Everything / Models and research / LLM training
LLM training Cooling
On TutorHero's Everything Index, LLM training is cooling on GitHub: 0.03 of every 10,000 new public repos over the last three months (29 a month), -38% vs a year ago.
Since 2019, and the range to 2032
New repos per 10,000 (3-month average, log scale)12-month forecastRange to 2032: middle half and 90% of past similar lines
The range is not a forecast of this line: it is what GitHub lines in the same state (n/a) did 1 to 5 years later, 2019 to 2026, applied to today's level. Past 5 years it is stretched, not measured. Checked out of time (learned before 2022, scored after), 71% of outcomes landed inside the 90% range, so read it as wide. Open it in the Everything chart.
Week by week
New public repos with the llm-training tag each week (Monday to Sunday, UTC). The outlined bars, last week included, are still filling in, as tags are often added days after a repo is created, so their height will grow a little. "Against the field" divides the change by what the typical tag did, so a wave that lifts every tag does not count. Hover for the numbers. Fewer than 40 new repos in 4 weeks: read the change with care.
Built vs talked about
Each line is its own share (of all new GitHub repos, all Hacker News stories, all English Wikipedia views, all new stars on the repos we track), shown as a share of its own peak, so the timing and shape compare even though the units differ. Hover for the counts. Tap a name to hide or show a line.
What it means
- Learn the ideas more than the name: LLM training is cooling, -47% below its Jul 2025 high.
- Start by reading Human-Agent-Society/reef (+7.5k stars in 3 months) and bojieli/ai-infra-book (+5.7k stars in 3 months).
- People to follow: Jelly Lee, Human Agent Society, Volcengine.
- Room to build: attention +16% vs a year ago while the share of new repos went -38%.
- Attention: 0.30% of new stars on the repos we track, +16% in a year.
- Hacker News: 2 stories a month, -19% share in a year.
- Not investment advice: this measures what people build on GitHub, not any company.
- Most-starred tools here lately: Human-Agent-Society/reef, bojieli/ai-infra-book (vouched).
- Check that a tool is still maintained before you adopt it: the field is shrinking.
Repos carrying it now
People building it
Jelly LeeLLM/LLMOps，博客：吃果冻不吐果冻皮
Influence 50
Human Agent SocietyBuilding engines for human-AI agent societies that collaborate and evolve.Influence 41
VolcengineAI agent context, search, and distributed training tools.Influence 74
0securityWe make software secure itself.Influence 3More in Models and research
Questions
Is LLM training growing on GitHub?
No, it is shrinking. Its share of all new public GitHub repos is -38% vs a year ago and -25% over the last three months (Sep 2026). Phase: Cooling.
How many new LLM training repos are created each month?
About 29 new public repos a month carry the llm-training tag (three-month average to Sep 2026), or 0.03 of every 10,000 new repos.
What will LLM training look like in 2032?
Not enough history for a range yet.
Not investment advice. Measurements of what people build and star on GitHub, a statistical forecast, base rates and a model's reading of them. Nothing here is about any company's revenue, valuation or stock.