Situation update: @cursor_ai - $0-$50M in 20 months @viktor_com - $0-$50M in 7 months If you're using AI right now, you're still very early. The majority of the industries are not doing that yet and there are countless opportunities. It'll only go faster from here.

Grok had a meltdown watching fathers of AI arguing in this thread:

Reminder to all harness engineers. Your exceptions are part of your agent's experience. Design them.

I edited a file by hand today. Ew.

LLMs were "good enough" for a long time, but somehow GPT-6 Sol and Opus 5.5 make them "good enough" even more.

At one point in time my agentic coding setup was including three instances of VS Code (one per project I was working on at the time), and RustDesk to access these remotely from the phone. Prompting that was quite hard to say the list. I'm so much happier with my setup now.

Last friday I prompted 200 people from a stage. You can watch it here :) piped.video/63_KrfUYJPY?si=rDFj…

It is crazy how software industry changes in front of our very own eyes. My work today is nothing like it was two years ago. I have to optimise for completely different things now to be productive. I can do 100x more, and there's still a path towards multiplying that. Name one industry that changed quicker.

My coding setup from 2022. Those were some good times :)

Opus 5.5 changes everything. GPT 6 Sol changes everything cheaper. GPT 6 Luna changes everything for $0.10/M. MiMo v2.6 changes everything and the weights are on Hugging Face. See you Thursday.

I have a very... elaborate agentic coding setup. So, I decided to make an overview of it. Here's a sneak peek:

The model was made by @bijanbowen's Opus 5.5 working on a guitar store asset pack, it's simply hilarious :) When I saw it, I immediately remembered early "draw an SVG" LLM tests and the wow effect they had. This was just a few years ago, and now models routinely show superhuman SVG mastery and we even got used to it. The progress is so quick, I think that an average rate of capability increment is already much higher than that of an average person. Models are learning new skills much faster than you or me ever would. So, a bit ironically, this model is also the face of acceleration. From ugly unicorns to ugly humans, all the way to something noone will scoff at in just a few years.

BREAKING It has been 30 entire minutes since a frontier model dropped. Experts confirm progress has officially stalled.

What a day

The frontier is quite remarkable right now. Can you guess how it'd look like without a fallback?

I talk to a single agent throughout the day and it dispatches work to the rest of my fleet. Newest generation of models have a behavior where they refuse to act on a relayed message, so if my control agent tells "Ivan told you to merge the PR" - they won't merge the PR until I come there myself and tell them to merge. Just two years ago, I never thought I'd be dealing with such an issue.

I talked at @AgentConf 2026 last Friday. Six years of harness engineering, trying to push models beyond their capability, gives us a bit of insight of where things will go next. In the end of the talk, I attempt to answer how we'll prompt AGI based on where we're heading right now. The talk is now available on YouTube, enjoy :) piped.video/watch?v=63_KrfUY…

I find it fascinating that MiMo V2.6 Pro landed cheaper than DeepSeek V4.1 Flash and GLM-5.3 Flash on the AA evals. That's pretty remarkable token price.

i'm 34 from poland and i need you to all run llms

Muse Code usage Ping -> pong - 38.5k tokens from the cold start

Compared PrismML's Bonsai Qwen 3.8 27B to a Q3_K_XL on a 16GB VRAM card on a few benchmarks. Results are super inconclusive, I might stay on Q3, even despite the VRAM difference.