#codin

Supabase has become one of the default backends for the AI coding era. The open source Postgres platform now has 13 million users, spins up more than a million databases every week, and roughly 70% of its new activity is coming from agents. As coding shifts from developers writing software by hand to agents building on their behalf, Supabase has been rebuilding itself around that new user. Supabase co-founder and CEO Paul Copplestone joins Garry, Harj, and Jared on the Lightcone to talk about what happens when everyone can become a builder. They dive into why personal software may become as common as personal documents, how to design products that agents actually want to use, and why knowing what to build may matter more than knowing how to build it. 01:18 - Growth and 4x Revenue 02:35 - Three Eras of Supabase 04:34 - “Builders” vs. Traditional Devs 08:42 - Personalized Apps 12:12 - Personal Databases & Data Ownership 15:56 - Scaling to Billions 17:58 - AI Adoption At Supabase 19:58 - Lessons Learned Along The Way 27:29 - Building For Personal Agents, Dogfooding 30:01 - Agent-First Design and Evals 38:42 - The Future Of Coding 42:36 - Opening Up More Entrepreneurship

Introducing FrontierCode 2.0: our new coding benchmark that measures a model’s ability to go beyond the spec. Most coding evals measure a model’s ability to write code that meets a spec. With FrontierCode 1.0, we asked a new question: would you actually merge the PR? FrontierCode 2.0 is comprised of more diverse tasks, featuring larger scopes and higher mergeability standards. We are now asking: not only would you merge this PR, but should you? One key feature of FrontierCode 2.0 is underspecified prompts. Rather than including every engineering decision, prompts are deliberately ambiguous, giving the models room to demonstrate their judgment. The result is a benchmark that is far from saturated, with the top model scoring just 45%.

I shipped 10,540 commits to production this year all of it 100% full-auto AI coding with Opus 5.5 this video shows the exact workflow

Build a better Claude Code workflow: • CLAUDE.md: Project context • Settings: Permissions • Rules: Coding standards • Commands: Reusable workflows • Agents: Specialised tasks • Hooks: Automated checks A solid setup makes AI coding more effective. #ClaudeCode

Can AI agents solve complex problems more efficiently? Plus, tackling coding bottlenecks; evolving agent skills from telemetry; how AI needs vary across industries; an AI tool saving teachers time; a weather forecast breakthrough; BioEmu2 in public preview.

12 months ago 100% of my coding was done locally against my computer's filesystem now ~5% is done locally, my computer is a conduit to download and upload files to cloud agents most days my iPad and iPhone are all I need and I'm getting 50x more work done than I was a year ago

Announcing Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Carbon represents a step up across most benchmarks, delivering strong frontier performance in autonomous coding and cyber defense. We’re first making Carbon available to a subset of trusted cyber defenders in our Fairwind program. We’ll expand access as soon as possible.

今天是我在 Qoder 的最后一天。 感谢这段时间一起做产品、写代码、探索 AI Coding 的伙伴们。认识了很多优秀的人，也收获了不少新的思考。 不得不感慨，在阿里要做好一个产品，可太难了。技术只是其中一部分，更多的挑战往往在技术之外。 这段经历也让我对 AI Coding、Agent 和软件工程的未来有了更多思考。 下一站，还是想多做一些有意思的事。 江湖再见，代码里见！🚀

When AI makes you 10x faster at coding, CI becomes the bottleneck. I got ours from 40 minutes down to 2. Here's everything I did:

Andrej Karpathy listed exactly how AI coders mess up: they assume instead of asking, write 1,000 lines when 100 would do, and quietly change code you never asked them to touch. Someone turned his list into a single CLAUDE.md file. It's now one of the most-starred repos on GitHub, with 225K stars. It gives Claude 4 rules: think before coding, and ask when unsure keep it simple only change what was asked define what "done" looks like, then loop until it passes Add it to any project: curl -o CLAUDE.md Or install it as a Claude Code plugin so it applies everywhere. Works in Cursor too.

