Dev.to WebDev πŸ›  Dev πŸ‘ 0 πŸ“– 2 min read

I just soft-launched my startup, for a second time

Lots to say about the platform I've been nurturing and growing over the last 15 months. A platform intended to handle (at scale) the long-video to short-video pipeline for content creators and streamers. It's been noth

Lots to say about the platform I've been nurturing and growing over the last 15 months. A platform intended to handle (at scale) the long-video to short-video pipeline for content creators and streamers.

It's been nothing short of painful, but the complement of rewarding sits perched next to the torment of bugs and user-issues.

"If you shipped your product in a state you liked it at, you shipped too late" -- someone who made a lot of money launching broken products

Soft-launched one week ago, ChatClipThat had a record 50 trial users day-of. No ads, all organic. All completely 100% SEO.

This is the second time I've soft-launched, all with the intention of advertising at scale and en masse; although yet to do that, I've been slowly coming to reality with the concept of nothing ever being perfect, and people who want to be a part of the journey.. well, they'll be there regardless.

I was in beta for a very long time collecting user-data, feedback, and testimonials. I had stepped through several evolutions of architecture, processing, and distribution, ultimately to end up in a constant tail-spin as I chased bugs, doubted my development ability, and prayed to God.

So, what changed?

The models, the agentic harnesses, and simplification.

Firstly, Muse and Claude have stepped up their game from the days I started. With Muse not even being a glimmer in Zuck's eye at the time, and with Claude having rolled out Sonnet 3.5 just then, the work was mostly MVP at best.

Secondly (and thank God for capitalism), with what OpenAI and Anthropic were able to create from sheer competition, the harnesses that steer agents and their decisions were also made better. Independent of the model, the problem solving in the models' harness became a pillar of refuge.

Lastly, the over-engineered pipeline resembled a mad scientist’s attempt to recreate an intuitive body from rigid schematics, each part sensible on its own but refusing to function as a cohesive whole.

That's to say, it should have worked in theory.

The audio, the video. The amplitudes, the changes in color, the motion. And much, much more.

All these are raw numbers and pieces of data that we can use to our advantage. The heuristics, if you will.

Too bad none of that tells us whether the moment is actually worth watching!

So, let's stop collecting all this data until we have a supercomputer, and instead let's fall back to a simple orchestration between LLMs, where instead of having gates upon gates and revisions and checks, we have one agent tell us what it saw, and one agent pick the windows.

That's it!

(Small side tangent: they aren't "agents". They are LLM API calls, rigid in structure with no tool calling ability, but as the idea of simplification stands, so too is language afflicted by our lust of reduction)

So, ChatClipThat now is soft-launched for the second time, and it's because of better models, a better harness, and a simplified approach.

Stop thinking, start building.

ChatClipThat

πŸ“° Read the original article on Dev.to WebDev

Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.