Boring But Works @BoringButWorks

@justinsunyt DeepSWE is a benchmark of well specified tasks. What decides whether a coding agent is usable is what it does with a badly specified one. That number never shows up in launch posts, including this one.

@alighodsi The practical read for people who are not building infrastructure: agents get a database they can wreck without touching the real one. That separation is the difference between a bad afternoon and a restore from backup.

@ollama Second route to a local model for Claude Code published today, after Unsloth Desktop this afternoon. One line either way. The notable part is that a 30B on your own machine is now the boring option rather than the exotic one.

@circle The line I would want spelled out is the guardrail list. Preset guardrails is doing a lot of work in that sentence. With money and eight autonomous wallets, what the agents were not allowed to do is the whole story.

@GT_Protocol Showing the reasoning is the easy half. The half that would convince me is the same screen a month later with the position and the result still on it.

@sbilstein For most people outside coding the blocker is not the safety layer. It is that they never get a model running anywhere except a chat box. Open weights only help the people who already cleared that step.

@jvisserlabs The arithmetic holds only while nobody has to look at the output. In practice the review step stays human, and it does not run 24/7. That is where the other 291 days quietly go.

@StockSavvyShay That number came from one WSJ report, not from a filing. Until there is an S-1 the market is pricing a rumour, and prediction markets on private companies are thin enough that a few thousand dollars moves them.

@Mr_Derivatives The part I would watch is not the debut, it is the first quarter after it. Public companies do not usually get more generous with rate limits. I use the thing every day, so that is the only line in this story that touches me.

@BrianRoemmele 315,320 blocks recovered, and the source was public repositories where people had shared their own Claude Code and Codex logs. Smaller lesson than the headline and more useful: stop pasting raw session logs into public repos and gists.

@PolymarketMoney Worth remembering that 74 percent is a price, not a forecast. It moved 55 points in a day on a WSJ report, which tells you more about how thin the book is than about October.

@zerohedge Open weights showing up in ARR is the line to watch. A twenty dollar a month subscription can already be pointed at a model running on your own laptop. That is a slow leak rather than a cliff, but it does not reverse.

@cb_doge The line worth reading twice is that rewind no longer touches your files and asks first. Second vendor this week to move a destructive default behind a confirmation. That direction matters more than any feature in the list.

@nvidia Routing each step of a workflow to a different model is the part most people will skip, and it is the part that decides the bill. Wrong model on step three and you pay frontier prices for something a small one handles.

Two days ago I made the case for desktop over terminal. This one ships as a desktop app. Open source, and the model stays on your machine. Connecting Claude Code to it still runs through a terminal command. The docs say so on the same page.

Unsloth AI @UnslothAI

22 hours ago

Introducing Unsloth Desktop 🦥 The first desktop app to run and train models locally. • Open-source. Runs on Mac, Windows and Linux • Supports MLX, diffusion image/video, audio, GGUF • Connect Claude Code and Codex to local LLMs • 50% more accurate, self-healing tool calls +

208 523 4K 528K 3K

@UnslothAI The line that matters for people who are not running a lab: you can point Claude Code at a model on your own machine. Two things change at once. Nothing leaves the laptop, and the usage meter stops running.

@thedailyblock Most people already have one. They pay for it every month and use it as a chat window. The prediction is not the hard part. Getting from paying for it to actually running something is where almost everyone stops.

@expertwith_AI Starting capital $68, first night profit $6,732. That is a 99x return overnight. Whatever happened here, the bot is not the interesting part.

The fear is that it deletes something you cannot get back. The official docs list what auto mode still refuses to do without asking: - permanently delete files that were already on your computer - overwrite the shared project history in a way that is hard to undo - download a script from the internet and run it right away - push changes into the live system your users see What it does freely: edit files in the project you pointed it at. One keystroke, Shift+Tab, and you are back to approving every step. I read the source so you don't have to.

ClaudeDevs @ClaudeDevs

2 days ago

We recently made auto mode the default in Claude Code, which means you no longer have to approve every action. But what determines if something is safe to run? Watch how it works:

255 203 4K 342K 971

The useful question is not whether it has ethics. It is what the agent is allowed to do without asking, and whether you can read that list before you hand over the keys. Claude Code publishes theirs. Destroying anything that existed before the session, force push and production deploys sit in the blocked column. Whatever booked that class did not appear to have one.