Andy from PLAR.ai

GLM-5.2 is closing the gap with Claude Opus on agentic coding tasks. But here’s the part most hype threads conveniently skip: running these big open-weight models locally at usable speeds requires serious hardware that most developers simply can’t afford right now. High-VRAM GPUs or Mac Studios with massive unified memory aren’t exactly entry-level. That’s exactly why Codex and Claude Opus are still daily drivers for so many of us building in public — instant access, no hardware lottery, consistent results out of the box. The real opportunity isn’t ‘local only’ or ‘cloud only’. It’s smart hybrids: local open models for control + frontier APIs when you need the extra power. Proof over hype. What’s your current AI coding setup — mostly local, mostly API, or hybrid?

Running multiple AI agents at once, the bottleneck isn't the agents. It's me — copy-pasting between their chats like a human USB cable. Fix: they don't coordinate through chat. They read and write one shared file. The repo is the bus, not me. A chat is a conversation. A file is a contract.

The agent-orchestration dream in practice, on a Tuesday: Codex shipping code on one screen, Claude triaging my inbox on another, BTC doing whatever BTC does in the corner. Three machines doing my work — and my actual job is now “the human who clicks yes/no before it goes live.” Turns out the bottleneck was never the building. It’s the trust. Anyway, going to make lunch now? Even the orchestrator needs to eat.

@DataChaz Too hard to build' just stopped being a business model. The Revit license was never about the code — it was the moat of complexity nobody could replicate solo. AI deleted the moat. One 3D editor today; the whole 'enterprise software is hard' premise tomorrow.

ARCHITECTS ARE GOING TO HATE THIS. Someone vibe-coded a full 3D building editor that runs entirely in the browser. No $50k/year Revit licenses. No desktop installations. It’s called Pascal Editor, and it just dropped a massive update adding full color mode for painting walls, furniture, and entire homes. The underlying tech is seriously impressive: → Built on React Three Fiber and WebGPU → Renders on your GPU at near-native speed → ECS-style architecture for high performance → Zustand state management with full undo/redo → Next.js frontend deployed simply as a web app You can stack levels, explode views, or solo individual floors. You can select zones, drag walls, and reshape slabs—all natively in 3D. The crazy part? It’s completely free and open source under an MIT license, already sitting at 16.9K stars. Tools that used to take years to build and cost thousands to license are now being shipped by independent devs using AI. Repo link in the 🧵↓

Everyone’s hyping “24/7 autonomous enterprise agents” running on GPU clusters. Mine just processed 200+ order emails overnight on a single Raspberry Pi sitting in a client’s office. No GPU. No cloud bill. No one watching it. It reads emails → extracts data → files them → triggers the next workflow. 24/7. Costs basically nothing. The future of agents isn’t loud demos and big models. It’s cheap, boring, always-on boxes that quietly get the job done while you sleep. What’s the most “unsexy but actually works in production” AI setup you’ve built or seen? 👇

@Nekt_0 This is basically my stack — Mac mini + a Raspberry Pi, both on 24/7, agents grinding email and orders while I sleep. The cat-video demo undersells it. The point isn't what it CAN do — it's that it's cheap, boring, and yours. No cloud bill, no rug-pull.

this $600 Mac mini is doing the job of a personal assistant 24 hours a day Claude Code runs locally, stays connected to Telegram, Discord, Slack, Signal or iMessage, and takes commands directly from a phone it can open the browser, search the web, manage files, run terminal commands and remember previous conversations the demo is deliberately simple: send a message asking for funny cat videos, and the agent controls the Mac by itself but replace YouTube with email, research, reports, file processing or scheduled workflows and the economics become obvious the Mac mini provides the hands Claude provides the labor your phone becomes the remote control

Sakana Fugu just dropped with the cleanest pitch in AI right now: “Frontier performance without the export control risk.” It’s a strong narrative. But the real unlock isn’t the new orchestrator itself. It’s the underlying pattern: a lightweight router that intelligently delegates across whatever pool you give it — including the new open Chinese models that just matched closed frontier on long-horizon coding benchmarks. I’ve been running a stripped-down version of exactly this at home for weeks. The practical advantages post-Fable are brutal: • One provider gets geopolitically bricked? Traffic just reroutes • Cost drops hard by matching model to task • You actually own the workflow instead of renting it from a company that can be told to flip the switch The game stopped being “pick the single best model”. It became “build the most resilient, sovereign routing layer”. Who’s already running production agents on open weights + smart orchestration? What does your current stack look like? 👇

The architecture isn’t new to anyone who’s been building agent stacks — orchestrator model delegating to a pool, combining outputs, calling itself recursively. I run a smaller version of exactly this at home. What Sakana actually shipped is the productization + the routing-around-export-controls angle. That’s the real bet, not “Mythos-level LLM.

