Peder Aaby

Can't help but feel this screenshot from Github Status is a bloody lie. There's been way more trouble the past 5 months than not working 0.3% of my time

@BlumeDotCodes Love the UI! I am currently searching for something similar calm and minimalistic for Patronus

@housecor Housekeeping is not fun, but needed! Working on monitoring and making these cleanups automatic, it's crazy how much outdated AI instructions ends up in the codebase for long living projects

Habit: I ask Claude to audit our instructions. "Audit AGENTS.md. Is the information in it valid? Does it match reality?" Claude found 7 instances of bogus, outdated instructions. We were literally lying to AI. Problem solved.

Blume.codes is launching on Mac! 8 days in we have 76 people on the waitlist and ready with the MVP. Today: Agent monitoring for Cursor, Claude Code, OMP, or Codex. Tomorrow: Codebase/chat analysis and context+setup fixes that make agents behave.

Team Claude, Cursor or Codex? Choosing Gemini is the AI equivalent of saying your favorite beer is water.

Setting up a container to automatically harness performance: cursor, claude, omp and codex. Interesting to look at the different system prompts, ie. from Codex: "NEVER use destructive commands like `git reset --hard" You just know that went really bad once💣

@ProgenticEngine And it works pretty well :) But there's still a long way for cursor to truly be an autonomous like harness

@pederaaby A harness + Kimi 2.6 is basically what Cursor 2.5 is…..

A good harness multiplies the effectiveness of the model. A proper harness + Kimi 2.6 can work better than Gpt-5.5 without an agentic codebase setup A good starter harness for devs: Claude code or Codex or with Andrej Karpathy skills (most popular repo on Github right now).

@vela_gao It matters so much!

@pederaaby For real. The model matters, but the harness decides what the model sees, which tools it can call, what state survives, and how failures become visible. That layer can change the whole result.

Whenever AI tells me confidently I'm right, I get extremely sceptical

We are creating an app that monitors your Claude/Cursor/Codex locally and improves your codebase. What features would you want to see? 1. Context pollution removal 2. Security steering 3. Statistics (how am I doing vs others) 4. Cross harness chat search Other suggestions?🌻

@won__sikkk Agreed. It helps with a good codebase setups, but the agents are sneaky

@pederaaby AI-generated code especially needs aggressive review because it looks clean but hides structural debt

Vibe coding comes with dangers! If you just relax and code without obsessively rigging your codebase and review code, your project will turn into a landfill of debt before you even realise it.

@rauchg Everyone can vibe code these days, but taste and competency still wins by far

How do you build a great brand? Build a great product

@imjcmartin 🍝

@pederaaby Just let the vibe coders get to 1000 hours on a project and they'll see

@NachoGZ80 I used to work on a gigantic codebase/as at a consultant. The entropy rockets correlated to number of devs as well. I remember my team laughed when we staffed from 20 to 60 people

@pederaaby The hardest and most intense battle we engineers face ( and one we never fully win) is the battle against ENTROPY in software. Uncontrolled entropy destroys any project with a force that no human or machine can overcome.

@cryptorover This is just a fear reaction by management / a wakeup call not to use Claude naively. With proper setups it is just so god damn useful and not that expensive, it will be back

🚨 THE AI COST CRISIS HAS STARTED. Microsoft reportedly told engineers to stop using Claude because AI bills were exploding, while Uber says its entire yearly AI budget was already destroyed by April.

@elonmusk 3x parameter size is🔥 Big hopes on utilizing the Cursor data, should be a proper goldmine

Grok foundation model V9-Medium (1.5T) has finished training. Evals look good. A lot of Cursor data was added in supplementary training and there is more to come. Fine-tuning is underway and reinforcement learning begins in a few days. 2 to 3 weeks to public release. This will be a major improvement over the 0.5T v8-small that currently serves all Grok production traffic, especially for difficult coding tasks.

@_kendev Then it feels like flying straight into a wall

@pederaaby It feels like flying until you try to refactor a broken real-time hook

@krishdotdev Opus 4.7 for complex tasks, GPT 5-5 as the workhorse. Sonnet is an amazing model, but OpenAI really stepped up

Who's your daily driver ? For me it's: Sonnet 4.6