X open-source algorithm update: weights unchanged, video checks start at 64 likes
ThePeterMick · x · 2026-09-19
The author diffed two new releases of X's open-source algorithm code with Claude and found the scoring weights barely moved.
Key changes
- Both releases changed the weights file by exactly one line — the timestamp. Scoring logic is unchanged.
- New video moderation: once a video passes 64 likes, an AI compares it against reference frames to detect disguised or spliced explicit content; past 512 likes a deeper version of the same check runs. Flagged posts stop being shown to non-followers.
- The check also applies to videos you quote.
- X removed the alternative scoring model that sat disabled in the code, leaving exactly one scoring path: a weighted sum of predicted user actions.
- The Under the Hood report now shows whether your account or posts were withheld due to a legal demand and from which country, plus the effect (content warning, hidden from non-follower recommendations, hidden from logged-out and underage users).
Action list from the author
- Copy-link share remains the single biggest signal at 20.
- Originals beat replies and reposts; only originals reach non-followers.
- Replies between mutuals go from 5 to 20.
- A DM share is worth 5; early likes (1 and 32) unlock reach to strangers.
- A second post in the same feed drops to 62%; under 1,000 followers there is a free lift to roughly 16th place.
- Native video over 10 seconds is the line for video lists; converting a reader into a follower is worth 4.
- A pinned post linking to a poorly rated site labels your account for a week.
Avoid
- Hiding anything in a video; quoting videos you haven't watched; chain-replying under big accounts; picking up an account label (it cuts off non-followers entirely); pinning a post pointing at a badly rated site.
More from Infra
- Cloudflare saved 100TB of RAM by shrinking a consistent hash ring — blaizedsouza · 2026-09-19
- Higgsfield Open-Sources Fault-Tolerant GPU Orchestration for Trillion-Parameter Training — higgsfield-ai · 2026-09-19
- Cloudflare's Quiche: A Rust Implementation of QUIC and HTTP/3 — cloudflare · 2026-09-19
- Cache Hit Ratio Explained: The Formula and Why It's the Key Metric for Cutting Latency — blaizedsouza · 2026-09-19
- MLC's New GPU Execution Model Chapter Explains Blackwell Threads, Memory Spaces and TMA — blaizedsouza · 2026-09-19
- We burned 1.5B OpenAI tokens — here's the postmortem that cut our API bill 65% — Careful-Key-1958 · 2026-09-19