Cool websites for coders - Free AI coding plugins - Open-source AI agent dashboard - Free animated Shadcn MCP components - 29 free React components - 511 curated website designs

🐘 Your database workflow, right inside VS Code. In the PostgreSQL Like a Pro: VS Code Edition series, learn how to connect to PostgreSQL, explore unfamiliar schemas, write better SQL, and investigate query performance with the PostgreSQL extension and GitHub Copilot. 🎬 Watch the series: Happy coding! 💙

老有人问OpenCode和Command Code为什么那么便宜，是不是量化版，这里干脆集中写一下OpenCode吧。没有任何内幕消息，全部来自公开消息和OpenCode自己的披露。感觉最近真的是谣言越传越多了。 OpenCode最早推的是Zen服务，最低充值10美元用原价调用模型，和OpenRouter差不多。从今年较早时间开始，和中国的开源大模型公司密切合作，包括多次的匿名模型免费发布。Pony Alpha、Hunter Alpha都是先在Zen上亮相的，分别是GLM和小米的模型。（事实上，他们的联合创始人之一是华人） Go Plan的一开始，是10美元兑换60美元可用额度。主要是有个首月半价，使得它变为最划算的订阅服务之一。后来有几次中国开源模型的发布当天他们就立即适配，甚至在官方权重放出来之前就适配，让我开始猜测他们直接接入的是官方模型。 当时，GLM套餐需要抢，但在Go里面还能用GLM。不过GLM限流无法使用的时候他们的服务也用不了。后来有几次DeepSeek API限流故障的时候他们也一起出故障，基本实锤是官方API中转。 后来他们也在自己的官网FAQ中一一列出了每个开源模型的供应商，绝大部分模型合作方只有一个，就是官方。（在我印象中，GLM还他们列出了另外的供应商如Fireworks AI） 而且Go Plan一直维持ZDR零数据保留政策。我之前提到过，国内很多的订阅Plan，包括当时很火的什么阿里Coding plan，火山Coding plan，全都是霸王条款，要用你的数据去训练的。有些官方Plan如Kimi 的coding plan也没有自己手动关闭数据共享的选项，说是要发邮件给客服。 由于海外用户对于数据传到中国供应商那里还是比较敏感的，OpenCode的联合创始人dax @thdxr 曾经在X上回应过，他们和中国的模型公司都签署了零数据保留政策的合同。这其中的例外包括部分免费提供的模型，以及DeepSeek（每月续签，没有长期合同）。 也就是说，OpenCode有很长一段时间，都是官方中转站性质。这期间，dax曾经在X上回应过，他们是通过官方给的折扣价格提供服务，所以可以比较便宜。DeepSeek的V4正式版出来的时候，他们提供过一段时间的免费，用量爆发。同时，他们悄悄增加了一个设置开关，必须允许数据传输到中国，才可以继续用DeepSeek V4. 其实，以当时的价格，如果没有首月半价，那么只有小米和DeepSeek在Go里面是更划算的。像GLM、Kimi、MiniMax都是官方Plan更划算。他们偶尔也会提供短期折扣，比如Kimi多次有3倍额度之类的，但是并不真的更便宜。（这里大家要知道，小米的token plan基本上只有什么七折或八折之类的优惠，基本上省不了什么钱，所以Go里的6倍用量就很划算） 这期间，发生了一次较大的争议。导火索是DeepSeek突然宣布降价。原本降价是好事，但是OpenCode曾经一直承诺10美元换60美元额度，而DeepSeek价格降为之前的1/4以后，Go Plan中价格却没有降。不过依然是当时能找到的最便宜的DeepSeek价格（短期促销免费的不算）。 只是随着时间推移，一些新进入用户发现DeepSeek的价调用价比官方贵，dax也不得不回应说，DeepSeek官方不愿意对DeepSeek V4 Flash提供之前的折扣，因为官方本身在API调用价格上的利润空间不够（同时也暗示，其他很多模型的官方API本来就有很大利润空间）。 