The gap in human-agent communication

Over the last few weeks, I’ve noticed a trend in assistant design: language models are being shaped to communicate in a way we can actually take in. A lot of hot new services like Instinct, Poke, Muse, Grok Bot and Dots are trying to mimic the quick, natural rhythm of texting. One of Opus 5.5’s most widely praised improvements was how much more natural its responses felt. I also recently came across Echo by Fulcrum, a writing model that’s great at matching human styles.

Most frontier models still aren’t good at naturally communicating their thoughts to us. I find it an annoyance to read through the verbose, complicated outputs of coding agents. They surface a lot of stuff I don’t want and miss what I would’ve wanted. Think about what you’d get back if you hired a capable human to do the job. It’d be three lines and maybe a question.

It’s hard to tell if the issue is the model misunderstanding what I want or something ingrained in its vernacular. If it’s misunderstanding, that’s a much harder problem. A solution would almost need a personalized agent that keeps learning from you. If it’s ingrained, we might be able to see traces of it in the pretraining data.

High quality data isn’t all you need

If models picked up this way of talking from what they read, it should show up in what they read. Frontier labs don’t publish their data mixes, but NVIDIA published the one for Nemotron 3 Ultra. Here’s a rough estimate of that mix, sorted by writing format. It’s built on a lot of assumptions and guesses, put together by Astra:

Estimated pretraining formats · Method (Percentage of tokens, classified by each whole training item’s format. Astra classified a random sample of about 14,000 items from NVIDIA’s public pretraining data., Very rough, low confidence)Percentage of tokens, classified by each whole training item’s format. Astra classified a random sample of about 14,000 items from NVIDIA’s public pretraining data.Very rough, low confidenceHoverTap a slice

52.1%of tokens

• Reference & explanatory documents22.6%
• Questions, answers & worked tasks29.5%
• Code files & formal source12.1%
• Articles, blogs & stories16.8%
• Lists, records & other formats14.3%
• Social posts, comments & reviews3.6%
• Dialogue & speech transcripts1.1%

Reference material + worked tasks

Written for a stranger who needs everything explained.

Roughly half of the estimated mix.

Reference & explanatory documents

Teaches the model to define every term, and to assume the reader knows nothing.

Wiki entries, papers, manuals and how-to guides.

Questions, answers & worked tasks

Teaches the model to show every step, then restate the answer.

Exercises, solutions and task responses. Embedded source passages and code count toward this category.

Code files & formal source

Teaches the model to be exact, because every token has to do something.

Source code, configuration and formal code. Repository prose and data records are classified separately.

Articles, blogs & stories

Teaches the model to have a voice, and a point to make.

News, essays, personal blogs, opinion pieces and stories.

Lists, records & other formats

Teaches the model to break everything into bullets, tables and headings.

Extracted fact lists, catalogs, tables, structured records and fragmented or mixed content.

Social posts, comments & reviews

Teaches the model to keep it short, for readers who already saw the thread.

Public posts, comments and user reviews. Worked question-and-answer threads count under tasks.

Dialogue & speech transcripts

Teaches the model to take turns with someone who already knows the context.

Conversational exchanges, interviews and speeches. Task conversations count under tasks, so this is not the total conversation share.

Reference material + worked tasks, 52.1% of tokens. Written for a stranger who needs everything explained.

About half of it is material written to be complete: documents and worked answers where someone (or a synthetic rewriter) spent real effort covering every detail for a reader they’d never met. That’s great data. It just teaches models to write for a stranger, when what you want is a collaborator who already knows the context.

Chasing this purer form of communication isn’t easy. The data that would teach it best (i.e. text messages between people) usually contains sensitive information and is understandably held to high data security standards. But hard problems haven’t stopped this industry yet. Opus 5.5 was a great step, and the next few generations will hopefully close more of this gap than people expect.