In the pursuit of the perfect AI model (spoiler alert: It may never happen, but let’s keep going), or at least AI model optimization, where the moving parts of automation intelligence are in their best possible state, most attempts still fail at the data layer. This often happens because datasets themselves are too thin, noisy, or the techniques needed to make them more efficient are locked inside frontier labs.
Adaption is a new ‘neolab’ company (an organization defined as part of a new generation of frontier AI specialists that follows the more established companies like OpenAI, Anthropic and Google DeepMind), and the firm is teaming with Together AI (a cloud platform for open source models) to streamline fragmented generative AI workflows for enterprises using open-weight models.
Together AI fine-tuning is now integrated directly into Adaption’s Adaptive Data platform.
Fast Intelligence, No Preconfigured Puff
According to Adaption co-founder and CEO Sarah Hooker, the way forward sees us remember that “intelligence should not arrive preconfigured”, so, instead, it should be fast, efficient and owned by builders all around the world.
Together AI runs the training infrastructure so that teams don’t have to, scaling from LoRA runs (standing for Low-Rank Adaptation – this process freezes pre-trained model weights and injects trainable rank-decomposition matrices, significantly reducing parameters for more efficient, low-memory fine-tuning) to full fine-tuning of the latest and strongest open models.
Making use of “adaptive data” (branded and productized as Adaptive Data in this case) to shape datasets toward a target model behavior, Hooker and team say that teams can evolve what they already have, expand what’s missing and rebalance where failures live.
The two now work together, end to end, in one platform to provide the shortest path from raw data to a deployed fine-tune that behaves the way a business needs.
Hooker’s comments appear in line with commentary made by her fellow co-founder, Sudip Roy, and Max Ryabinin, Together AI, VP of research and development.
Rebalancing The Long Tail
“Adaptive Data evolves your dataset to match the behavior you want, and shapes the data to your objective. It rebalances the long tail, fills in missing diversity, and grows your seed data into a training set large enough to move the model. Early deployments see an 82% average lift in dataset quality against target objectives, across 242 languages,” note the pair.
The companies promise model training that keeps up with a team’s latest iteration.
Using these technologies, they say that AI developers and data scientists can launch and fine-tune on Together AI’s infrastructure without leaving Adaption. Adaptive Data reads the dataset, model and volume, then recommends hyperparameters tuned to that combination, so you’re not starting from defaults. Training cycles drop from weekly to hourly,
Model Behavior, By Design
Most AI products are built from the same handful of off-the-shelf models, tuned by the same recipes, optimized for the same benchmarks. Hooker and team say that this new offering means the engineering group can now pull ahead and be able to treat model behavior as something they design, not inherit.
