Pour tous ceux qui vibecodent des apps et des SaaS allez suivre ce mec.
Il vous partage des conseils de fou pour que votre SaaS ressemble à une licorne au lieu de ressembler à tous vos concurrents
Bu fiyatlandırmayı kimse konuşmuyor ama asıl bomba orada.
Model daha güçlü olmuş, tamam. Ama önbellek okumaları %75 daha ucuz. Yani Claude Code ile gün boyu ajan çalıştıran adamın faturası %25, ağır iş yapanın %45 düşüyor.
Eskiden "Ai pahalı, ancak büyük şirketler kaldırır" derdik. Şimdi tam tersi, ne kadar çok kullanırsan o kadar ucuza geliyor.
Bu işi ciddi yapan için maliyet duvarı yıkıldı. Üretenin önü açık.
Hayırlı olsun fable 5.1 ve mythos 5.1 yarın testleri yapıp paylaşacağım.
很多人用 Pi,但未必知道背后这个人。
Pi Agent 的作者叫 Mario Zechner,X 上是 @badlogicgames。奥地利人,以前不是做 AI 的,是做游戏框架的。libGDX 就是他搞出来的,Ingress、杀戮尖塔这些都用过。
后来还做过 RoboVM,公司卖了又被微软关掉,社区反噬那一套他经历过,所以现在特别不愿意自己去融一堆钱当 CEO。
2025 年底他自己折腾了一个极简 coding agent,两晚写出来,本来只给自己用。默认就四个工具:read、write、edit、bash,系统提示词压得很短,其他全靠你自己加。口号也写得很直:There are many agent harnesses, but this one is yours。
后来 Flask 作者 Armin Ronacher 的公司 Earendil 把 Pi 收了,Mario 入股并加入团队。他自己发文标题就叫 I've sold out,写得很坦率:不想再走一遍创业那套高压,家里有小孩,但又希望 Pi 能有人养着别停更。技术方向还是他拍板,核心继续开源。
这人说话冲,bio 写的是 Old man yelling at Claudes。讨厌 Agent 越做越重、提示词越写越长。他自己还专门测过 MCP 和 CLI,结论也和很多人实际感受一样:CLI 往往更省、更稳。
所以 Pi 长成现在这样,不是产品经理堆出来的功能清单,是一个老开源作者按自己脾气做的壳。你喜欢它干净、可改、不绑死模型,基本都能从他这个人身上对上��。
感兴趣的话可以去看他博客 https://t.co/DEJ5r0GRU6,或者直接关注 @badlogicgames。产品号是 @pidotdev。
الصين تُطلق أول موظف ذكاء اصطناعي مكتمل يعمل لحسابك 24 ساعة يومياً وبشكل مجاني 100%! 🚨🇨🇳
أعلنت شركة ByteDance رسمياً عن طرح مشروع DeerFlow المصنّف كأقوى نظام وكلاء مفتوح المصدر يعمل محلياً على جهازك الشخصي دون الحاجة لسيرفرات خارجية!
ليس مجرد شات بوت يجيب على الأسئلة، بل نظام يتكفل بإنهاء الأعمال الصعبة بالكامل! 🧵👇
Así se ve un sueldo de 750.000 dólares al año: un tipo en camiseta blanca, un pizarrón y 2 horas y media.
Stanford, CS336. Percy Liang construye un LLM desde cero. Lo que hay debajo de Claude y ChatGPT, y arranca por la parte que todos saltan: el modelo no lee tu texto, lee números.
Anthropic paga ese sueldo a los ingenieros que entienden esa capa.
Lo único que cobra Stanford son 2 horas y media de tu atención.
QWEN 3.6-27B JUST EMBARRASSED A MODEL 14X ITS SIZE.
Alibaba built a 27B parameter model that beats its own 397B flagship on coding.
But the size difference isn't even the most useful part.
Why Qwen 3.6-27B matters:
→ 27B dense model — all parameters activate on every token
→ 77.2% on SWE-bench Verified for real-world GitHub bug fixing
→ 94.1% on 2026 math reasoning
→ 82.9% on multimodal academic reasoning
What you actually get:
✓ 262K token context window
✓ Text + image + video inputs
✓ Apache 2.0 license
✓ Can run locally instead of sending everything to third-party APIs
And the workflows are practical:
✔ Analyze landing pages from screenshots
✔ Build 7-day email sequences
✔ Feed it dozens of community posts and extract recurring pain points
✔ Turn those insights into content scripts
✔ Use thinking mode for deeper strategy and planning
The lesson:
Stop judging local AI by parameter count.
A focused 27B model can now outperform a 397B model on coding while being realistic to run yourself.
And Qwen 3.8-27B is already the next model to watch.
Your AI agent forgets everything the second it finishes a task.
Every fix. Every answer. Gone. Like it never happened.
That just changed.
Claude managed agents can now carry memory across runs — even in self-hosted sandboxes.
→ It writes down what it learns
→ Picks up right where it left off next run
→ You can view, export, or roll back the memory anytime
Before: agent = new hire every single day, same mistakes on repeat.
Now: agent = teammate that actually improves.
Save this video, you'll stop rebuilding the same agent from zero every time. ✔️
Want the SOP? DM me.
This blog post was written for those who may be curious to know what an agent harness is, but don’t, and might not know where to start. Read the full post here
https://t.co/pUnUnQvEHJ
A harness turns a model into an agent. At it’s core it provides 4 things:
- a system prompt
- tools
- an agentic loop
- a translation layer across models
New blog post from Earendil co-founder @colindaymond on what a harness is, and how you can own yours. Full post below
نموذج مجاني يعمل على جهازك المحالي يُسقط أقوى نماذج الذكاء الاصطناعي المدفوعة في العالم!" 🧵🚨
صرح Alex Finn بتجربة مرعبة: نموذج Qwen 3.8 27B يعمل محلياً على كارت RTX 5090 ويتغلب على نموذج Opus 4.8 في الاختبارات!
نحن أمام عصر الذكاء الخارق المجاني وغير المحدود على مكتبك 🧵👇