Foundational values for foundation models

The Great AI Kitchen: Why We Choose Between Master Chefs and Scratch Cooking

Imagine you are trying to build a robot that can look at an X-ray and tell you if a patient has a broken bone. In the world of medical science, this is a high-stakes game where the robot's "brain" (its algorithm) needs to be incredibly smart, but also safe and understandable. For a long time, scientists built these brains from scratch, teaching them every single rule of anatomy and physics from the ground up. But recently, a new trend has exploded: "Foundation Models." Think of these as giant, pre-trained super-brains that have already read the entire library of human knowledge. They are like master chefs who have already cooked every dish in the world; you just ask them to make a specific soup, and they tweak their recipe slightly instead of starting with raw ingredients.

But here is the tricky part: just because a master chef can make your soup doesn't mean it's the best choice for your specific dinner party. Maybe you need a dish that uses very little water (because your well is dry), or maybe you need to know exactly why the chef added salt so you can trust the taste. This is where the question gets deep. Scientists are now arguing over whether to use these giant, pre-made brains or to build their own smaller, custom ones. It's not just about which one is faster or smarter; it's about what values matter most. Do we care more about saving energy? Do we care more about being able to explain our decisions to a doctor? Or do we care about making sure scientists in poorer countries can participate in the game? This paper dives into that exact debate, mapping out the hidden values that drive these technical choices.

---

The Map of Values: Why We Choose Our AI Tools

The authors of this paper, John Baxter and Elodie Germani, decided to play a game of "Why?" with the scientific community. They asked: "Why do researchers choose to use these giant Foundation Models, and why do others choose to train their own models from scratch?" To answer this, they used a method called the "Socratic approach." Imagine a detective asking a suspect, "Why did you do that?" and then asking, "And why did that reason matter?" over and over again until they hit the very core belief. They did this to build a giant map (a graph) showing how different values connect to technical decisions.

Their investigation revealed that the choice isn't a simple "good vs. bad" switch. Instead, it's a tug-of-war between two sides of a spectrum. On one side, you have using Foundation Models with minimal changes (like hiring the master chef). On the other side, you have training a model from scratch or making huge changes (like cooking from scratch).

The Case for the Master Chef (Foundation Models)

Many researchers lean toward using these pre-trained giants for a few very practical reasons.

First, there is Data Efficiency. Imagine you are trying to learn to diagnose a rare disease, but you only have five patient records. A model trained from scratch would be confused and fail. But a Foundation Model has already seen millions of images and patterns. It's like having a chef who has tasted every spice in the world; they can guess the flavor of your new dish with very little help. This is a huge win for researchers who don't have access to massive amounts of data.

Second, there is Model Reuse. Why reinvent the wheel? If a model already exists that is good at understanding medical images, you can just use it. This saves time and effort, leading to Research Speed and Facility. Medical research is hard; it requires strict rules, ethical approvals, and complex data collection. Using a pre-made tool lets scientists skip the heavy lifting of building the engine and focus on driving the car.

Third, there is Model Recognition. If you use a famous model that everyone knows, your work gets cited more often. It's like using a well-known brand name; people trust it, and it's easier to explain your work to others.

The Case for Cooking from Scratch (Training Your Own)

However, the paper finds that many scientists argue against using these giants, and for good reasons.

The biggest issue is Explainability. If a master chef makes a soup and says, "It tastes good," but you can't ask them why they added that specific herb, you might not trust it. In medicine, trust is everything. Doctors need to know why an AI thinks a patient is sick. Foundation models are so huge and complex that even their creators often can't explain exactly how they make a decision. If you build your own model, you can design it to be transparent, making it easier to earn Clinical Trust and meet Regulation rules.

Then there is the problem of Speed and Memory. Foundation models are like giant, heavy trucks. They are powerful, but they eat up a lot of fuel (energy) and need a huge garage (computer memory). If you just need to deliver a small package, a tiny, efficient car (a custom model) is better. The paper notes that foundation models often have "dead weight"—parts of the brain that aren't used for your specific task but still take up space. If you care about Environmentalism (saving energy and water) or if you have limited computer power, training a smaller, custom model is often the better choice.

There is also a value called Research Sovereignty. This is about fairness. If only a few rich countries or big companies own the giant master chefs, then scientists in poorer countries can't really participate unless they are allowed to use those specific tools. The paper suggests that while Foundation Models could help level the playing field by giving everyone access to a powerful tool, there is a risk of "infrastructure lock-in," where everyone becomes dependent on a single provider. Building your own models allows for more independence.

The Tricky Middle Ground: Fairness and Novelty

The paper highlights that some values are messy and don't fit neatly on one side.

Take Fairness. You might think a giant model trained on millions of people is fairer because it sees more diversity. But the authors point out that most of these models are actually trained by single centers on single datasets, which might still be biased. So, using a giant model doesn't automatically fix unfairness; it might even hide it. The paper suggests that ensuring fairness might actually require more control over the model, which leans toward building your own.

Then there is Model Novelty. In science, people love to discover something new. Using a pre-made model feels like using someone else's invention. If you want to be famous for a new discovery, you might want to build a new architecture from scratch. However, the paper notes that "novelty" often means making things more complex, which clashes with the desire for simple, energy-efficient models.

The Final Verdict: It Depends on What You Value

The paper concludes that there is no single "best" answer. The choice depends entirely on what the researcher values most.

• If you value speed, data efficiency, and ease of use, you will likely choose a Foundation Model.
• If you value explainability, environmental impact, and total control, you will likely train your own model.

The authors suggest that the scientific community needs to be honest about these trade-offs. We can't just say "AI is good" or "Foundation Models are the future." We have to ask: "Good for what? And at what cost?" By understanding these underlying values, scientists can make better decisions that aren't just about what works technically, but what works for society, the environment, and the future of medicine. The paper doesn't claim to have solved the problem, but it provides a clear map to help everyone navigate the debate.