粉丝与涨粉
快照 11月4日 · 7 日榜 273 个账号可排值得借鉴
- 近 120 天带图或视频占 73.6%、回复占 0%，形式以图文为主。
- 每千曝光涨 1.4 粉，同粉丝档（5–20 万）中位 16.8，低于中位。
- 日均 9.4 条、浏览中位 806，这个节奏可以照抄。
近 7 日代表帖浏览前 5 + 互动率前 3，最多 10 条
- 1
HOLY SMOKE, BUILDERS ARE GONNA LOVE THIS ONE 🤯 @hypescribe has MCP built in, so you can: > search recordings > read transcripts > even start new transcriptions from Claude, ChatGPT, Claude Code, Codex or Cursor. and more plus API keys for your own scripts ↓
- 2
WAIT... OPEN WEIGHTS FOR AN AI SCIENTIST?! @chenru_duan and Deep Principle just dropped Hydrogen-1.0. It's pretty mighty, built on GLM-5.3, post-trained on lab records and discovery trajectories. Chemistry, materials science, experiment planning.. .. and yes!! you can build on its weights 👀
- 3
NO F*CKING WAY. 9X ARR GROWTH. 100X USAGE GROWTH. @traversal_ai just raised $120M at a $1B valuation. AI writes the code. Traversal understands production, finds incident root causes and helps prevent the next failure. Nubank, PepsiCo and Harvey already use it 👀
- 4
THIS THING IS STRAIGHT OUT OF SCI-FI 🤯 type at 200+ words/min without a keyboard or saying a word Just a pair of glasses.
- 5
every software engineer after reading this:
- 6
AI AGENTS ARE COMING FOR THE DRAWING BOARD @snaptrude goes from a brief to a detailed building model, plumbing, electrical and drawings. It shows you the plan first. You approve, steer or take over 🔥 very impressive.
- 7
Most OCR models come as-is. If they can't handle your old scans, tough luck. @OpenBMB ships MiniCPM-OCR with open weights *AND* fine-tuning code. Now even small archives and university libraries can adapt the model to their own documents. Smart move.
- 8
THIS 4B MODEL JUST MADE LONG-CONTEXT MEMORY WAY MORE INTERESTING. 256K context. A fixed 72 MiB KV cache. No cache growth as context gets longer. Meet EdgeRock-4B, the new open-source model from Rock Matrix. Here's why it matters. Traditional LLMs need more KV cache as context grows. EdgeRock takes a different approach. It combines recurrent state, local attention, and Rock Memory Anchors (RMA), preserving learned KV records in a fixed-capacity bank. This lets the model retrieve information from much earlier in the context, even after it leaves the local attention window. Now, look at the reported results: → 6.42× the output throughput of Qwen3.5-4B. That's 12,124 tokens/s at 512 concurrent sequences on a single 80GB Hopper GPU. → 99.1% less KV cache at 256K context. Just 72 MiB per sequence, constant from 4K to 256K. → 233.7 tokens/s during long generation, even with 128K input tokens followed by 128K generated tokens. → 82.9% mean pass@1 on AIME 2024–2026. The strongest evaluated bounded-state baseline reaches 18.8%, including comparisons with 14B-class models. And that's the real breakthrough. Instead of growing KV cache with context length, EdgeRock keeps it fixed while retaining access to selected older information. Long context. Constant KV cache. Strong reasoning. All in an open-source 4B model.