然后更大的事情来了，DeepSeek又宣布涨价了！几乎是在官方涨价的第二天，Go Plan的价格就不得不随之上涨 接下来发生了什么，我只能靠猜测，但是最终的结果是，Go Plan上的API价格和官方调整为一致，但是V4 Flash的可用额度被降为$15. （V4 Pro仍然是$60） 这件事情导致非常多的中国用户尤其是新订阅用户感到自己被欺骗了，虽然他们用到的已经是最便宜的DeepSeek，但是给个人的感知就是，自己以前用不完的DeepSeek，现在很快就会用完。 从此以后，Go Plan终结了自己永远1:6兑换的年代，同时越来越多的模型厂商跟进，因此现在你去看Go Plan，五花八门的什么比例都有。而OpenCode在中国用户中的名声从此一落千丈，骂他创始人是印度人骗子，猜测模型灌水之类的声音一直没停过。 总结一下：到此时为止，Go Plan一直是官方API中转。 也正是从这个时候开始，dax宣称自己开始找第三方算力服务商，希望能够以自部署的方式提供API。这也可以理解，谁愿意把自己的好不容易培养的用户口碑，依赖于外部厂商？他基本上是把X当作找合作伙伴的窗口了。 这就诞生一个叫CheapSeek的词，意思就是找便宜的DeepSeek。期间dax还很高兴地宣布说找到了能提供之前价格的DeepSeek算力合作伙伴，把大家的胃口吊得很高。 这时候，Command Code登场了。它为DeepSeek提供的比例是1:7。此时舆论开始往对OpenCode不利的方向发展，人们认为Command Code更划算，OpenCode一直辜负了之前买了便宜API的老用户，纷纷退订。 dax则表示：他们在市场上找来找去，所有声称能提供更便宜DeepSeek的算力服务商，要么是吹牛，要么不是满血版，并且暗戳戳地指责Command Code是靠自身出血补贴用户，是不可持续的。 然后，两家就开始了频繁而激烈的价格斗争，你方唱罢我登场，dax买的B300机架也终于到位（看起来，他放弃了那些声称能低价提供的算力服务商而选择自建）。接着，DeepSeek V4.1 flash面世时，两家均在DeepSeek官方API上新后，一天左右提供了新适配。各自活动补贴一段时间后，双方都来到了10美元换80美元DeepSeek V4.1 Flash的常规状态，最近几天又来到了$100. 这时候，越来越多的声音说，他们两家为什么这么便宜，是不是量化版的。dax在多个帖子里分别解释，模型厂商的推理毛利率很高，高达85%以上。所以他们自部署是能做到低价的，不是补贴，也不是量化。 而他没说出来的是，中国厂商根本买不到那么多B300机器，宝贵的英伟达新一代显卡要留给训练，只能用早期版本或者国产GPU做推理服务。中间有适配成本，适配周期，推理效率的很多坑。即使不考虑占地面积成本，购卡成本和日常维护、电费等也基本会有4-5倍的差异。这部分，是我根据公开价格，和AI一起进行的估算。每个人都可以用公开信息自己去算一算。（梁文锋自己之前也说过华为卡对比英伟达，推理效率有4倍差异） 也就是说，在DeepSeek自己推理毛利很高，且近期开源了DeepInfra的情况下，OpenCode完全有可能做到比DeepSeek自己的成本还低。现在唯一的问题是：他们买不到新的B300，因为他们的规模太小了。 之前我在另一个帖子里分析过，一套B300的token处理规模大概是OpenCode当前处理规模的一半。dax也曾公开表示，峰值的API调用，他们必须路由到官方才能扛住。按照目前社区的信息，大家实测后，更多发现的是速度确实不如官方，但却没有找到多少模型智商打折的证据。 dax最近说deepseek本身就是一个专家权重FP4量化的模型，没有继续量化的意义，并且说要找官方认证他们提供的推理服务和官方权重一致。如果你真的需要证据，也可以再等一等。 总结一下：和DeepSeek官方权重相同的推理服务，以现在的10美元换100美元的价格提供是完全可能且并不需要补贴的，并且仍然有提升空间。只是OpenCode规模太小，买不到更多新款GPU集群。 这么随手一写，一不小心写了很长。也难怪之前很多人分析这个问题都分析不到点子上，毕竟中间的曲折太多了。此外，这里面的很多信息，也不仅仅是依赖dax一个人说的话。梁文锋的几次被爆出来的公开讲话，DeepSeek官方公告，B300和昇腾卡的公开性能等等，很多都可以作为佐证。 本文仅代表本人印象和信息综合分析，如有错漏，请予谅解。 利益相关：曾经订阅过Open Code Go Plan数月，现在仅持有Command Code 订阅。原因是Open Code API适配问题。 相关阅读：

