How Adaption's New Tool Could Speed Up AI Training

A company called Adaption has released a new tool called AutoScientist. It is designed to help people build and train AI models faster by automating much of the tedious work involved.

The tool works by handling two normally separate jobs at the same time: preparing data and adjusting how the AI learns. In typical AI development, these happen in sequence—first a human prepares the data, then an engineer spends weeks tuning the AI itself. AutoScientist does both together. Adaption says their tests show the tool has more than doubled how well their AI models perform, though they have not yet released detailed numbers or explained exactly how they measured this.

How It Works

Think of training an AI like baking a cake. Normally, a baker gathers and measures all ingredients first. Then they adjust the oven temperature and baking time separately. AutoScientist tries a different approach: it adjusts the ingredients and the oven settings at the same time, checking how the cake turns out and tweaking both in response.

The tool builds on another product Adaption already makes, called Adaptive Data, which helps create high-quality datasets. AutoScientist takes that further by also automatically adjusting how the AI learns from that data.

In real AI work, training takes a long time. Teams cycle through preparing data, cleaning it, tweaking how the AI processes information, and testing again and again. AutoScientist aims to compress this—instead of doing each step once after the other, it tries adjusting multiple things at once and watches for improvements.

Who Else Is Doing This

Several other companies make tools that help with AI training. Most tools like MLflow and Weights & Biases treat data preparation and model training as separate problems, handled by different tools. AutoScientist focuses specifically on doing both at once.

The company was co-founded by Sara Hooker, who previously led AI research at Cohere, a company that helps other businesses use large AI models. Her background brings credibility, since she has faced these exact problems before—training powerful AI systems and then customizing them for specific customers.

This timing matters. As AI models become easier to copy and share, companies win by customizing general-purpose AI for their own needs. Tools that can do this quickly and cheaply are valuable.

What Could Go Wrong

There are some real challenges that AutoScientist needs to handle. When you optimize two things at once instead of one thing at a time, the search space grows much larger. This can make the system get stuck at a local dead-end rather than finding the best solution.

Second, when the system selects which data to use, it needs to avoid overfitting—a problem where an AI learns patterns specific to the training data that do not apply to real-world situations. Humans are good at spotting these traps. An automated system might miss them.

Third, there is the question of compute cost. Optimizing two things at once typically requires more computing power than doing them separately. The benefit of faster training might be offset by higher electricity bills and longer hardware waits.

Finally, most companies already have established systems for managing data, tracking experiments, and deploying AI. AutoScientist would need to fit into these workflows rather than replace them entirely.

What We Have Learned From This Before

The AI industry has tried this kind of automation before. In the early 2010s, tools like H2O.ai and DataRobot promised to automate large parts of machine learning, removing the need for specialized data scientists. These tools did find customers in certain industries, but they ended up working alongside human experts rather than replacing them. Specialists still made the big decisions; the tools handled the routine optimization.

AutoScientist might follow the same path. The AI models available today are much more powerful than they were back then, which gives the tool more to work with. But the core pattern is familiar—automating the repetitive parts while leaving room for human judgment.

What Happens Next

For AutoScientist to succeed in real business use, it will need to do several things. First, it needs to work smoothly with the tools companies already use. Second, the optimization process needs to be transparent enough that teams can understand what happened and debug problems. Third, it needs to be cost-effective.

Adaption has made strong claims about performance improvements. Proving these claims work across different types of models and real business problems—not just in their own labs—will be crucial for adoption.

Broadly speaking, companies want to build more AI systems without hiring twice as many AI specialists. Tools that can compress the time and guesswork between identifying a problem and getting an AI solution running are addressing a real need. AutoScientist's focus on adjusting data and learning together is a logical next step in making AI development faster and more practical.