同期站内报道里提到它
近 7 天的公开报道里没有按名字提到它的。
本周涨粉观察
全部观察每周一生成；中文区、7 日涨粉超过 1000 的账号会有条目。
发布节奏
最近 12 周每周中位浏览
最近 12 周发布时间
颜色 = 中位浏览爆款 vs 日常
爆款 = 浏览前 10%| 指标 | 爆款 | 日常 |
|---|---|---|
| 帖子 | 33 | 296 |
| 带媒体 | 51.5% | 76% |
| 回复 | 0% | 0% |
| 平均字数 | 280 | 265 |
| 中位浏览 | 3,514 | 761 |
打法建议
只陈述相关性- 互动最好的时段是 01:00 前后（北京时间），中位浏览 1,299。
- 一周里 周五 的中位浏览最高。
- 形式上 纯文字 的中位浏览更高（775 vs 1,215）。
- 爆款里带媒体占 51.5%、回复占 0%；日常分别是 76%、0%。
- 当前节奏约 9.4 条/天，日均浏览 1,523。
内容结构
按本站分析的分类（近 120 天）还没有分类分析。
形式与语言
帖子占比- 视频125
- 图片117
- 纯文字87
回复占比 0% · 引用 55.9% · 带媒体 73.6%
互动最高的帖子
按浏览排序- 1
HOLY SMOKE, BUILDERS ARE GONNA LOVE THIS ONE 🤯 @hypescribe has MCP built in, so you can: > search recordings > read transcripts > even start new transcriptions from Claude, ChatGPT, Claude Code, Codex or Cursor. and more plus API keys for your own scripts ↓
- 2
WHAT?! Unreal Labs published an awesome playbook on building cloud agents 1M+ agents launched. One year of battle-testing. @unreallabsai breaks down the stack, security choices and integration headaches. super practical lessons for anyone taking agents into production 👀
- 3
AI AGENTS CAN BUILD YOUR APP. MO CAN HELP YOU FIND OUT WHERE IT BREAKS. Mo from @momenticai explores any app using a URL, a testing brief, and test accounts if needed. No test scripts required. It reports test cases, verdicts, coverage gaps, and the bugs it finds. Every bug report gets console/network logs, detailed reproduction steps, and a video replay. That part excites me: you get evidence to investigate, beyond a list of suspected issues. I built http://llmexplorer.org in 2025 as a fun project and am planning a refresh, so I put it through Mo. In one single run: → Multiple agents explored pages in parallel, including search filters and model listings → They tested external links and UI edge cases, completing 55 test cases → The audit surfaced 8 reported bugs, including broken links, clipped organization chips, and filter-trigger issues After fixing a bug, I can use the reproduction steps to write a regression test in Momentic and run it locally or in CI. Explore → reproduce → fix → add regression coverage. Mo doesn’t guarantee a bug-free app, but it gave me a concrete starting point for the refresh. Watch the demo below. Seeing the agents work through my own site and surface actionable findings is pretty great 👀 Try Mo in the first reply 🧵↓
- 4
WAIT... OPEN WEIGHTS FOR AN AI SCIENTIST?! @chenru_duan and Deep Principle just dropped Hydrogen-1.0. It's pretty mighty, built on GLM-5.3, post-trained on lab records and discovery trajectories. Chemistry, materials science, experiment planning.. .. and yes!! you can build on its weights 👀
- 5
MY PREDICTION FOR 2028: Most creatives will be managing teams of AI agents before they ever hire another human. Crazy? Maybe. But I can see it happening. What's your prediction? Share it in @contraben's thread🧵 and you could win a MacBook Pro 👇
- 6
MEET YOUR AI MARKETING DEPARTMENT. → Strategist researches your niche → Creator makes videos → Analyst tracks conversations → Manager plans and publishes Freaking. cool. That's Muse from @Adant_ai. Four roles, one team, working 24/7 👇
- 7
NO F*CKING WAY. 9X ARR GROWTH. 100X USAGE GROWTH. @traversal_ai just raised $120M at a $1B valuation. AI writes the code. Traversal understands production, finds incident root causes and helps prevent the next failure. Nubank, PepsiCo and Harvey already use it 👀
- 8
THIS THING IS STRAIGHT OUT OF SCI-FI 🤯 type at 200+ words/min without a keyboard or saying a word Just a pair of glasses.
- 9
best use of AI I've seen this year
- 10
every software engineer after reading this:
- 11
PERSONAL AGENTS HELP YOU WORK. BUT WHAT ABOUT PLAY? Just found Yyo by @YorollAI, a personal agent site with a whole cast of characters to hang out with. You can chat, call, play games, or share your screen and watch YouTube together. game night has a new guest 🤘
- 12
i don't think people realize how big this is… A US lab just released a 40B open-weight coding model leading its class on three major coding benchmarks. 125+ tok/s on a MacBook. 150+ on RTX Spark. no per-token fees. no usage limits. your laptop is now the endpoint.
点赞、转发、浏览是采集那一刻的快照，不是帖子最终数据；时间为北京时间。KPI 卡片里的同比只在账号有 8 周以上数据时显示， 否则只看近 4 周。“打法建议”只说明数据里的相关性，不代表因果。