Tomas Halgas

Ever wanted to talk to someone who’s completely transparent with you?

You can now have a chat with Chimera.

Built with @Lovable of course.

chimera-chat-voice.lovable.app

We just replaced Firestore subscriptions at @Lovable with a system built on NATS, cutting application payload by 97% compared with sending full results.

Over 2 million concurrent subscriptions at peak. An estimated 40 TB → 1.11 TB actually sent in 30 days.

What’s crazy is that @AtaMarzban94 built the core in 40 days, and we rolled it out without a single user-facing incident.

Using proprietary software that 80% fits used to be fine, now you can just build your own that fits you perfectly. It’s like why would someone buy somewhat-fitting mass-produced clothes if they could easily make their own bespoke?

Software engineers want a spotless codebase but are fine with five forgotten mugs on their desk.

Then I watched this video about people who clean as they cook, so the kitchen’s done when dinner’s ready.

And it all makes sense now.

Why train new models when you can just ask an LLM to write the weights?

An LLM wrote a model that plays a simplified version of the brick-breaking game Breakout and scores 20× as much as one trained for 500 million game steps. Other demos got perfect scores on small tests of addition and sorting.

All without a conventional training run. Machine learning will never be the same again!

Now for real.

A lot of these demos are basically an LLM wiring up little circuits inside a neural network. Programming in a very awkward language.

One demo used roughly a million neural units to store the answer for every allowed input to a basic arithmetic problem. A few lines of ordinary code could calculate the answer directly.

These are useful test cases for understanding neural circuits. But if the claim is that they replace training, why encode the algorithm in a neural network when you could just write it in a normal programming language and run it?

Circuit demos by @csinva
csinva.io/weight-space-p…

The results I find more interesting are the repairs to existing models.

In one experiment, an LLM edited just 0.016% of an existing language model’s weights. Its accuracy on a maths benchmark jumped from 7.9% to 48.1%.

The model had been getting stuck repeating itself. The edit helped it stop looping and finish its answers. This took tools and repeated tests, and the same approach didn’t meaningfully improve another model that rarely looped.

Breakout and Qwen loop-repair experiments by @rikiyatakehi and colleagues
rikiyat.github.io/blog/weight-sp…

These demos haven’t shown us how to build a smart brain from scratch, but they did show how we can use one AI to diagnose a problem inside another…

…and perform the AI equivalent of targeted brain surgery.

I’ve built products before AI and with AI, and the difference in how fast you have to rethink things is pretty wild.

At @Lovable, it’s even more obvious. Things we couldn’t imagine building three months ago are becoming foundational building blocks.

100,000 pull requests later, we have a few stories to share.

Today we’re launching the Lovable Tech Blog. @antonosika’s kicking it off with how we built Lovable and why we keep rebuilding it. After that, more of the team will share what they’re building, trying and learning along the way.

I get to see this work up close every day. Really happy we’re opening it up for everyone else to see too.