WisGate

@Baidu_Inc The interesting number isn’t just “90% generated.” It’s what percent shipped cleanly after reviews, tool runs, and production fixes. Generated code is easy. Generated code that survives contact with users is the real benchmark.

The Miaoda App and Miaoda Enterprise Edition are here, giving more builders and businesses access to our coding agent! And the most interesting detail? 90% of Miaoda App's code was generated by Miaoda itself. Coding agents are making on-demand, purpose-built software commercially viable. To date, Miaoda-generated apps have served 10M+ users, with total application value reaching RMB 5B.

@Baidu_Inc The interesting number isn’t just “90% generated.” It’s what percent shipped cleanly after reviews, tool runs, and production fixes. Generated code is easy. Generated code that survives contact with users is the real benchmark.🤜🤛

@Baidu_Inc Daily active agents is a good top-line pulse, but we’d pair it with task completion rate, human intervention rate, and repeat invocation by the same team. A busy agent is not automatically a useful agent. Some are just very talented at failing often.😂

Baidu Advances Agent Portfolio to Embrace the Agent Era, Champions Daily Active Agents as Key Metric prnewswire.com/news-releases/…

@ClaudeDevs This kind of quota change matters less for “how many prompts” and more for what kind of work you offload. Refactors, test repair, and repo spelunking burn budget very differently. Teams should map usage by task class before they decide a plan got cheaper. #wisgate

Starting June 15, paid Claude plans can claim a dedicated monthly credit for programmatic usage. The credit covers usage of: - Claude Agent SDK - claude -p - Claude Code GitHub Actions - Third-party apps built on the Agent SDK

@opencode If people are jumping in for the free window, the best use is a 30-minute bakeoff: one greenfield feature, one ugly bugfix, one test-writing task, one tool-heavy task. The “free” part is nice. The comparable workload is the real answer.

OpenCode x Qwen 3.6 Plus - free, again Last time y’all treated our capacity like an all-you-can-eat buffet. We found more GPUs. Round 2.

For front-end work, we’d test three things before getting excited: text rendering accuracy, diff discipline in existing code, and whether the model can recover from a broken first pass without rewriting the whole page. That’s where a lot of “great model” stories get humble.

@SenseTime_AI @lindahua The useful part of a report like this is usually not the headline score. It’s training recipe transparency, inference tradeoffs, and failure modes. First things we’d check: context behavior, image edit consistency, and what breaks under messy prompt chains.

Led by our Co‑Founder and Chief Scientist Dr. @lindahua , our passionate AI pioneers have brought vision to life.🔥 Access the 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝗨𝟭 𝘁𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗿𝗲𝗽𝗼𝗿𝘁 for the architecture, training recipe, and innovations behind this breakthrough.

Separate agent credits are a useful signal: agent runs are infrastructure spend, not casual chat. Track cost per completed task, set model fallback rules, and keep one OpenAI-compatible path ready. @wisgate_ai can help there. Follow us for practical gateway notes.

@berryxia The encouraging part for smaller teams is that “more compute” is still not the only playbook. Architecture choices, data discipline, and training efficiency can still move the board a lot. The depressing part is you still have to be good at all three.

Moonshot AI创始人杨植麟最近放出了一个40分钟视频。 这位92年生、清华计算机本科第一、CMU博士、Transformer-XL和XLNet共同作者，前Google Brain和Meta研究员，坐在镜头前平静拆解了Kimi K2的整个训练过程。 他们只花了460万美元。 上周一场8模型实时编程大战，Kimi K2直接拿下第一，GPT-5.5排第三，Claude Opus 4.7第五。 我看完后最大的感受是，AI竞赛的规则已经在悄然改变。 所有人还在拼谁敢烧更多钱、堆更多算力，他却用极致优化、线性注意力、子代理这些硬核架构，把资源差距直接抹平甚至反超。 40分钟全是干货，零废话，把关键打法讲得清清楚楚。 如果你正在做AI代理，或者准备2026年入场大模型赛道，这段视频强烈建议存下来周末慢慢看。 小团队靠聪明架构，正在把大厂的传统玩法一点点颠覆。 你还觉得只有堆钱才能赢吗？

@vista8 @OpenSquilla Smart routing is great right until users can’t predict why answers changed. The fastest way to lose trust is hidden routing with no fallback trace. If teams expose route rules, cache hits, and fallback reasons, costs go down without the “why is this worse today?” drama.

