kotevcode

Spent the last few months building telebiz.io It's your entire business stack (CRM, AI agent, automations) - but inside Telegram where you already work.

In a world where most business tools monetize your data in some form, Telebiz was built on a different principle. Here's exactly what happens to your data: → Your Telegram messages go directly from your browser to Telegram's servers. → Your AI requests go directly from your browser to your LLM provider. → Your CRM data syncs directly between Telebiz and your CRM. The entire codebase is open source and publicly auditable on GitHub. Check it out yourself: github.com/kotevcode/tele…

Proof Of Prior Delivery. Soon

It’s a massive week for AntSeed! 🐜✨ The v0.1.15 desktop release is officially live with some heavy-hitting updates: * OpenAI Responses API Support: Now serving Codex, ChatGPT agents, and more via /v1/responses. * Code Cleanup: Deleted ~1,700 lines of custom SSE parsers—native streaming via pi is here. 🧹 * Windows is Live: Full support for the desktop app on Windows! 🪟 * UX Polish: Refined Chat/Discover UI + smarter, "sticky" peer pinning. Building the p2p AI network, one commit at a time. 🚀

@AvinoamOltchik 🫡

I made a trivia app that teaches you how to become a Claude Code Architect. 351 questions based on Anthropic's (rumored) certification. Every answer teaches you something — right or wrong. Public leaderboard. Free. claude-architect-trivia.onrender.com #ClaudeCode #Anthropic #AI #Claude

Most AI tools live in a separate tab you have to remember to open. The Telebiz AI agent lives inside your Telegram conversation- where the actual work happens. Here's what it does in practice: - Reads your full conversation history for context - Drafts a reply that sounds like you, not a robot - Logs the interaction directly into your CRM - Sets a follow-up reminder based on what was said The best BD tool is one that works where you already work. Read more here: telebiz.io/docs/ai-agent/

Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers. 🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. 🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale. 🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead. 🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains. 🔗Full report: github.com/MoonshotAI/Att…

Here's how @telebizio dataflow works: → Messages go directly to Telegram. → Our database stores only encrypted Chat↔CRM ID mappings. No data access. → AI requests go from your browser to the selected provider. Your key, your billing. → CRM OAuth tokens are encrypted at rest with AES-256-GCM. We built a business layer on top of Telegram without ever touching what's inside it.

🐜 cooking something big.

POV: you're running BD on Telegram. Your pipeline is split across 40 DMs and 3 group chats. Your CRM has been "almost up to date" for two weeks. Your follow-up reminder system is a folder with chats you keep ignoring. And somehow you're still closing deals. Imagine what you'd close with the right tooling behind you. TeleBiz brings CRM, AI, automations, and reminders directly into your Telegram - where your deals already live. Start today at- Web.Telebiz.io 🫡

the goal is simple: any AI tool, any provider, no lock-in the P2P routing proxy is that invisible layer, it just works, regardless of which peer handles your request infrastructure should disappear. we're getting there. antseed.com/?v=16 github.com/AntSeed/antseed

fact

@bcherny Finally 😅

🎁 Happy Friday - Opus 4.6 1M is now the default Opus model for Claude Code users on Max, Team, and Enterprise plans. Pro and Sonnet users can opt in with /extra-usage.

gm. did you update the CRM? no more tab-switching. no more forgetting. everything is aligned and visible across your team. with @telebizio your CRM lives inside Telegram, where your conversations happen

Imagine having complete confidence that you're using the top-tier AI service for a specific task, like building a website, without having to fuss over finding the exact 'best model' yourself with new models dropping everyday. leave it for the provider. Choose based on reputation.

starting with @telebizio the first step is to set up your organization. by creating your organization, you can also add your team members and stay on top of their actions and synchronize with the notes they are taking. no need anymore to go to your crm to check the notes. Everything is happening inside your favorite app.

New to Telebiz? Set up your Organization first. 🏢 It's your team's central hub for: - Shared CRM & productivity integrations - Role-based access (Owner, Admin, Member) - Centralized settings & preferences How to get started: Telebiz Settings → Organizations → "Create Organization" → Done. Add your team, connect your tools, collaborate, and stay on top of your team actions — all in one place. Learn more here: telebiz.io/docs/organizat…

how ai agents actually work. think of a conversation as an array of objects: system → user → assistant → user → assistant… the system prompt defines the role. example: “you are a coding agent that solves programming problems.” then the user asks something. the llm receives the entire conversation as context and starts generating the next tokens. (what are tokens? small pieces of text. a token can be a whole word, part of it, whitespace, etc) but there’s another piece: tools. tools are just functions the llm is allowed to call. for example: • read_file(path) • edit_file(path, content) • search_web(query) • bash(command) these are passed to the model as part of the context. so instead of only returning text, the llm can return tool calls. example: “call read_file with path /home/user/project” the agent receives that request, runs the function, and sends the output back to the llm as more context. then the llm continues thinking… so the loop looks like this: llm → tool call → tool output → llm → tool call → … the llm is the brain. the agent is the hands and legs. when cc or codex writes code for you, that’s exactly what’s happening. the model asks the agent to read files, write files, run commands, see results. back and forth in a loop. now there’s another layer called skills. skills are basically a special tool call. the agent injects summaries of available skills into the prompt and tells the llm: “you have skills. to use them, call the skills tool with the skill name.” when the llm decides a skill is relevant (for example writing python), it calls the skill tool. the agent then loads the full instructions for that skill and injects them into the context. so skills are really just prompt bundles loaded through a tool call. a minimal agent implementing this loop can be written in ~20 lines of code. most real implementations are larger because they handle edge cases: timeouts, retries, nudging the model when it gets stuck, adding context, preventing loops. but the core idea is still just this: llm thinks → calls tools → sees results → thinks again.

Every deal starts with a conversation. With Telebiz, that's also where it gets managed. Connect your CRM directly to Telegram and update deals, create new opportunities, log notes, and set follow-ups, all from within the chat. No switching tabs. No forgetting to update the pipeline. No information is lost between conversations. Your CRM stays current because you never have to leave the place where business actually happens. Supporting @NotionHQ, @HubSpot and @pipedrive Start today at: Web.Telebiz.io

@karpathy we are working on a p2p inference network. @AntSeedAI , will always find you a replacement

My autoresearch labs got wiped out in the oauth outage. Have to think through failovers. Intelligence brownouts will be interesting - the planet losing IQ points when frontier AI stutters.