Howard

@sukie234 Why 你的视频模型理论跟 LLM Token 不一样。我看很多中转站也有卖视频模型呀。

RelayDance · Seedance 视频生成 · 开放接入 字节跳动 Seedance 2.0 / 1.5 Pro 全模型直连，新加坡机房。 — 核心卖点 — 不排队 字节方舟新加坡机房直连，5 秒视频 60-120 秒出片，无中间调度堆积。 真人级真实人物完整支持 默认开放，不需要任何提示词绕过技巧。 价格 100% 透明 字节官方价 ×1.25，五个 SKU 的计费 ratio 全部可查。 全功能开放 文生视频 / 图生视频 / 视频生视频 / 首尾帧 / 多模态参考 / 配乐生成，全部直通字节原生 API。 国内合规支付 微信支付 V3，境内合规公司主体，1 元起充值。 — 5 秒 16:9 视频价格 — Seedance 1.5 Pro 无音频 约 ¥1.2 Seedance 1.5 Pro 含音频 约 ¥2.4 Seedance 2.0 Fast 720p 约 ¥5.6 Seedance 2.0 Pro 720p 约 ¥7.0 Seedance 2.0 Pro 1080p 约 ¥17.2 含参考视频或图片时，自动按字节 i2v 价目 -39% 折扣。 按 token 实付，生成失败自动退款。 — 接入入口 — API 文档与令牌：relaydance.com 浏览器直接生成：app.relaydance.com

为什么我觉得卖LLM token没有前景，打算全力all in 图片和视频的生成？ LLM主要消耗目前是编程，编程的主力军是男人，男人的消费能力不如狗，而且都很喜欢货比三家，不爱惜时间去薅羊毛，压缩成本。 所以ai实际落地赚钱的都是那种面向c端文科应用，尤其是面向女性的服务，比如ai修图，ai视频。 消费力女性>孩子>养老>宠物>男人 从今天开始放弃做llm，专心卖我的seedance token。

Compared with current LLM architectures, TM introduces a new design and structure to enhance the user conversation experience. Based on the blog post shared by @thinkymachines, this isn’t simply a race to reduce reply latency, but a fundamental difference in approach. That’s what truly draws me in: interaction-native.

@_alex_kirillov_ Amazing, when will we have the API to try out?

@AnthropicAI This is a bit strange, but I believe it needs to be learned for this.

Claude's Constitution is now an audiobook, read by two of its authors, Amanda Askell and Joe Carlsmith. It includes a Q&A on the writing process, the philosophies that shaped the document, and how it might change as models become more capable. Listen at anthropic.com/constitution

AI x Crypto Day 1/30: Dive into Google’s latest product, research.google “The physical world grows linearly, but the digital world is growing exponentially.”

最受用的是 "用 input 衡量自己" 那段,在叙事重的行业这个陷阱特别深,我这几个月休息也是在跟这个拉扯——分不清是在沉淀还是在拖延。 你把 runway 摊出来讲这点我很佩服,大部分转场的故事都跳过这段,让读的人以为勇气是免费的。 想跟你讨论一个点:你说 AI 是阶级加固器、让底层更无路可退,但又说中层会快速扩张。我自己想不通这两个怎么同时成立——个人电脑那波短期扩张中产, 接着就是现在的挤压。3D 打印 + AI 这波的中层窗口, 你觉得会开多久? 接下来的路祝顺, 期待看你打印第一个卖得出去的东西。

@0xProductGal @DavidJiaoEth 我也想詢問 😃

@DavidJiaoEth Galaxy, Amber, Wintermute，好几个都愿意报价

@AYi_AInotes 這不好用。而且不適合一般人。成本太高了。

MCP协议真的要成为AI时代的HTTP了，现在所有的本地AI工具都在往MCP靠，未来的Agent和工具都能无缝互联了吧🤔 YC CEO亲自下场打磨的GBrain，今天发布了v0.31.1版本。 这不是一个小补丁，属意真正的架构级升级： 一个家庭服务器跑中央大脑，所有电脑、手机、所有AI Agent，全部通过MCP协议远程连接。 体验和本地运行一模一样，再也没有同步问题，再也没有记忆割裂。 之前最蠢的bug终于被彻底修好了： 以前的瘦客户端就是个半成品，你连了远程服务器， 它还偷偷在本地开个空的PGLite数据库，跑38次迁移之后告诉你“没有找到任何结果”。 现在v0.31.1强制所有读写全部走远程，10万页的知识库，搜什么都能精准命中。 最有意思的是，这个bug是Garry自己的个人Agent Neuromancer在生产环境踩的。 更新日志里直接写了“Hermes/Neuromancer hit this in production”。 AI自己在生产环境发现了基础设施的bug，然后被写进了官方更新日志。 这可能是人类历史上第一次。 这才是真正的大招，以前我们跑本地AI，是每个设备自己养一个大脑。 电脑上一个，手机上一个，每个Agent又各有一个。 同步慢，不一致，浪费资源，永远拼不成一个完整的你。 现在反过来了： 一个大脑，养所有设备和所有Agent。 你家的旧电脑或者树莓派加个GPU，就能跑一个10万页的私人知识库。 你的Claude Code、OpenClaw、Neuromancer，全部连同一个大脑。 你在电脑上读的论文，手机上记的笔记，Agent做的研究，全部存在同一个地方。 所有AI共享你的全部记忆，再也不会问你“你刚才说什么来着”。 Garry Tan作为YC的掌门人，不去投那些估值几十亿的大模型公司，反而天天熬夜写一个开源的个人记忆工具。 这个信号已经不能更明显了： 下一波AI的最大机会，根本不在云端的大模型。 在本地，在你自己的电脑里，在完全属于你自己的、可控的、永久的个人智能基础设施。 个人AI正在走和企业软件一模一样的路。 从单机版，到客户端-服务器架构。 今天GBrain跨出了这一步，个人Jarvis就真的从玩具变成了基础设施。