前段时间小龙虾、Hermes爆火，一个特别大的痛点就是太烧Token了。 关于怎么省Token，很多人研究了很多方法，比如用qmd等本地语义搜索，换便宜模型等。 最近刷到一个开源项目@OpenSquilla，把省Token这件事儿做得很不错。 核心逻辑：智能模型路由 + 本地向量检索 简单问题，用便宜模型，复杂任务，用更厉害的模型。 智能路由本地完成，不消耗Token，换模型也是自动判断，不需要手动切。 后台还有模型调用成本统计，随时查看用了哪些模型，花了多少钱。 连续对话，让它写个抓取 Paulgraham 最新文章脚本，只消耗了5500 Token。 完成后会显示 COMBO ×2 ，像游戏的连击反馈，有意思，哈哈哈 相比完整重发，每轮只增量发送，缓存命中机制也实际传输 token 减少了 90%+ 记忆系统做得也不错，快到上下文上限时，子 Agent 筛除关键内容再压缩，支持BM25 + 向量混合检索。 自动整理白天对话，第二天也能记得上下文，让 Cron job 定时抓新闻、跑任务，很省心。 安全上也有考量，高风险工具跑在沙箱里，按来源直接不明工具、Skill调用。 支持 Openclaw 一键迁移，记忆、配置、技能全能移过来，切换零成本。 安装很简单，跟Claude Code或Codex说： 带我安装配置：github.com/opensquilla/op…

@DeRonin_ Our usual order is: cut context bloat first, route simple jobs down one tier second, then fight about model pricing third. If you need one OpenAI-compatible layer for routing and fallback, that helps too. Follow us for practical gateway notes.

Andrej Karpathy: "90% of your AI coding bill is paying for context you didn't need to send" Here are 10 things senior AI engineers stopped wasting tokens on: 1. Auto-context loading 50 files for a 30-line fix: $1.20/turn for tokens you'll never read. 80% input waste, every session 2. Running Opus on lint, format, and rename tasks: $0.60 for what Haiku nails at $0.02. 30x overpay on the cleanup tier 3. Tool call loops that re-send the full repo on every retry: 5x context cost per agentic flow. fixing these alone cuts 30-50% of bills 4. Sonnet as the default model: Kimi 2.6 matches its quality on most coding tasks at 1/6 the cost. defaulting to Sonnet in 2026 is leaving 60-70% on the table 5. Streaming responses on stable-prefix workflows: kills your prompt cache. you pay 10x for tokens that should have cost cents 6. "Just in case" file includes: 80,000-token prompts that should be 3,000. context bloat is the silent budget killer 7. Per-session knowledge rebuilding: 10 min writing a SKILL.md once vs paying agents to re-figure out your environment every run. $4 vs $0.30 per execution 8. Single-model setups: premium tier on every task is the most expensive mistake in AI coding right now 9. Asking 10 small questions one at a time: 10 separate input prefix charges vs one batched call. 70-90% savings on routine workflows 10. Buying Claude Pro + ChatGPT Plus + Cursor Pro: you seriously use one. the other two are habit, not utility what actually compounds instead: - context discipline (grep before fetching, always) - prompt caching on every stable prefix - multi-model routing (Kimi 2.6 default, Opus for the 10%) - graduated skills via SKILL.md files - profiling tool calls before optimizing prompts - the routing mindset (right model for right task) in 12 months, the gap between developers shipping on $200/month and $4,000/month budgets won't be skill it'll be how well they route study this.

@NousResearch Very interesting result if it holds up broadly. The thing teams should watch next is not just speedup, but whether downstream finetuning and eval stability stay boring. Infra wins are best when they’re powerful and a little disappointing to talk about later.

Today we release Token Superposition Training (TST), a modification to the standard LLM pretraining loop that produces a 2-3× wall-clock speedup at matched FLOPs without changing the model architecture, optimizer, tokenizer, or training data. During the first third of training, the model reads and predicts contiguous bags of tokens, averaging their embeddings on the input side and predicting the next bag with a modified cross-entropy on the output side. For the remainder of the run, it trains normally on next-token prediction. The inference-time model is identical to one produced by conventional pretraining. Validated at 270M, 600M, and 3B dense scales, and at 10B-A1B MoE. The work on TST was led by @bloc97_, @gigant_theo, and @theemozilla.

@AYi_AInotes A good product check here is simple: what exactly leaves the page on submit, and can your team prove it? “We only send limited identifiers” usually sounds less comforting once query content, page titles, and third-party pixels start sharing a room.

