Just shipped HermesOffice — an open-source office suite (Word, Sheets, Slides, PDF) where the AI is a real agent running 100% on your machine.
No cloud account. No third-party proxy. The assistant that edits your documents is the same agent that knows your files, your memory, your tools.
Built on the excellent GenOffice engine (Apache-2.0), forked and rebranded. → https://t.co/eZsenv9dwZ
@Teknium@NousResearch
This is F*cking gold
andrej karpathy joined anthropic to lead a team using claude to accelerate its own pretraining research, teaching claude to help build the next version of itself.
that's real, confirmed directly. the "two anthropic seniors made his loop 1000x better with graph engineering" claim is not something i can verify anywhere, and i'm not going to invent a personal testimonial to make it sound more credible than it is.
what is genuinely worth reading: anthropic's own published claude cookbook on knowledge graph construction, extract, resolve, assemble, query, real and public, no leaked breakthrough required.
read that, then check the article below.
MCP protokolü bugün en büyük güncellemesini aldı. Kimse ne olduğunu tam anlamadı ama bu tarih değiştiren tarafta.
Eski MCP: her session'ın state'ini serverda tutmak lazımdı. Yani serverless çalışmıyor, edge'e atamıyor, load balancer arkasına sticky session'la bağlıyordun. Enterprise için can sıkıcı, indie geliştirici için pahalı.
Yeni MCP: stateless. Artık server'ı Cloudflare Workers'a, Vercel Edge'e, tek satır YAML ile Kubernetes cluster'a atıyorsun. Load balancer round-robin, sonsuz yatay scale.
Üstüne üç yeni katman geldi:
> MCP Apps: agent'ın kendi UI'ını sandbox iframe içinde render ediyor
> Tasks: uzun süren async operasyonlar
> Enterprise Managed Auth: identity provider'dan merkezi erişim kontrolü
MCP artık sadece "AI'ın USB-C'si" değil. Server'ı stateless, UI'ı render eder, uzun task'ları arka planda çalıştırır, SSO ile enterprise kilidini açar.
Yani her AI agent artık aslında bir mini SaaS platformu.
Vibe coder'ın hayali gerçek oldu: bir prompt yaz, agent hem hesabı hem UI'ı hem auth'u tek başına kurar.
Kaydet, 2026'nın en önemli teknik hamlelerinden biriydi.
quick update on how this is going: they have gone back to linear because maintaining their internal tool that they vibecoded was taking away from their actual work’s bandwidth.
When processing text for large language models, you must first tokenize the text before mapping it to vectors.
Marcel got in touch with me months ago. He wanted to make it faster. I told him I had tokenizers on my todo but had not done anything with it. He proceeded to seemingly crack the problem.
Happy to see he got good results.
You’d think I could just get AI agents to work on all my clever ideas… but no. Human understanding is still a key bottleneck.
So let us build collectively. There is so much to do.