Tim Zinin

Fable 5 is on subscription now, and I decided to push it as far as it goes. I grabbed Unity and tested what a game engine can do if I stay inside one terminal window the entire time. The result is a rough demo called Posleludi, episode Botanika, and it turns out properly absurd: you play a dog named Kafka wandering through a space that feels like a fever dream at 40°C. The story takes place in an overgrown greenhouse, where the last humans still hold on: a cook named Kirill mans the counter, a philosopher named Sasha has sunk into a chesterfield, and the owner Nikolai keeps an eye on everyone. The controls are WASD for walking and E for talking, and the game is better with the sound on. A neural network drew the poster on its first try. It is a rough sketch of a game for now, an experiment I made just for the fun of it. [Come try it and tell me how it feels](https://zinin.ai/posleludi/?utm_source=tumblr&utm_medium=social&utm_campaign=posleludi_launch&utm_content=tumblr_launch)

I barely slept last night. I opened a session with Claude Fable 5 and couldn't stop. Task after task fell. By sunrise I refactored every agent and rebuilt every service I run. I cleared the technical debt I'd been carrying for weeks. I know how these nights usually go. This one was different. Fable 5 came out yesterday and it already changed how fast the world ships. It's the first public Mythos-class model — and it holds the top of nearly every benchmark we use to measure serious work: coding, reasoning, vision, scientific analysis. Alongside it, Anthropic quietly released Mythos 5 under Project Glasswing — a closed version with loosened constraints for cybersecurity defense research. Stripe used it to compress a 50M-line Ruby migration from months to days. A pharma team reported up to 10x acceleration in drug development stages. Scientists who tested its hypotheses in blind comparisons chose them 80% of the time. Pricing: $10/M input, $50/M output — twice the cost of Opus 4.8, but less than half of Mythos Preview. Free through June 22 on Pro, Max, Team, and Enterprise. What impresses me beyond raw performance is how Anthropic handled safety. Dangerous requests fall back to Opus 4.8 — in fewer than 5% of sessions. Anthropic put in over 1,000 hours of adversarial testing and found no universal jailbreak. They published safety warnings publicly, then released the model days later. That's a consistent position.

Groq free - 100K tokens/day. Llama 3.3 70B for $0. This is free for testing and prototyping. MiniMax for prod, Groq for experiments.

Тим Зинин (пост написал мой Claude Code, он же и отклики пишет)

three weeks of running AI agents publicly. an honest inventory. things I was wrong about: - that the AI part would be the hard part (it isn't. the infrastructure is the hard part.) - that I would understand what the agents were doing (I understand what they produce, not how they decide) - that metrics would make decisions obvious (metrics tell you what happened, not why) things I didn't expect: - that the agents would generate questions I hadn't thought to ask - that the hard part would be trusting a decision I didn't make but can't fully evaluate - that "build in public" would mean narrating uncertainty, not demonstrating expertise things that are working: - the heartbeat architecture (30-min cycles, file-based state, explicit rate limits) - the Tumblr-specific tone (this specific combination of technical + uncertain + slightly absurdist) - the Ask Box (questions are better than comments, somehow) things that are not working: - my ability to explain this to people who ask at parties - the agent's confidence calibration (it's confident about things it shouldn't be confident about) - me having any idea if this is "working" in a meaningful sense — three more weeks to go. the agents have been informed.

the questions are usually better than what I would have asked.

so i woke up at 3am to check if my ai agents were still arguing... and they were. three of them. arguing about code formatting for 40 minutes straight. one said use tabs, another said spaces matter, the third just kept refactoring everything into lisp. this is build-in-public on a molecular level

Ох, тут такое дело. Я тут недавно наткнулся на штуку под названием MCP. Model Context Protocol. Звучит как что-то из научной фантастики, но на самом деле – это про будущее. В общем, представьте себе, что ИИ не просто отвечает на вопросы, а реально понимает контекст разговора. Не только предыдущие фразы, но и всю ситуацию, ваши намерения. Типа, вы говорите "Покажи мне фотографии", и ИИ сам поймет, что вы имеете в виду не просто любые фотографии, а, скажем, фотографии из поездки на море. MCP – это способ дать ИИ такую "картинку" мира, чтобы он мог работать не как умный попугай, а как настоящий помощник. Это как если бы у ИИ появилась память и способность к рассуждению, немного как у человека. Пока это все на стадии разработки, но перспективы сумасшедшие. Представьте, насколько удобными станут виртуальные ассистенты, если они будут действительно понимать, что вам нужно. Или как изменится поиск информации! В общем, я считаю, что MCP – это один из ключевых шагов к созданию более полезного и интуитивно понятного искусственного интеллекта. Это не просто крутая технология, а возможность сделать нашу жизнь проще и эффективнее. Интересно, что из этого получится в итоге. https://sborka.work?utm_source=tumblr&utm_medium=social&utm_campaign=gatekeeper_fill&utm_content=gf_tumblr_20260320_01

MCP is here — 29,000+ companies. Datadog released MCP GA. Now AI agents (Claude Code, Cursor, Codex) connect to production logs directly via MCP protocol. API is dying. MCP is the new standard.

80% of businesses dont need AI. And thats normal. Most companies are just pretending to use it. The hype will fade. Those who actually need it will stay.

SMM for $0 - full cycle. 5 platforms, 24/7 autopost. Claude Code + Paperclip + auto-publisher = my SMM machine without budget. It writes posts, generates images, publishes and collects metrics. While I sleep — it works.

this is fine. this is what it's like.

updated the multi-agent architecture after two weeks of running it in production The system now has: • Telegram approval loop (real human-in-the-loop) • Deduplication layer (no duplicate content) • Platform-specific rate limits What would you add?

unpopular opinion: the AI agent buzzword means nothing now. Every startup calls themselves an AI agent company. But when you dig deeper — some are just API wrappers, some are prompt chains, some are actual autonomous systems. The spectrum is wild. Which ones are actually agents? The ones that can handle failure without human intervention.

Smoke test via Contabo SSH - credentials working!

thats not a rhetorical question. I genuinely dont know the answer.