How I Messed Up Our AI Strategy: The Big Pivot That Didn't Pack a Punch
Superthread founder David Hasovic gets blunt about the failed first attempt at going all-in on AI, from model selection errors to the 'better product' trap.
May 17, 2025
|
David Hasovic
The Big Pivot That Didn't Pack a Punch
At the end of 2023, I made the decision to radically change Superthread and go all-in on AI. However, I messed up. Here’s what happened.
When we started, before the modern LLM era, our goal was simple: build the fastest team collaboration tool that’s actually useful. That said, my friend Hugh Hopkins from Delegate (Acquired by Landbase) was already obsessed with ChatGPT 1, and I should have paid way more attention!
Anyhow, we made some big strides and the result was a product that was about 20% better than the competition and in some cases 100X faster. But it turned out, users neither cared about small improvements nor speed that much.
Coming from B2C, I learned a hard B2B lesson: the better product doesn’t always win. That’s a topic for another time.
Ultimately, we didn’t get the knockout punch I originally envisioned. That breakthrough moment, the big leap forward, felt further away than ever. And that’s when I hit a wall. I was stuck, trying to crack a new model for collaboration dynamics, but nothing was clicking. It was honestly a low point.
Then, suddenly, it hit me, the answer was staring me in the face all along. I told the team, “Starting January, we’re going all-in on AI.”
But, well… things didn’t exactly go as planned.
Catching Up vs. Pushing Ahead
It’s January 2024, and we’re kicking off the year with fresh energy.
I’ve always had a thing for AI. Back in secondary school in 1997, I built my first chatbot using QBasic, and my first job was with a startup run by one of my AI professors. With new LLMs, the time felt right; we jumped in and committed to making our company AI-driven.
Soon enough, we had our own set of tools up and running:
RAG search
Text rewriting for different tasks and documents
Duplicate detection
Auto-tagging
It was a solid start, and within a few months, we’d caught up with the industry. But there was a snag: we were only catching up, not pushing ahead. We were doing the same standard AI features as everyone else, but nothing groundbreaking. We weren’t focused on where AI could go next; we were stuck where it was.
Execution Failures and the Fizzle
Execution was another challenge. We initially picked the wrong models for vectorisation, and our auto-tagging model, trained on GitHub issues, didn’t work well either, it could only handle bugs and requests!
Eventually, my enthusiasm fizzled, and the project stalled after a few months. I’ve been here before; I need that spark early on or I find it hard to stay excited.
Looking back, not pushing Superthread to become a full AI company was a big miss. Now, we’re making up for that and aiming to get it right. Once we’re there, we’ll be sharing some of the magic we’re creating.
Stop using tools that are stuck in the past.
Superthread is built for high-performance teams that don't have time for clunky, legacy workflows. Experience the speed for yourself. Sign up for Superthread for free.
