Andrew Olsson

art/code/video. Building to give agents cloud superpowers. Making videos on my AI adventures...

[Andrew Olsson @andrewjolsson]

unusually is the new goblin

[Andrew Olsson @andrewjolsson]

@sama Will this be available to all API users or just special pre-selected enterprise customers?

[Andrew Olsson @andrewjolsson]

I just stumbled upon the CMS of the AI era: I'm working on a personal project (https://t.co/vOWiheIvsw) and I had Codex setup the site with Astro. I had it create a system where each post is a markdown file, and codify the steps for publishing a new post in the AGENTS.md file. So now when I want to post, I just give Codex the text/image. It makes the post, updates the site, and redeploys. If it wasn't clear already, Wordpress is dead.

[Andrew Olsson @andrewjolsson]

For 30 years I thought soap just killed germs, then I learned that it mostly just makes them less sticky. I just had Codex built me a visualization of this process: https://t.co/iOtdbvgWfe

(the things we do with 40% usage remaining and a weekly reset incoming...)

[Andrew Olsson @andrewjolsson]

@alexgetmancom @thsottiaux @OpenAIDevs @ClaudeDevs I made a video about this happening to me in APRIL and I still haven't heard back after appealing, a thousand comments on that video from others in the same situation, it's probably not from using the harness but who knows... https://t.co/Da2ltg7GoU

[Andrew Olsson @andrewjolsson]

@Mappletons You too?! I thought it just hated me...

[Andrew Olsson @andrewjolsson]

@sl1ma4 I won’t change your mind, I’ll just spam my video about Jevons Paradox that explains why this is happening…
https://t.co/6mMPnJahLH

[Andrew Olsson @andrewjolsson]

@kelvinbuildss I’ve built a tool to do just that and yes I trust it b/c it’s built on a planning protocol that makes it safer than just hoping the model does the right thing.

[Andrew Olsson @andrewjolsson]

@FlowHaa It’s great! it feels very magical to just tell codex to deploy it and it just works, even with databases and storage connected. I’d love to hear your thoughts if you try it out!

[Andrew Olsson @andrewjolsson]

@alaphati_t I'm building Bahama (https://t.co/45V4AeEXrd). Cloud toolkit for agents. handles hosting, DBs, connection strings, everything you need to tell your agent "deploy it" and it’s done.

Open source, creates one standard that works with popular cloud tools, feels a bit like magic..

[Andrew Olsson @andrewjolsson]

Today I’m launching Bahama, a cloud toolkit that helps coding agents deploy the apps you build.

It's open source, installs with one line, and works with cloud providers you already use (like @vercel and @neondatabase).

https://t.co/8q9HF5oNgb

[Andrew Olsson @andrewjolsson]

I made a new video on the Jevons Paradox that explains how AI broke software as we knew.... https://t.co/6mMPnJahLH

[Andrew Olsson @andrewjolsson]

@ProfRobAnderson Have you read the book, Cheap Land Colorado? It’s a very good account of these type of lots and the people who live on them.

[Andrew Olsson @andrewjolsson]

Does YouTube really think this is how I want to get new video ideas? More importantly, do they really WANT me making videos like this?

[Andrew Olsson @andrewjolsson]

I made my video about how "harnesses are the next big thing" a few days too early...

[alex zhang @a1zhang]

Transformers struggle to generalize to tasks they were not explicitly trained on. Instead, we propose in 2026 that it is the job of the harness to generalize through composition.

We observe a powerful property when training RLMs: for tasks with shared structure that look different, the root model naturally learns the same trajectory, meaning it views the two task trajectories as the same! In other words, the Transformer does not need additional generalization capabilities to transfer capabilities from one task to the other, the harness induces it.

We find that well-designed harnesses form a quotient set over task trajectories, meaning their individual LLM calls can see structurally “similar” tasks as near-identical, token-for-token! Harnesses can effectively generalize for the Transformer during training, without relying on any intrinsic generalization capability from the model.

For example, RLMs can see problems of different lengths as the same: we show that RLMs can train exclusively on short tasks, and fully generalize to similar but unseen tasks 8-32x longer because it produces near identical trajectories for both.

Taking this further, we show that tasks across different domains (e.g. math solutions vs. essay writing) that share a decomposition strategy exhibit the same generalization effect. RLMs can train on the problem of finding which essays belong to the same author and improve performance on finding math problems that share similar solutions.

The full blogpost, experiments, and discussion are in the thread below.

[Andrew Olsson @andrewjolsson]

In 2026, how you use the AI model matters more than its raw intelligence... but how do you use it better? With a harness. These guys took frontier models from 13% to 99% with a better harness and I made a video about it: https://t.co/tz4K6S10aK

[Haven Feng @HavenFeng]

Today, we’re introducing [schema]: a harness reaching 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on ARC-AGI-3 Public set.

[schema] makes an LLM think like a physicist. 🧵

[Andrew Olsson @andrewjolsson]

@HavenFeng I was so impressed by the implications of this that I just made a video about it: https://t.co/tz4K6S10aK

[Andrew Olsson @andrewjolsson]

@steipete You were right to pause here because—and this is the key part—I found several smoking guns.

[Andrew Olsson @andrewjolsson]

Fable is the first model that can read my mind. As a model is doing UX work, I typically preview the changes and write out fixes to send on the next turn. Fable just fixes everything before I have a chance to ask.