Fixed
Smoother Desktop Filter Scrolling
Fixed content jumping when the desktop filters switch between expanded and compact modes, with steadier behavior near the transition point.
Fixed content jumping when the desktop filters switch between expanded and compact modes, with steadier behavior near the transition point.
The More Filters panel now closes with Escape, outside clicks, and breakpoint changes, then restores keyboard focus.
Currents grows from a content product into a complete, discoverable one: Agent Access and global Feedback are live, and site search no longer covers only the blog.
A new Agent Access page: five read-only tools, MCP connection configs, Skill installation, verification questions, and troubleshooting. Invite-only, read-only, bilingual.
A new standalone Feedback page: product issues, feature ideas, source suggestions, and Agent access problems — all anonymous. In-context content correction stays on detail pages, renamed to Report a content issue.
All Currents pages share a persistent left rail — Featured, All Updates, Hot Board, AI Daily, Topics, Favorites, and Agent Access are one click away. On mobile it folds into a drawer.
Site search no longer indexes only the blog: feature pages, topics, and the changelog are all searchable — searching MCP, Skill, Agent, or Feedback takes you straight to the right page.
The Hot Board moves from real-time aggregation to a persistent event layer: every event now has a stable identity, a lifecycle, and a bilingual summary. The main board gains an explicit quality bar, and low-confidence clusters move to a separate Watching section.
Withdrawn or invalidated reports now correctly leave their events and trigger a recalculation of heat and qualification, instead of lingering as stale content. Mistakenly merged events can now be split reliably once their semantics are corrected, with the original event identity always traceable.
Reports of the same story are merged into one event with a stable identity that keeps tracking new developments. Merges are traceable, and report updates refresh the event's real activity time.
The Hot Board now switches between All, News, Products, and Research, so you can focus on the kind of events you care about.
Clusters that don't yet qualify — single-source, aggregator-only, or lacking community traction — are listed under Watching, keeping the main board clean.
The main board now requires at least two independent reports, or one official source with significant community discussion. Paper listing entries (arXiv / HF Daily Papers) no longer qualify on their own.
High-confidence events get auto-generated bilingual summaries and progress notes. Heat combines independent report count with community traction, sampled every 3 hours.
Fixed repeated fetches from one source being counted as multiple reports. Different outlets referencing the same URL are now credited independently.
This batch moves Currents beyond a plain timeline — you can now browse by events and topics, and deep reading is completed with full-text translation. It also includes the Phase 7 layout refinement (three densities and a re-layered timeline).
A new Hot Board groups events reported by multiple outlets within 48 hours, showing each event's status, heat score, and distinct source count. When nothing qualifies, it automatically falls back to the last 7 days instead of showing an empty page.
A new Topic Map organizes 38 topics — OpenAI, Anthropic, agents, reasoning, robotics, and more — into company & model, technical direction, and content-type groups, each linking to a full topic feed.
Every topic now has its own page with a description, a full paginated timeline, bilingual SEO metadata, and sitemap coverage.
Detail pages gain an Original Translation tab alongside the AI summary and deep read. Featured archive entries have been backfilled, and new entries are translated automatically as they arrive.
The timeline now offers compact, standard, and comfortable densities (standard by default). Your choice is remembered in the browser and synced across tabs.
Both new and archived entries now carry fuller source text. Papers use clean abstracts instead of noisy scraped page content.
Recommendation reasons, tags, and multi-source info are re-layered; the desktop timeline breathes more and high-score items stand out at a glance.
Fixed read cards collapsing their titles in standard and comfortable densities — now only the summary folds away, and the title always stays visible.
Fixed detail pages showing untranslated category names and incorrect source or author labels.
The largest single-day upgrade since launch: Currents grew from a working MVP into a full AI news product.
Every item now has its own shareable, bookmarkable URL with full SEO metadata and structured data. Legacy overlay links redirect automatically.
Three top-level views — Selected, All Updates, and Papers — plus category, source, and minimum-score filters introduced in this phase, so you can browse at your own depth.
A daily digest is generated automatically: one headline plus five sections — models, products, industry, research, and tips & opinions — with browsable archives.
The timeline opens with one lead story and four supporting stories; the section hides itself automatically when there isn't enough material.
Items you've read are dimmed automatically (tracked locally in your browser) so you can focus on what's new.
Thirteen field-verified sources were added — official blogs from Google, Microsoft, NVIDIA, GitHub, and Cloudflare, plus international and Chinese tech media — with AI-relevance filtering on general outlets to keep noise down.
The archive was backfilled day by day to July 1, 2026 — over two thousand entries — so the timeline has no gaps.
Cards are now real links, so open-in-new-tab, middle-click, and other native browser behaviors work as expected.
Fixed the All Updates view silently hiding lower-scored items — it now genuinely shows everything.
Currents officially launched — from kickoff to production in a single day.
A brand-new AI news section: content is collected daily from arXiv, Hugging Face, Hacker News, and official research blogs, then enriched by AI into bilingual titles, summaries, a 0–100 recommendation score, a one-line reason to read, and an in-depth analysis. Multi-source coverage of the same story is merged and deduplicated automatically.
Full-text search in both Chinese and English, with category filters.
Bookmark items without an account — favorites are stored only in your own browser.
Chinese and English RSS feeds, with per-category subscriptions available.
Feasibility checks before the build began, paving the way for same-day launch.
Verified that the AI collection and processing pipeline could run safely on existing infrastructure — at zero added cost and with no impact on services already in production. The build began the same day.