasons or otherwise — are building models on custom data from scratch, and spending tens of thousands to millions of dollars in compute in order to train and run them.

“Companies have collected treasure troves of data and want to train efficient, performant, specialized AI models that can maximize the benefit to their business,” Morcos said. “However, making effective use of these massive datasets is incredibly challenging and, if done incorrectly, leads to worse-performing models that take longer to train and [are larger] than necessary.”

DatologyAI can scale up to “petabytes” of data in any format — whether text, images, video, audio, tabular or more “exotic” modalities such as genomic and geospatial — and deploys to a customer’s infrastructure, either on-premises or via a virtual private cloud. This sets it apart from other data prep and curation tools like CleanLab, Lilac, Labelbox, YData and Galileo, Morcos claims, which tend to be more limited in the scope and types of data they can process.

DatologyAI’s also able to determine which “concepts” within a dataset — for example, concepts related to U.S. history in an educational chatbot training set — are more complex and therefore require higher-quality samples, as well as which data might cause a model to behave in unintended ways.

“Solving [these problems] requires automatically identifying concepts, their complexity and how much redundancy is actually necessary,” Morcos said. “Data augmentation, often using other models or synthetic data, is incredibly powerful, but must be done in a careful, targeted fashion.”

The question is, just how effective is DatologyAI’s technology? There’s reason to be skeptical. History has shown automated data curation doesn’t always work as intended, however sophisticated the method — or diverse the data.

LAION, a German nonprofit spearheading a number of GenAI projects, was forced to take down an algorithmically curated AI training dataset after it was discovered that the set contained images of child sexual abuse. Elsewhere, models such as ChatGPT, which are trained on a mix of datasets manually and automatically filtered for toxicity, have been shown to generate toxic content given specific prompts.

There’s no getting away from manual curation, some experts would argue — at least not if one hopes to achieve strong results with an AI model. The largest vendors today, from AWS to Google to OpenAI, rely on teams of human experts and (sometimes underpaid) annotators to shape and refine their training datasets.

Morcos insists DatologyAI’s tooling isn’t meant to replace manual curation altogether but rather offer suggestions that might not occur to data scientists, in particular suggestions tangential to the problem of trimming training dataset sizes. He’s somewhat of an authority — dataset trimming while preserving model performance was the focus of an academic paper Morcos co-authored with researchers from Stanford and the University of Tübingen in 2022, which earned a best paper award at the NeurIPS machine learning conference that year. 

“Identifying the right data at scale is extremely challenging and a frontier research problem,” Morcos said. “[Our approach] leads to models that train dramatically faster while simultaneously increasing performance on downstream tasks.”

DatologyAI’s tech was evidently promising enough to convince titans in tech and AI to invest in the startup’s seed round, including Google chief scientist Jeff Dean, Meta chief AI scientist Yann LeCun, Quora founder and OpenAI board member Adam D’Angelo and Geoffrey Hinton, who’s credited with developing some of the most important techniques in the heart of modern AI.

Other angel investors in DatologyAI’s $11.65 million seed, which was led by Amplify Partners with participation from Radical Ventures, Conviction Capital, Outset Capital and Quiet Capital, were Cohere co-founders Aidan Gomez and Ivan Zhang, Contextual AI founder Douwe Kiela, ex-Intel AI VP Naveen Rao and Jascha Sohl-Dickstein, one of the inventors of generative diffusion models. It’s an impressive list of AI luminaries to say the least — and suggests that there might just be something to Morcos’ claims.

“Models are only as good as the data on which they’re trained, but identifying the right training data among billions or trillions of examples is an incredibly challenging problem,” LeCun told TechCrunch in an emailed statement. “Ari and his team at DatologyAI are some of the world’s experts on this problem, and I believe the product they’re building to make high-quality data curation available to anyone who wants to train a model is vitally important to helping make AI work for everyone.”

San Francisco-based DatologyAI has 10 employees at present, inclusive of the co-founders, but plans to expand to around ~25 staffers by the end of the year if it reaches certain growth milestones.

I asked Morcos if the milestones were related to customer acquisition, but he declined to say — and, rather mysteriously, wouldn’t reveal the size of DatologyAI’s current client base.