We should teach machines the Feynman Technique!

On the
decreased
communication
abilities
of frontier
LLMs

On the decreased communication abilities of frontier LLMs

Frontier models, while working on complex tasks, are producing answers increasingly harder to read, using a more complex and opaque language. People have been increasingly complaining or making fun of this. Here's a few rigorously cherry-picked examples:

• "Quintile" instead of "group of five" (Opus 5)

• "IV" instead of "initialVerification" (A recent Claude)

• An impossible dense sentence (a recent Claude): "Sign-off is per-decision on the canonical structure; tiers are projections gated by an automated caveat-drop/entailment check at export."

• GPT 5.6 Sol does not hold back: "failure-atomic create with credential-first ordering + rollback, recursive redaction, temp-rename publish, per-agent save/list generations, sparse-only persistence.". These terms may help the model keep track of its work, but shouldn't reach the user in this form!

Then, people are seeing this as a problem to solve. Yesterday I found a clever and effective idea on X by Andrew Carr that makes the problem even more visible and ironic: he made a skill to fix the jargon, with another "jargon". The skill rewrites the answer into "ASD-STE100" Simplified Technical English, a specific language originally designed to simplify technical manuals. It relies on short sentences, active verbs, and words with only one clear meaning. There is something funny about borrowing a language standard created to prevent mistakes in aircraft maintenance just to make sense of our most advanced language models!

It would be easy to dismiss this as a prompting problem. You can tell a model to use plain English, but the default still matters. The direction models take when we do not steer them is interesting in itself. The interesting part is that clarity seems to be declining while capabilities improve. Newer models solve harder problems, but often explain them less plainly. Earlier this year, in the generation after the big jump of December 2025, I think they had the best communication abilities.

What could be the reason? I'll give a possible interpretation. To work on a large problem, you have to compress it. You notice that several details belong together. You give the group a name. You can then think about the group as one thing instead of ten separate things. A beginner sees many facts, while an expert sees a few larger shapes. A specialist can use one word to refer to an idea that would otherwise take a paragraph to explain, and when two specialists talk, this saves time. The problem begins when the specialist forgets that the word is compressed and the listener gets lost. In human experts this is called "the curse of knowledge".

Great teachers can reverse this process. Richard Feynman is famous for this (and his Feynman Technique), but he clearly exercised his teaching abilities constantly. I wonder: are frontier labs trying to mitigate this? What's happening is creating a new kind of gap between benchmarks and user experience. A model can solve a harder task and still feel worse to use.

The change may come from training on synthetic text, rewarding professional language, optimizing agents for long tasks, asking them to compress more work into fewer tokens. If you work in the field you know that the last training phase, Reinforcement Learning with Verifiable Rewards, works well in domains where answers can be checked objectively. This is one of the main reasons LLM capabilities have continued to scale over the past year. But clarity of writing (and more broadly, taste) is qualitative, it cannot be verified this way.

Maybe models are improving faster where success is easy to measure, while other qualities are being left behind...