Tomas Halgas (@TomasHalgas) header
Tomas Halgas (@TomasHalgas) profile picture

I lead technical evangelism at @lovable_dev. Built Sutro (acq. by Lovable) and Sphere (acq. by Twitter). What stays scarce when software gets cheap?

Joined May 2012
Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)Photo by Tomas Halgas (@TomasHalgas)
My conversion to marketing is now complete. Years as a software engineer, CTO and CEO. Then a few months running technical evangelism at @Lovable and suddenly I'm in "Marketing Lead" category. Am I still allowed to call myself an engineer?
11
28
1,806
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?
2.6 billion updates in 30 days. Over 2 million concurrent subscriptions at peak. That’s @Lovable Realtime, the system keeping our browser tabs in sync. I built the core in 40 days. We rolled it out to every user without a user-facing incident. Here’s how it works.
2
5
1
33
1,669
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.
3
16
668
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-progr… 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/weigh… 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.
5
1
16
407
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.
4
2
2
30
1,766
It’s finally happening: companies are scaling revenue without growing their teams. I don’t think people realize how big this is. More small teams will be able to run huge global businesses. But it’s uneven. Tech stands out and the US stands out. Two takes: 1. Big Tech can grow without adding jobs. If promotions still depend on bigger teams, the internal politics could get ugly. 2. Europe needs to move fast or get left behind.
7
26
609
If you’re like me, you have dozens of unfinished projects. Built in Replit, Bolt, Base44, v0, Manus, or locally with Claude Code or Codex? Bring a code ZIP to @Lovable. Lovable rebuilds it, hosts it, then helps you with content and ads. Maybe this is the one you finish. lovable.dev/free-import
5
11
440
London has become a bit more @Lovable. Our new London office is on track to be one of our largest. And yes, we’re hiring 👀 lovable.dev/careers
14
8
184
10,956
Only in San Francisco do tech founders get the rockstar treatment. @FabianHedin and I got invited on stage with @tiesto and @TheChainsmokers. The duo also run a VC fund. So technically, this was a founder-investor meeting.
18
2
1
78
5,102
San Francisco is a magical place. You message a friend to catch up and 20 tech founders show up. Thanks for hosting me for a fireside chat today, @vinit!
1
23
604
Spent the day with @FabianHedin at @Cloudflare Connect in SF. AI web traffic passed human traffic in May. Cloudflare forecasts 1,000x more AI browsing within 5 years. You spent a decade designing for thumbs. Your next visitors might not even have hands.
3
1
11
559
“Grow my app. make no mistakes” is finally a thing! @Lovable now helps you grow your business!
Lovable can now help you grow your business. Launch ads directly to Google, YouTube, Facebook, and Instagram. Create images, video and marketing copy from your builds.
1
17
471
Offsite costs scale roughly linearly with headcount. Possible new connections scale with n(n−1)/2. Two weeks after @Lovable Croatia offsite, I’m still finding out how useful those new connections are in my work. Quite a good return on getting everyone in the same place!
3
14
652
How to make sense of the future of AI? Here’s a starter pack: - website Less Wrong - like a Wikipedia for making rational decisions - books Superintelligence and The Singularity is Near on how close to AGI we are - essay The Bitter Lesson - the idea that maybe more data is all it takes to get to AGI - read about scaling laws - it scales fast! - Effective Altruism philosophy - a take on what’s rationally good and how to measure it Credit to @benjaminvrbk et al.
4
2
1
20
1,528