What I Mean When I Talk About a Machine’s Language

I talk to plants, I talk to my dog, and I spend a considerable amount of time typing requests to machines. The last one has its own peculiar frustrations. I can describe what I want, receive something beautifully organized that misses the point and then spend the next message explaining which part of “casual” apparently needed more attention.

That experience makes the idea of an AI whisperer appealing. There must be a better way to communicate what I mean. But when I started reading about a machine’s “language,” I found several different subjects being bundled together. Giving a model instructions, studying its internal calculations and training separate agents to send messages are related questions with very different answers.

Conceptual illustration of two weathered robot props facing one another in a workshop; the article discusses software research

The robot images are conceptual illustrations. The research discussed here concerns software models and agents.

The instructions I can actually improve

When I’m using a writing tool, I want to give it enough context to make a useful attempt. “Make it sound like me” leaves a lot unresolved. I get more specific when I explain the reader, provide an example of the tone and say what needs to change. For a blog rewrite, that might mean keeping the personal anecdote, explaining the science plainly and replacing a long list with connected paragraphs.

I’d then read the result against the request. Did the useful details survive? Did the tool introduce a new claim? Does the piece sound like something I’d actually say? The next instruction can address the part that failed. I find that much more practical than hunting for a phrase that supposedly unlocks perfect answers.

Clear instructions give me a way to evaluate the response too. If I ask for three short paragraphs, I can see whether I received them. Checking a scientific explanation takes more work because fluent writing doesn’t tell me whether its facts hold up. I still need the original research or another suitable source for that part.

The patterns inside a model

Anthropic’s March 2025 interpretability research looked at internal activity in Claude 3.5 Haiku. The researchers reported evidence of shared conceptual representations across different languages, along with examples of planning and cases where an explanation of an answer failed to reflect the mechanism they observed.

That makes “language” an interesting metaphor, but it needs context. Internal numerical activity is different from a sentence I can read. Studying it means developing methods to connect that activity to the model’s behavior. Anthropic also described limits to what its approach could capture, so I’d read the findings as a partial view of a particular model.

The detail that stays with me is how much work goes into interpreting even a small part of the process. Asking a model to explain itself gives me more text to examine. It doesn’t automatically give me a reliable account of the computations that produced the answer. That’s why the research interests me beyond anything I can learn by chatting with the tool.

When separate agents learn to communicate

A 2016 paper by Jakob Foerster and colleagues investigated agents learning communication protocols for cooperative tasks. Each agent had only part of the relevant information, and a limited communication channel let them share information to improve the joint result. The researchers examined the protocols that emerged from that setup.

I can follow the appeal of the experiment without imagining a secret conversation hidden inside every chatbot. The agents had a defined task, available signals and a training objective. A useful code for that environment can be a learned way to coordinate. It doesn’t need to resemble an ordinary conversation to do that job.

Meta’s 2017 negotiation research provides another example. Its agents negotiated over collections of items with different values. The team reported that training both agents could lead their exchanges away from ordinary language, and described training choices used to preserve language people could understand.

The detail I’d keep is the relationship between the objective and the behavior. If a task rewards reaching an agreement, keeping every exchange readable has to be part of the design too. The researchers’ account describes that training problem. It gives me a more useful explanation than the familiar story about frightened engineers pulling the plug on scheming bots.

What I want from the conversation

I’m interested in making these tools easier to work with, and in understanding enough to notice when a metaphor is doing too much. I can improve a prompt by being clearer about the work. Researchers studying internal activity need different tools, and researchers training communicating agents are asking another set of questions again.

If I wanted to spend more time with the technical side, I’d compare the syllabus of an introductory language-model book with the topics I actually want to learn. I’d be looking for examples I could understand and check, rather than promises of a secret shortcut.

The ordinary work of communicating well turns out to be useful here. I can state the task, show what I mean, read the result and ask for a specific correction. There’s plenty to learn in that process. My article on artificial pain sensors follows another familiar word into the engineering behind it, which is usually where I find the better story.

Michele Edington (formerly Michele Gargiulo)

Certified Sommelier, Certified Specialist of Wine, WSET Diploma student, writer & storyteller. I blend wine, science & curiosity to help you see the world as strange and beautiful as it truly is.

https://www.michelegargiulo.com/bio-page
Previous
Previous

Octopus Cities: The Underwater Metropolises of the Future

Next
Next

Hot Dogs to Millions: The Wild Entrepreneurial Journey of Sam Parr