Alexander Basis

This is my favorite translation 🤣🤣

@mattpocockuk @theo Curious if it affects Paperclip usage …

Anthropic has given us a "dedicated monthly credit" Which, in effect, slashes AFK usage limits of Claude Code by ~5-20X Here's how it affects you:

@dotta @rileybrown I’m just curious if the latest Claude news about credit limits to programmatic use affects Paperclip? Or Agents on the Claude Code will still be using sub limits?

@rileybrown You gotta try Paperclip (with codex) - you'll do 95% of your work on it but be able to keep track a week later x.com/dotta/status/2…

And I’m not talking about point of no return to Claude Code. If they release a model that’s better, great. I’m talking about using an agent interface (super-app) for ALL tasks. I don’t go to a browser anymore. I don’t go to email. I don’t go to notion. Every task that is started is kicked off via an agent chat. I believe over the next year this is inevitable for all knowledge work. If your app doesn’t open in the superapp browser you may be… Cooked.

@HeyGen @HyperFrames_ Shader

Another @HyperFrames_ catalog update Wednesday: html-in-canvas effects Today: textures and shaders Every effect is one command to add for your next video npx hyperframes add <name> Comment + RT "shader" for this videos source code (must follow)

Max reasoning effort makes your coding agent dumber, not smarter. Codex shipped with medium default. Arena's BullshitBench (155 nonsense prompts) showed high-effort GPT and Gemini burn 50K tokens on confidently wrong answers. I'm pinning mine to medium. Pin yours too.

Claude Code's subagents see a 2K-token summary of your conversation. Forked subagents see all 180K — for 10% the cost of fresh tokens. I fork for design, research, recap. Not for code review: a fork justifies the code it just wrote. Flag: CLAUDE_CODE_FORK_SUBAGENT=1.

My audit for this week: list every task your agent does. Mark each easy / hard / impossible to verify. Every "hard" with no verifier is where your agent silently breaks. Full talk (20 min): youtube.com/watch?v=XNtkiQ…

The sharpest line: "I don't think planning is going to stay around." Plans are brittle — at plan time you don't know the special cases. Skills encode judgment at the node. When skills run out, the agent asks, then writes to a decision log so it never blocks.

Legora ships AI agents to 600+ enterprise legal clients without using chat. Their CTO says chat is structurally wrong for agents. It's low-trust AND low-control, the worst quadrant of agent design. Here's the framework he uses instead.

On April 22, Qwen's 27B open-weights coding model beat its own 397B flagship on every major agentic benchmark. SWE-bench Verified: 77.2 vs 76.2. Terminal-Bench 2.0: 59.3 vs 52.5. SkillsBench: 48.2 vs 30.0. A dense model 14× smaller, under Apache 2.0, fits in 28GB VRAM. One RTX 4090 with offload. Comfortable on a 5090 or an M4 Max. Same day: • Anthropic pulled Claude Code from the $20 Pro tier, then reversed within hours after developers flagged the pricing-page edit. • Anthropic's Enterprise repricing swaps the $200/month bundled Code seat for $20 plus API-rate overage. A single 500K-input Opus 4.7 session now bills $22.50, more than the old monthly plan. • GitHub paused new Copilot Pro and Pro+ signups and removed every Opus model from the $10 tier. • OpenAI's Codex crossed 4 million weekly developers. Altman committed to resetting rate limits at every additional million up to 10M. Anthropic's Head of Growth compared the situation to a power utility shedding load — the first time a frontier lab has named the analogy publicly. The flat-rate $20 coding subscription is repricing toward the ceiling on every vendor at once. Same week, a single-GPU open-weights model outperformed a 397B system on the exact workload that's making the economics break. I'd run Qwen3.6-27B against my own repo this week, before the next price-page update.

3. Three lint rules: no dynamic imports, no raw SQL outside one query file, unique function names. 4. WHEN_TO_SLOW_DOWN.md listing where agents propose, not commit. 5. 200K is the real context ceiling. 1M is marketing. Start with 1. 30 min. Hits first.

2. For any LLM output you ship: writer → reviewer in a FRESH context window → editor. Same model fine. Different prompt, no exposure to the writer's reasoning. Fixes the model's bias toward liking its own work. Prompts can't fix this. Separate context can.

Flask's creator shipped broken code through an AI agent last week. From the stage: "I committed it with the agent and just didn't really see it." The fix isn't slowing down. It's five structural practices that rebuild friction where agents delete it.

Meta: 'Token Legend' leaderboard. Salesforce: $175/mo AI-spend floor. Coinbase fired a $400K engineer for slow adoption. Orosz closed AI Engineer: lines-of-code was this mistake 30 years ago. Tokens repeat it. Ask your lead: what outcome are tokens a proxy for?

Not sharing the YouTube links is intentional 😂

For those who want to stay in control — not just be a background agent for Claude Code