In Defense of the Prompt

A recent post from Thinking Machines and Bridgewater made a strong statement:

"An explicit prompt can only convey the intuition an expert is able to put into words, while the judgments that matter most are often the hardest to articulate. Fine-tuning sidesteps this: rather than contorting the expert's intuition into a static prompt, the training process lets the model develop its own judgment."

We think this claim is both unjustified and dangerous.

The authors present an example where fine-tuning works well, but never justify the much broader statement that expert judgment is better learned implicitly than expressed explicitly.

One of the biggest advantages of LLM-based AI, especially when compared to older machine learning techniques, is that prompts are readable and interpretable. By replacing prompting with fine-tuning, model builders open a can of worms: overfitting, poor generalization, false causality, and a much harder debugging problem when something inevitably goes wrong.

As an example, consider the task of distinguishing pictures of dogs from wolves. A reasonable human might use features like jaw shape, fur, or eye placement to make a decision. Researchers found in a now-famous 2016 paper that an AI system had instead learned to look for snow in the background of the image. Many of the wolf photos happened to be taken in snowy environments, so "snow" became a surprisingly effective shortcut.

]This is a subtle failure mode. The model learned something that seems to work, but wasn't actually the concept it was supposed to learn. The same thing can happen with LLM fine-tuning. The model is free to learn whatever patterns best explain the training data, whether or not those are the patterns a real expert ought to use.

Consider the task Bridgewater describes in their post: determining whether a news article is relevant to a macro investor.

Suppose the model learns that articles mentioning “Jerome Powell” are unusually likely to be macro-relevant. During his tenure as Chair of the Federal Reserve, this was a perfectly reasonable statistical pattern. However, now that there is a new Fed Chair, this model likely will miss important content about the new Chair, Kevin Warsh.

Say the system instead learned explicit prompts, like:

Articles mentioning Jerome Powell are relevant to the economy.

An economist could read this statement and note: "That’s not exactly right, it should say that articles mentioning the current Fed chair are relevant." They can then update or improve the prompt, fixing the root-cause of a potential issue before it leads to a bad business decision.

Documenting a complex thought process is time-consuming work. Therefore, the real alternative to fine-tuning isn’t manually crafting prompts, but automatically optimizing in the prompt space. Frameworks like DSPy have become increasingly popular as convenient ways to learn and optimize prompts based on labelled examples and human feedback. These frameworks still allows algorithmic learning, while keeping the results in a form humans can inspect, challenge, and improve.

We’d recommend reaching for these techniques first, rather than jumping to fine-tuning. Thinking Machines and Bridgewater were vague on how they optimized their “expert prompts” but doing so effectively is an active area of AI research: prompt optimization methods like GEPA outperform fine-tuning on a variety of difficult tasks.

At Sphinx, we've taken this idea further. Rather than optimizing a single prompt, we automatically learn and maintain thousands of pieces of interconnected knowledge from enterprise work products collected across systems of record. But in many ways, these are a large collection of optimized prompts.

The important difference is that they're still written in natural language. A subject matter expert can read them, edit them, approve them, reject them, or remove them entirely. If the system starts learning something questionable, people can catch it before it becomes institutionalized.

There's another advantage too: future-proofing your work against the rapid pace of change in AI.

A fine-tuned model is tied to a specific base LLM. If next year's LLM is twice as good or ten times cheaper, much of the work has to be repeated.

Knowledge written in natural language doesn't have that problem. While there may be minor stylistic changes needed for prompts, natural-language learnings can be used with Claude, GPT, Gemini, open-weight models, or whatever comes next. Improvements in AI are automatically captured without additional effort.

This isn’t an argument against fine-tuning in general. Fine-tuning is an excellent tool for certain well-defined tasks, very stationary objective functions, and for capabilities that genuinely can't be expressed in language.

However, when the objective is capturing expert judgment, making that judgment explicit is one of the biggest advantages LLMs have over previous generations of machine learning. Abandoning this to chase a flashy technique and optimize questionable metrics is an unfortunate step backwards.