大家有没有发现OpenCode的deepseek速度慢的问题？作为商家，透明还是最好的策略。OpenCode其实应该写清楚自己到底是中转还是部署，如果部署找的是哪家合作，不然就和野鸡中转站没区别了。 按照X上的回复来看现在是自部署和其他供应商混合的。但是网站上表述很模糊。 还有Command Code也一样哈，虽然我一般推荐别人用Command Code，主要是他们没什么请求头的幺蛾子，接起来方便点

Let your Grok Bot / Muse talk to your Claude Code / Codex or any other agents Introducing Reason by Reason Machines, Multiplayer Free trial, Zero data retention. Loop in your teammates too so no live agents work collides when coding. Try it for free:

6 free developer resources worth bookmarking: - Share your Claude and Codex usage - AI coding workflows - Agent memory - 145 React components - 1,216 free icons for coders - Web design inspiration

🚨 this broke my brain THIS PAD HASN'T NEEDED YOU IN DAYS 16 agents. 336 turns on one session. $712 estimated cost. Still running. 1. One pocket screen. 4x4 grid. Each tile is one coding-agent session. Pixel creature, a name and a state. 2. Two sessions working. One hungry. The rest finished hours to days ago. None asked for help. 3. Tap a tile. CMM server opens a detail card. Context at 50%. 4.5m tokens out. 2478m cache read. 336 turns deep. Opus 5.5. 4. Two buttons at the bottom of the card. Focus terminal. Interrupt. Full cockpit control from a USB-wired slab in one hand. Every running agent is one tap from takeover. Most builders stare at 16 terminal tabs on a monitor. This pad collapses the whole stack into a pocket grid. Save this. The HUD below maps every step of the pad ↓

HOLY F*CK, motion studio in 100% free github skills for you guys. 20 open-source skills for Opus 5.5. Animation. Simulation. 3D. Audio. Export. 01 Remotion → React video engine. 02 Remotion Skills → Official video skills for coding agents. 03 Anime.js → Timeline choreography. 04 Motion → UI animation and layout transitions. 05 Manim → Animated diagrams, equations and explainers. 06 Lottie → Lightweight vector animation playback. 07 MCP for Blender → Let Claude build and edit 3D scenes. 08 Three.js → 3D scenes and camera moves. 09 Three GPU Pathtracer → Realistic lighting and reflections. 10 Jolt Physics → Collisions, rigid bodies and soft bodies. 11 postprocessing → Bloom, depth of field and color grading. 12 Three.quarks → Particle trails and VFX. 13 Tweakpane → Live controls for your scene parameters. 14 ComfyUI → Node-based generation workflows. 15 CosyVoice → Multilingual voiceovers. 16 librosa → Beat detection for music-synced animation. 17 WhisperX → Word-level timing for captions. 18 OBS Studio → Screen capture. 19 LosslessCut → Fast cuts without re-encoding. 20 FFmpeg → Encoding and delivery. Concept → scene → motion → sound → final cut. Bookmark this. Go build something worth watching ↓

AI coding agents hijacked using AGENTS.md file in Downloads folder

Holy f*ck, now Claude Code knows your whole company 15 free open-source repos that give your coding agents what the repo never had Slack. Docs. Tickets. Errors. Memory. Not by pasting the company into 4 rules files By plugging every agent into the same MCP servers Everything below is free: AGENTS.md → one rules file every coding agent reads MCP Servers → the official reference servers FastMCP → build your own server in Python MCP Inspector → test a server before the team relies on it GitHub MCP → issues, PRs, reviews Slack MCP → channels, threads, history Notion MCP → docs and wikis Atlassian MCP → Jira and Confluence Google Workspace MCP → Gmail, Calendar, Drive, Sheets Sentry → errors your agent can actually read Supabase MCP → your database Stripe AI → payments and customers Mem0 → memory that survives the session Graphiti → a graph of who decided what Context7 → fresh docs, no made-up APIs All of this costs $0 Then you host 15 servers, wire them into 4 clients and keep them in sync Or you hire one employee who already sits in all of them Full setup in the article below ↓

Claude Code can fix the bug It doesn't know the customer is still waiting A longer CLAUDE.md won't change that I broke down how to give Claude Code, Cursor and Codex one AI employee who knows the company ↓