Adaption's AutoScientist: A New Tool to Speed Up AI Model Training

Adaption has released AutoScientist, a new tool designed to automate and accelerate the process of training AI models. The platform takes a different approach than most existing tools: instead of handling data preparation and model training as two separate steps, AutoScientist optimizes both at the same time. According to the company, this method has more than doubled performance across various models in internal testing, though Adaption has not yet published detailed benchmark results or disclosed which specific model types were tested.

How AutoScientist Works Differently

To understand why this matters, it helps to know how AI model training usually happens today. Normally, data scientists spend weeks or months curating and cleaning datasets—removing errors, finding patterns, selecting the right examples. Once that data is ready, machine learning engineers then take it and spend more weeks optimizing the model's internal parameters to work as well as possible on that fixed dataset. These are treated as two separate jobs, handed off from one team to another.

AutoScientist collapses these two distinct phases into one simultaneous process. The system can adjust both which data points are being used for training and the model's internal weights at the same time, with each adjustment informing the other. This is a more efficient way to work, in theory, because the tool can identify which pieces of data are actually useful for making the model better, rather than including everything that looks clean on the surface.

The approach builds on Adaption's existing product, Adaptive Data, which focuses on creating high-quality datasets. AutoScientist takes that capability further by automating the entire loop—data selection plus model optimization happening together.

Why This Matters Now

Adaption's co-founder, Sara Hooker, previously led AI research at Cohere, a company focused on building language models for business use. That experience gives her credibility on this problem. At Cohere, she would have seen firsthand how much time teams spend going in circles: cleaning data, training a model, realizing the data had problems, going back to fix it, training again, and repeating this cycle dozens of times before reaching something ready for production.

The industry is moving toward a future where base AI models become commodities—freely available from multiple vendors—and competitive advantage comes from customizing those general models for specific tasks. Organizations need ways to rapidly adapt a general-purpose model to their particular problem without requiring an entire team of specialized machine learning experts in-house. A tool that can automate this adaptation process addresses a real need.

The broader context here is that automated machine learning tools have been around for a decade, but they've traditionally focused on doing one thing well: automating hyperparameter tuning, or automating feature engineering, or automating architecture search. AutoScientist's bet is that simultaneously optimizing data and model together produces better results than optimizing them one at a time.

What Could Get in the Way

Any new automation tool faces practical hurdles in real-world deployment, and AutoScientist is no exception. When you're optimizing multiple things at once—both your data and your model—you create a vastly larger search space. This can lead the system to get stuck at local optima, finding solutions that look good in isolation but don't generalize well when the model sees new data. Detecting and avoiding this is harder when two variables are changing simultaneously.

There's also the question of whether algorithms can replicate the human judgment that goes into data curation. Experienced data scientists catch edge cases, spot distributional shifts, and notice quality problems that automated systems frequently miss. AutoScientist will need to handle these kinds of judgments without human oversight to justify the promise of accelerated training.

Running both optimization processes in parallel will require more computational power than traditional sequential approaches. For organizations already stretched on compute budgets, this could be a real cost concern that offsets gains elsewhere.

Finally, most companies have established data pipelines and MLOps workflows built around the assumption that data preparation and model training are separate stages. AutoScientist will need to integrate smoothly with these existing tools rather than forcing teams to rebuild their entire process from scratch.

A Pattern We've Seen Before

We encountered a similar wave of tools in the early 2010s, when the industry got excited about automating feature engineering. Platforms like H2O.ai and DataRobot promised to fully automate machine learning, dramatically reducing the need for specialized data science expertise. Some of these tools found homes in specific industries, but they ultimately supplemented human experts rather than replacing them. They handled routine optimization while leaving strategic decisions to the practitioners who understood the business problem.

AutoScientist may follow a similar path. The key difference is that today's foundation models and transfer learning techniques provide much stronger starting points for automated optimization than the tabular datasets that earlier AutoML tools worked with. That difference matters.

The co-optimization idea also draws on lessons from earlier research into neural architecture search and automated hyperparameter tuning, where optimizing multiple variables together often produces better results than tuning them one at a time—as long as you can manage the computational cost.

What Comes Next

AutoScientist's real test will come in production environments. The performance claims are promising, but the company will need to demonstrate these gains across a wide range of model types and real-world problems, not just in their own labs. Real adoption will also depend on three practical factors: how cleanly the tool integrates with existing enterprise workflows, whether the optimization process is transparent enough that teams can understand and debug what it's doing, and whether it uses compute resources cost-effectively.

The broader push to automate machine learning is accelerating as organizations look to scale their AI capabilities without hiring dozens of specialized machine learning experts. Tools that compress the time between identifying a problem and shipping a working model address a genuine bottleneck. AutoScientist's focus on simultaneous data-model optimization represents a logical next step in that evolution—moving beyond automating just one aspect of training toward automating how experienced practitioners actually work, adjusting multiple variables in tandem based on what the numbers tell them.

The real question is not whether the idea works in principle. It's whether it works reliably, affordably, and within the constraints of how organizations actually build and maintain AI systems.