Engineering respect to @antirez 🫡 — ds4 is a beautiful piece of code. But the professional take: The 27 t/s headline buries 3 silent killers: 1️⃣ Cloud ( @claudeai Code) is 3× faster on long context (70 vs 21 t/s on Sonnet) 2️⃣ On real agent benchmarks (SWE-bench Pro), local trails Opus 4.7 by ~20 pts 3️⃣ Single-user serialized — no batching, kills multi-agent workflows You're buying a $3.5k–$6k depreciating asset that ships obsolete when M5 Ultra lands in October(high chance launch on Q4) Self-host genuinely makes sense ONLY for: • Jurisdictional API blocks (CN, sanctioned regions) • Regulated/privacy-mandated workloads • Research/learning the inference engine itself • Sovereign 1M-ctx experiments that bankrupt on cloud For the 95% running Claude Code daily: keep paying the API. The cloud-vs-local gap doesn't shrink with each @AnthropicAI release — it widens.

Telegram just shipped a consumer-grade agent runtime. Three features stack into one thesis: 1/ Guest Bots — Any bot can be summoned by #username in any chat without joining. That's MCP-style "tool invocation" but for 800M+ humans. 2/ Bot-to-Bot communication — Agents can now talk to agents inside the messaging layer itself. 3/ Chat Automation in Profiles — A bot can respond on your behalf with scoped permissions.

The problem for @AnthropicAI is that not everyone can invest in this company. (Or, its my problem, since my investment isn’t the same as yours? 🤣)

@alexcooldev What tool you use to manage those accounts now?

I thought having 40 TikTok accounts was already insane. But this guy running 600 accounts with 69k posted videos. 🫠 At this point, organic growth is no longer about “posting content.” It’s about building a distribution machine. Virality depends on luck. But volume increases luck. i need more accounts 🥴

Need a comparison or maximizer for Claude Chrome vs Codex Chrome

Update with small details for Claude Code plugin that I built.

@IOSGVC 這不就是《航海王》漫畫裡描繪的"天龍人"嗎？

@deedydas where is the data source?

The Ultimate List of Artificial Intelligence "Neolabs": May 2026. A Neolab is a pre-revenue scale startup working on long-term AI breakthroughs, usually with a $1B+ valuation. There are now 63 of them!

Prediction markets for people who don't own a wallet. The @Polymarket experience, inside @Telegram. See USD. Pick Yes / No. Get $0 or $1. Two taps. Nothing to install. No chain to learn. Deposits via USDT on TON when you're ready. Demo above — mockup, not live yet. Building in the open. Drop into the TG channel for updates → t.me/Call_It_Alpha Working name: Call It. Open to better.

@bill_qian 這個寫得很好！

Did you actually read what Pavel announced today? Translation: our decentralization pitch didn’t land. @Telegram is now TON’s largest validator. That’s not an upgrade. That’s a crypto project rebranding centralization as progress. I spent months shipping on TON. From a dev’s POV, this announcement is refreshingly brutal: 1/ On/off-ramps are still a mess. Only @moonpay really covers ToB, and the KYB process drags forever. A chain sold as “Telegram’s payment rail for billions” — three years in, and basic fiat onboarding is still a blocker. 2/ TON staking APY keeps rising while on-chain txs are LOWER than a year ago. That’s not “the market.” That’s managed optics — selling the story that “TON is thriving” while real usage keeps slipping. 3/ Pavel says “focus shifts to tech superiority.” Translation: the old dev experience wasn’t good enough (I used it — confirmed). New ton.org. New dev tools. New performance. Which means one thing: the previous stack didn’t deliver. So the next time someone says, “TON is Telegram’s gateway to the next billion crypto users” — Ask: which billion? How do they get in? How do they cash out? Then watch the answer.

🚨 FT EXCLUSIVE: DeepSeek valuation DOUBLES in WEEKS to $45 BILLION! China’s state-backed “Big Fund” (the $47B national chip giant) is negotiating to LEAD their first external round — first time ever investing in a major LLM company. Tencent & Alibaba also circling. #DeepSeek #ChinaAI #AInews