Where practitioners spent their attention

Every Discord channel heading in this archive looks like this:

### **Unsloth AI (Daniel Han) ▷ #[general](…)** (1123 messages🔥🔥🔥):

AI News, a typical Discord channel heading

That number is not an editorial judgement. Nobody decided the conversation was worth 1,123 units of importance. A bot counted the messages that existed in a room on a day, and printed the count. Across the whole archive there are 31,607 of those channel-day records, covering 2,142,082 messages in 56 servers, from May 2024 to March 2026.

Everything else in this book measures what a publication chose to write about. This measures what people did, whether or not anyone wrote about it — and what people did turns out to be duller, more repetitive, and far more constrained by money than any announcement suggests.

Where the words actually are

Sort all 2.14 million messages by the kind of channel they were posted in and the result is lopsided enough to be worth staring at.

Five sixths of everything happens in a channel called general. A single channel name accounts for 1,397,089 messages — 65% of the entire corpus on its own. Structured help channels account for 3.2%. Research and paper channels account for 1.9%. Showcase channels, where people post the things they built, account for 0.6%.

If your mental model of a technical community is a place where people ask well-formed questions in the channel provided and receive well-formed answers there, the census does not support it. What it supports is that the structure is mostly decorative: the technical conversation happens, but it happens in general, unlabelled, next to everything else. That is a fact about where the talk sits rather than about its quality — the extract below is from an ordinary chat channel and it is as specific as anything in a help forum. The second-largest identifiable topic — after general, ai-discussions and off-topic — is hardware. hardware-discussion carries 46,273 messages, which puts it fractionally ahead of every channel named help in the archive combined.

Here is what that channel contains, on three consecutive working days in January 2025:

@octaviagoetia expressed a preference for using GPUs over online APIs, citing concerns over TOS restrictions and potential latency issues. @heychazza agreed, highlighting the benefits of using a beefy laptop for tinkering with LLMs without worrying about unexpected costs.

Users noted that running a 70B model requires more VRAM than available on common GPUs, leading to reliance on slower CPU inference. Recommendations included sticking to models under 16B with Q8 or Q6 quantization for better performance.

@octaviagoetia is currently using Llama 3.1 for generating quest lines, with plans to upgrade to Llama 3.2 1b, finding it more creative and efficient.

AI News, Discord recap, LM Studio #hardware-discussion, 2025-01-03 to 2025-01-07

Terms of service, electricity bills, VRAM ceilings, and someone generating quest lines for a game on a laptop. That is the ground floor of this field by volume, and none of it appears in any launch post.

The busiest room in the archive

Rank the 56 servers by total messages and the top of the list is not a frontier lab.

Unsloth AI is a fine-tuning toolchain, and it is the most active community in this corpus: 302,248 messages across 23 months, more than the OpenAI server and the HuggingFace server put together. It is busiest in exactly the period when the practice it supports had largely stopped being written about.

Two other entries are worth pulling out. LMArena and Cursor Community both appear in March 2025 and are immediately among the busiest rooms in the archive — evaluation and coding agents arriving as fully-formed communities rather than growing into them. And BASI Jailbreaking, which starts in November 2025, has the highest activity rate of any server measured: 19,062 messages per month, ahead of every established community including the ones that had been running for two years.

Two ways this census lies to you

The counts are mechanical, which makes them tempting to over-read. There are two specific traps in this data and both took me a while to see.

The identity key is a display name. Servers appear to die when they have only been renamed. Cursor IDE runs from November 2024 to April 2025 and stops; Cursor Community starts in March 2025 and continues. CUDA MODE ends in September 2024; GPU MODE begins the same month. OpenRouter (Alex Atallah) ends in August 2025; OpenRouter begins in August 2025. Nothing in the record links them. Take the birth and death dates literally and you will report a churn of communities that is partly a churn of strings.

The sample changes underneath the counts. This is the larger problem.

The number of servers sampled runs at about 30 for sixteen months and then drops to 23 in September 2025 and stays there. That single decision produces a cluster of servers whose last appearance is August 2025 — Cohere, LlamaIndex, Notebook LM, MCP (Glama) — and every one of them looks, in the raw data, like a community that died in the same month. They did not. They were dropped from a list.

The same decision inflates the other line. Messages per sampled server rises from 1,665 in May 2024 to 6,941 by February 2026, and some of that is real growth and some of it is the sample being pruned down to the busiest rooms. There is no way to separate the two after the fact.

What the practice layer looked like

Put the census together and the practice layer of this field, measured by message volume rather than by coverage, is remarkably consistent across two years. It is people in an undifferentiated chat room, talking mostly about what they can afford to run, on hardware they own, using models they downloaded — and the single largest such room in the archive exists to help them modify those models.

There is no version of that sentence that would work as a launch announcement, which is precisely why it is worth measuring separately. Announcements are about what becomes possible. This is a record of what practitioners were talking about while that was being announced — not a census of what ran anywhere — and the constraint shaping that talk was never capability. It was VRAM, terms of service, and the electricity bill.

Practitioner talk about a technology is usually duller than its coverage, more repetitive, and organised around what things cost.