@durov
 It happened. Our decentralized confidential compute network, Cocoon, is live. The first AI requests from users are now being processed by Cocoon with 100% confidentiality. GPU owners are already earning TON. https://cocoon.org is up.
 
 Centralized compute providers such as Amazon and Microsoft act as expensive intermediaries that drive up prices and reduce privacy. Cocoon solves both the economic and confidentiality issues associated with legacy AI compute providers.
 
 Now we scale. Over the next few weeks, we’ll be onboarding more GPU supply and bringing in more developer demand to Cocoon. Telegram users can expect new AI-related features built on 100% confidentiality. Cocoon will bring control and privacy back where they belong — with users.

@BetterCallMedhi
 
pour ceux qui ne comprennent pas forcément ce qui est en train de se passer :  
 
Durov vient de lancer Cocoon une infrastructure où des propriétaires de GPUs louent leur puissance de calcul pour faire tourner de l’IA de manière chiffrée et gagnent du TON (crypto)
 
mais ce qui se joue va bien au-delà de la décentralisation technique, je m’explique :
 
ce qui est fascinant c’est l’architecture de coordination systémique pudique Durov possède Telegram donc il contrôle un milliard d’utilisateurs qui génèrent une demande massive et instantanée pour du compute IA confidentiel
 
et via le protocole TON il orchestre l’offre en permettant à n’importe qui de monétiser ses GPUs idle avec des garanties cryptographiques Intel TDX qui assurent que même le propriétaire du GPU ne peut pas voir les données traitées
 
le résultat c’est une boucle de rétroaction autorenforçante où + il y a d’utilisateurs TG + la demande de compute augmente, ce qui attire + de fournisseurs de GPUs,réduit les coûts & rend le service + compétitif face aux hyperscalers
 
c’est exactement comme la Chine qui a construit son écosystème manufacturier en intégrant verticalement toute la chaîne depuis les matières premières jusqu’aux produits finis avec une coordination étatique sauf qu’ici la coordination se fait via des protocoles cryptographiques et des incitations économiques au lieu d’une planification centrale!!!  
 
la vraie disruption c’est que Durov peut exécuter cette vision précisément parce qu’il possède tout à 100% sans actionnaires qui demanderaient de monétiser les données utilisateurs ou de prioriser la croissance trimestrielle sur l’architecture long terme et ça C’ÇA CHANGE LITTÉRALEMENT TOUT
 
d’ailleurs, pendant que les startups européennes se diluent à 30% après leur Série B pour construire des features SaaS de merde, Durov construit une couche souveraine de compute qui pourrait redéfinir où et comment l’IA mondiale s’exécute
 
sachez qu’ici on parle pas juste de tech random et c’est ce qui est le + fascinant puisqu’on parle avant tout de géopolitique encodée dans des smart contracts et perso c’est ce genre d’alliance dont j’ai toujours rêvé entre écosystème crypto et IA

Prerequisites
[...]
NVIDIA GPU with CC support (H100+)
[...]

Made a CLI that lets Claude Code use Gemini 3 Pro as a "lead architect"
 
I've been using Claude Code (Opus 4.5) a lot lately and noticed it sometimes goes off in weird directions on complex tasks. It's great at writing code (especially Opus 4.5), but architecture decisions can be hit or miss. Gemini 3 Pro is INCREDIBLE at this.
 
So I built a CLI wrapper around Gemini that integrates with Claude Code. The idea is Claude handles the implementation while Gemini provides strategic oversight.
 
Since Claude Code auto-compacts it can run for very long. The /fullauto command takes full use of this.
You can send a prompt, go to sleep, and it will be either done or still working when you come back. So only Claude subscription / Gemini API key rate-limiting will stop it.
 
The Oracle maintains a 5-exchange conversation history per project directory by default so Gemini has enough context to make useful suggestions without blowing up the context window. Claude can also edit this context window directly, or not use it (oracle quick).
 
It will auto install a slash command `/fullauto` mode. You give Claude a task and it autonomously consults Gemini at key decision points. Basically pair programming where both programmers are AIs. Example:
 
/fullauto Complete the remaining steps in plan.md
For /fullauto mode, Claude writes to FULLAUTO_CONTEXT.md in your project root. This works as persistent memory that survives conversation compactions.
 
/fullauto also instructs Claude on how to auto-adjust if the Oracle's guidance is misaligned.
 
It can also use the new Gemini 3 image recognition and Nano Banana Pro for generating logos, diagrams, etc.
 
When Claude runs oracle imagine it will use nano-banana-pro image generation, and if it's region blocked the CLI automatically spins up a cheap US server on Vast.ai, generates the image there, downloads it to your machine, and destroys the server (you need vast.ai API key for this).
 
Example uses Claude Code can do:
 
# Ask for strategic advice
oracle ask "Should I use Redis or Memcached for session caching?"
 
# Get code reviewed
oracle ask --files src/auth.py "Any security issues here?"
 
# Review specific lines
oracle ask --files "src/db.py:50-120" "Is this query efficient?"
 
# Analyze a screenshot or diagram
oracle ask --image error.png "What's causing this?"
 
# Generate images (auto-provisions US server if you're geo-restricted)
oracle imagine "architecture diagram for microservices"
 
# Quick one-off questions
oracle quick "regex for email validation"
 
# Conversation history (5 exchanges per project)
oracle history
oracle history --clear
I used this tool to create the repo itself. `/fullauto` orchestrated the whole thing.
 
Repo: https://github.com/n1ira/claude-oracle

When you select Auto or Pro, Gemini CLI will attempt to use gemini-3-pro-preview first, before falling back to gemini-2.5-pro

Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents!
 
 DeepSeek-V3.2: Official successor to V3.2-Exp. Now live on App, Web & API.
 DeepSeek-V3.2-Speciale: Pushing the boundaries of reasoning capabilities. API-only for now.

@steph_palazzolo
Sam Altman told staff today that he was declaring a “code red” as ChatGPT faces growing threats from Google and other AI makers.  
 
He wrote that he’s marshaling more resources to improve model behavior and other features in the chatbot.
 
w/@erinkwoo

ChatGPT Android app 1.2025.329 beta includes new references to an "ads feature" with "bazaar content", "search ad" and "search ads carousel"

The reason why Deepseek V3.2 is so cheap
 
TLDR: It's a near linear model with almost O(kL) attention complexity.
 
Paper link: https://github.com/deepseek-ai/Deep [...] k_V3_2.pdf
 
According to their paper, the Deepseek Sparse Attention computes attention for only k selected previous tokens, meaning it's a linear attention model with decoding complexity O(kL). What's different from previous linear models is it has a O(L^2) index selector to select the tokens to compute attention for. Even though the index selector has square complexity but it's fast enough to be neglected.