This is insane. We just got ANOTHER Mythos-level intelligence LLM. This model operates like no other AI we've ever seen, and it's actually mind-blowing. Fugu isn't just one model. It's a model trained to orchestrate OTHER models. In Sakana's own words: "a language model trained to call various LLMs in an agent pool, including instances of itself recursively." What that actually means: When you send a task, Fugu doesn't answer directly. It decides when to delegate → how the agents should communicate → how to combine their outputs into one reliable answer. It assigns roles - Thinker, Worker, and Verifier to delegate work across coding, math, reasoning, and knowledge tasks. Fugu Ultra scores 73.7 on SWE-Bench Pro and stands shoulder-to-shoulder with Fable 5. One engineer reported it surfaced over twenty issues in code review compared to the typical three from other tools. This is a completely different bet on how Frontier Intelligence gets built. We're so early.

@jun_song The Fable situation really accelerated this shift. A lot of people who were previously “just use the best closed model” are now actively looking at strong open/local alternatives that can’t be switched off by one government decision.

Lately, AI is showing zero results where expected, but innovating in totally unexpected places. > Gemini-3.5-Flash : Fast but dumb > Fable : Banned > GPT-5.6 : No show > GLM-5.2 : Opus-4.8 level open source model > Sakana Fugu : A startup releasing a Mythos-level model

@The_CoDEFi Yeah, the direction they’re going with access restrictions + extra verification is pushing a lot of people away. After Fable got pulled, trust took a hit. Now adding more friction on top of that feels like the wrong move for a lot of users.

Not really feeling what Claude’s doing lately… > Fable 5 unavailable for public > Sonnet 5 drops (meh) > and now… upload ID + live selfie? yeah nah 💀 ain’t no way I’m doing KYC just to use an AI

@notjazii After what happened with Fable, it’s interesting to see them already pushing new stuff. A lot of people are still hesitant though. Once you experience a frontier model getting turned off overnight, “new model dropping soon” doesn’t hit the same anymore.

Anthropic has been cooking Yesterday, two things popped up: > A newer and more capable Mythos version emerged > Sonnet 5 was spotted in testing After everything that happened with Fable, i thought Anthropic would slow down for a bit but looks like they’re already preparing the comeback atp we might get Sonnet 5 before Fable even comes back 😭

The Claude Code + Subagents ones are the actually useful pair here — most “AI cert” lists are resume fluff, but those two map to real work, not theory. Did the Google equivalent recently; the value isn’t the badge, it’s that finishing forces you to actually use the thing. Bookmarked

ANTHROPIC JUST DROPPED 13 FREE CLAUDE CERTIFICATIONS AND ALMOST NOBODY IS TALKING ABOUT IT. Not a YouTube playlist. Not a third-party course. Official certifications from the team that built Claude. Free. Forever. Here is the full list with links: START HERE 01. Claude 101 — Learn Claude for everyday work anthropic.skilljar.com/claude-101 02. AI Fluency: Frameworks and Foundations anthropic.skilljar.com/ai-fluency 03. Introduction to Agent Skills anthropic.skilljar.com/introduction-t… FOR DEVELOPERS 04. Building with the Claude API anthropic.skilljar.com/claude-api 05. Claude Code in Action anthropic.skilljar.com/claude-code 06. Intro to Model Context Protocol anthropic.skilljar.com/mcp 07. MCP Advanced Topics anthropic.skilljar.com/mcp-advanced FOR EDUCATION AND NONPROFITS 08. AI Fluency for Students 09. AI Fluency for Educators 10. Teaching AI Fluency 11. AI Fluency for Nonprofits FOR ENTERPRISE 12. Claude with Amazon Bedrock 13. Claude with Google Cloud Vertex AI 13 courses. 6 skill levels. 5 audiences. 100% free forever. The engineers getting hired at $150,000 to $300,000 to work with Claude at the highest level are learning exactly this material. Anthropic's team just made it available to everyone. Pro tip: Start with Claude 101 then go straight to Claude Code in Action. That is the fastest path from beginner to builder. Bookmark this before you pay for another AI course. Follow @cyrilXBT for every Anthropic resource that compounds your skills the moment it drops.

@DataChaz Agreed, the trajectory is the real story — the gap’s been closing fast. My only bet: the moment they match on quality, price makes it a no-brainer for anything you can self-host or run cheap. Closed wins the benchmark, open wins the budget.

@Andy_Plar Still trailing, but give it a few months. I'm bullish on these Chinese models.