最近全网都在聊OpenAI的ChatGPT 5.5多厉害，Codex多好用之类的， 但没人注意到昨天刚爆的这个集体诉讼，这件事可能是真正炸穿底线的事。 南加州联邦法院昨天正式立案，原告代表所有美国ChatGPT用户起诉OpenAI。 诉讼文件里的实锤硬到爆， OpenAI在ChatGPT网站代码里直接嵌入了Facebook Pixel和Google Analytics， 你在输入框里敲下任何一个问题，按下回车的瞬间， 这个查询的完整主题会变成浏览器标签标题， 然后Pixel会把它和你的Facebook cookies一起，实时发给Meta。 那些cookies里包含c_user和fr字段，也就是你唯一的Facebook用户ID。 但这还不是最恐怖的地方， OpenAI自己在隐私政策里承认了这件事🌚 他们说他们只会分享有限的标识符用于推送Pro版广告，不会分享完整的对话内容。 诉讼方认为，查询主题本身就是最敏感的个人信息。 以前我们以为，免费AI的代价是你的数据用来训练模型，但其实模型只是诱饵，真正的产品， 是你每一次的好奇心， 和你完整的数字身份。 还有更讽刺的， 很多人用ChatGPT，就是不想被Google追踪自己搜了什么， 结果转头就把自己问的每一个问题，原封不动送给了Meta和Google🥹 #OpenAI #ChatGPT #隐私

@EvidenceOpen Shadow AI usually means one of two things: either the product solved a daily pain before IT caught up, or governance lagged behind real demand. The move for vendors is to preserve the speed while making auditability and admin controls catch up fast.

“We did the hardest thing in the history of American health care. We got the majority of American doctors to all voluntarily adopt a single technology platform.” NBC News on how that happened, what U.S. physicians actually do with OpenEvidence, and how partnerships with NEJM, JAMA, NCCN, and Wiley make it possible.

@hongming731 The clean default is still: isolated env, explicit approvals on side-effectful actions, and post-task logs you can actually read later. A surprising number of agent incidents are not model failures. They’re observability failures wearing a model costume.

For any new #codingagent: test three things before the honeymoon phase wins. Can it recover context, explain file changes, and stop before touching unrelated code? The last one is where the comedy usually begins.

@OpenAIDevs @AriX @romainhuet The first useful pattern here is boring on purpose: browser ops, admin backoffice work, QA repro, and file-moving glue. If teams start with “let the agent run everything,” they usually skip the approval model and regret it by Friday.

Computer use lets Codex work across your apps without taking over your Mac. @AriX talks with @romainhuet about what changes when agents can click, type, and keep working in the background.

@OpenAIDevs Hooks pay off when they enforce policy, not when they replace thinking. Good split: hooks for secret scans / evals / audit logs, agent logic for the task itself, CI for final enforcement. If all 3 collapse into one layer, debugging gets weird fast.

Codex is getting easier to automate and customize around your code. 🪝 Hooks customize the Codex loop with scripts that run at key points in a task: • Run validators before or after work • Scan prompts for secrets • Log conversations to internal systems • Create memories or customize behavior by repo or directory ⚙️ Programmatic access tokens provide scoped credentials for Business and Enterprise teams: • Create tokens from ChatGPT workspace settings • Use them in CI, release workflows, and internal automations • Set expirations or revoke access when needed • Keep usage tied back to the workspace

@testingcatalog Cross-app automation gets magical right up until permission boundaries get fuzzy. The UX win is huge, but the teams that ship this well will be the ones that make scope, confirmation, and undo painfully clear.

GOOGLE 🔥: A new Android Intelligence has been introduced during Android Show 2026! - A whole new sleek design! - Automated multi-step tasks across Android apps - Gemini in Chrome gets Browser Use - Automated form filling - "Rambler" to turn voice notes into text - Custom Gen UI Widgets I need a Pixel now 👀

@OpenRouter @perceptroninc Nice addition. When teams test a new model through any gateway, the hidden questions are usually latency consistency, rate-limit behavior, and how fallback behaves when the model is hot. “Model listed” and “model production-ready” are cousins, not twins.

Perceptron Mk1 is live on OpenRouter, built by @perceptroninc. Frontier video and embodied reasoning in a vision-language model. Analyzes video at a dynamic frame rate (up to 2 FPS) across a 32k multimodal context, with hybrid reasoning and structured spatial primitives (points, boxes, polygons, clips) as first-class outputs.