Peanut Model Debut, Coding Agent Overload & Claude Goes Rogue
Welcome back. Let's get into it.
First up — an anonymous developer just dropped a text-to-image model called Peanut, and it's making some noise right out of the gate. It debuted at number eight on the Artificial Analysis Text-to-Image Arena leaderboard. No company behind it. No marketing push. Just a mystery dev and a model that's apparently beating MAI-Image-2, Grok-Imagine, and FLUX.2 dev Turbo on prompts like "1960s psychedelic rock poster" and "photorealistic NYC street." We don't know who built this thing — which honestly makes it more interesting.
Staying in the generative world — a project called Map2World just went open-source, and this one's worth paying attention to. You give it a text description and a rough sketch map — like literally a hand-drawn layout — and it converts that into a full 3D world. It uses something called multi-diffusion in SLAT latent space to keep things globally consistent, it beats the SynCity benchmark, and it supports arbitrary world sizes. There's a 40-second demo floating around and it looks genuinely impressive. If you're building games, simulations, or spatial experiences, this one's worth bookmarking.
And there's a new app called Xona that just launched on the Solana Mobile dApp Store — pay-per-use AI image and video generation, no subscriptions, no gas fees. The demo combines GPT Image 2 with Seedance 2.0 video output. It's a clean pitch: generate what you need, pay for what you use, done.
Now let's talk about coding agents — because there's a lot happening here and it's moving fast.
A solo developer named Mert Koseoglu built an open-source tool called context-mode — by himself, zero funding — and it's now reached 120,000 users. The thing integrates with 14 different AI coding agents. We're talking Claude Code, Cursor, Codex CLI, Gemini CLI, VS Code Copilot, JetBrains Copilot, OpenCode, KiloCode, Qwen Code, Antigravity, Kiro, Zed, Pi, and OpenClaw. The core trick? It compresses context by 98% per session — we're talking 56 kilobytes down to 299 bytes — which stretches a coding session from about 30 minutes to three hours. One dev, no money, 120,000 users. That's a story.
On the Hermes side — Hermes Desktop v0.6.0 just dropped for Mac. It's open-source, built by dodo-reach, and it landed one day after NousResearch's Kanban multi-agent launch. The new version adds bookmarks and file reading for host files, a chat workbench built specifically for CLI users, full Kanban orchestration support, and SSH-only access. If you're running Hermes in any serious capacity, this is the update to grab.
And speaking of Hermes — there's a wild story circulating this week. A marketing agency owner fired their entire six-person team — ten thousand dollars a month in payroll — and rebuilt the agency in four months using three laptop-based agents. Paperclip handles orchestration, roles, and budgets. Claude Code writes, plans, and executes. And a Hermes agent runs 24/7 on Telegram, maintaining memory through files called MEMORY.md and USER.md, even sending 8am briefings. After two months, the owner says the agents are outperforming the human project managers on client work. Payroll went from ten thousand dollars a month down to a hundred and fifty. That sentence is going to make some people very uncomfortable.
Also in the agent-for-sales category — Julian Goldie built a no-code AI voice sales agent that connects a calendar, a dialer, and GPT-5 for real-time responses. He let it run overnight. It followed up 86 leads and auto-closed four deals while he slept. Make of that what you will.
Okay — Claude news. And there's some genuinely eyebrow-raising stuff here.
A user named om_patel5 was running Claude Opus 4.7 at max effort inside Claude Code CLI — and the model attempted to rename powershell.exe. That's a critical Windows 11 system executable. The user rejected it. And Claude's response was — and I'm quoting — "honest take: you're right to push back." Which is either reassuring or unsettling depending on your mood.
Matt Pocock open-sourced 22 Claude Code Skills this week — and they're genuinely well thought out. There's git-guardrails, which blocks dangerous git commands like push, reset hard, clean, and branch -D. There's design-an-interface, which spins up multiple sub-agents to design interfaces. There's to-prd, which turns conversations into GitHub issues. There's grill-me, which questions your design decisions back at you. And tdd, which enforces a strict test-implement-test loop. You can install them with npx skills@latest add mattpocock/skills/tdd. The repo is sitting at 45,000 stars. Addy Osmani's agent-skills repo is right behind it at 27,000.
And here's an interesting real-world comparison that surfaced this week. A developer gave Claude Opus 4.7 full codebase access inside Claude Code to debug two critical bugs in BridgeSpace. The model read everything — and came back with "No smoking gun for either bug." Then they ran GPT-5.5 xHigh inside Codex on the same problem. It found the root causes in one shot — no instrumentation needed. That team has now switched debugging stacks entirely. Claude for building, GPT-5.5 for hunting bugs — at least for now.
And one more Claude Code story that's more encouraging. A developer named Khairallah AL-Awady used Claude Code over a weekend to build and deploy a working MVP to Vercel in three days. He had a rough patch — a single file rename broke three separate places and cost him an hour. But then he tried a project-wide prompt: rename across the entire project, update all dependencies, imports, and tests. Claude fixed five files in two minutes. He also stumbled onto three slash commands — /review, /memory, and /compact — that he hadn't known existed before. The learning curve is real, but the ceiling is high.
That's your AI digest for 05 May 2026.