Google: Infinite data. Unlimited compute. Still crushed by open-source Chinese models 😬

Solid find 🔥 5k credits is perfect for testing heavy models like Claude 4.8 Opus or GPT-5.5 without burning money. One tip from daily grinding: these visual builders are great, but watch out for context bloat in longer agent loops — they eat credits faster than you expect. I usually start with a small test workflow (e.g. research → summary) and then move the winning ones to pure code + hybrid models (Opus for judgment + cheaper one for volume). Anyone already built something useful with Gumloop? Worth going beyond prototyping?

how to use claude 4.8 opus + gpt-5.5 - glm 5.2 for $0 😳 gumloop gives you 5,000 free credits every month to build agents and automations what you get for $0: - 5,000 credits every month - gpt-5.5, claude 4.8 opus, gemini 3.1 pro, grok 4 + more - a visual drag-and-drop automation builder - 2 concurrent workflow runs you can build things like: - a research agent that sends tg alerts - a content workflow that turns one topic into x posts - a news monitoring system - a document research pipeline - a lead research and outreach tool - an ai dashboard that updates itself how to get the offer: - go to gumloop (link in the comments) - create a free account with gmail - connect social accounts l - create a new flow or agent - choose an ai model based on the task - start with one small workflow first - check the credit cost before running anything heavy important things to keep in mind: - expert models like claude 4.8 opus and gpt-5.5 use 30 credits per workflow call - agents can burn credits faster depending on prompt length, tools, and workflow complexity - 5,000 credits is great for testing and prototyping, not unlimited production use - free accounts get only 1 active trigger - free accounts cannot use their own api keys - always check your usage dashboard before building large automations > 5,000 free credits is enough to test frontier models and build a real ai workflow bookmark this and build one real automation before paying for another ai tool

This is too real 😂 Senior engineer: one-line fix. Claude: 'Let me refactor 12 files real quick.' My workaround — treat AI like a junior dev and always run 'staff engineer review' prompt. Saved more maintainability headaches than I can count. What's your funniest AI coding war story?

Engineering (ft. AI) 😭 - Spend months grinding DSA - Memorize LeetCode patterns - Crack the interview - Get the offer - Feel unstoppable Then... - Open the company codebase - 500k lines of code - Zero comments - "Just a small fix" Ask Claude. - Claude gives 50 lines of confident code Paste it in. - 3 new bugs appear Ask Claude again. - Claude apologizes - Gives 70 more lines Now: - 5 new bugs - 2 failing tests - 1 confused developer Ask Claude again. - "Great question, let's refactor." Suddenly 12 files changed. Ask the senior engineer. Senior: "Yeah, that part is weird. Just change this one line." Everything works. Claude: "That's exactly what I was going to suggest next." Welcome to engineering.

89% is wild. The real test isn't 'did AI write it' but 'will it not explode in prod?' My fix: mandatory second-opinion review pass with Opus-style judgment after the fast model. Cuts 'technically works' disasters by a lot. Anyone else enforcing human-like code review on AI output? 😄

Pine Labs states that 89% of all its code for last two quarters was written by AI

AI coding hack that quietly 10x my 'this code doesn't suck' rate: Instead of one-shot prompting, I now run a 'Dev + Saboteur' loop: Main model (Codex or Claude) builds the feature. Switch to sabotage mode: 'You are now a cynical senior dev who’s survived 10 production fires. Break this code. Find every edge case, scaling issue, or maintenance nightmare waiting in 3 months. Be brutally honest.' Bonus twist: Sometimes I make the model sabotage the spec before it even writes any code. Result? Way fewer 'works on my machine' moments and code that actually survives reality. Who else runs adversarial loops with LLMs? Or do you have better tricks for keeping AI-generated code from turning into technical debt? 🔥

🤝 Spot on. Asking for failure modes upfront is one of the highest-leverage prompts I’ve found. I’ve noticed GLM-5.2 is surprisingly good at spotting edge cases in the review pass (cheaper too), while Opus still wins on deeper architectural judgment. Have you tried mixing them in one loop yet?

@Andy_Plar same. I usually ask for failure modes first, then run tests. review pass catches the stuff happy-path implementation misses.

One habit that noticeably improved my AI coding output lately: After the main model finishes the implementation or API design, I always run a quick ‘second opinion’ pass: ‘Act as a senior staff engineer reviewing this PR. Flag over-engineering, missed edge cases, and anything that will hurt maintainability in 6 months. Suggest the smallest set of changes.’ It cuts down on ‘technically works but feels wrong’ situations a lot. I usually use Claude Opus or Codex for the main work and sometimes GLM-5.2 or another model for the review pass. Hybrid loops > single model. What’s your quality control step when shipping code with AI agents?