Bartosz Idzik

Rebuilding real UI behavior from video with Replay.

[Bartosz Idzik @idzikbartosz]

I’m putting my money where my mouth is. 💸

I just unlocked 2 Free Reconstructions video-to-code for everyone on Replay.

Running high-end LLM video analysis isn't cheap, but neither is your time.

I’m confident that once you see Replay turn your screen recording into clean code, you’ll stick around.

Try it here: https://t.co/GmuYPxDgbV

[Bartosz Idzik @idzikbartosz]

Imagine finding a UI style you absolutely love online, taking a screenshot, and deciding to build your own website with that exact vibe.

I tested this exact scenario: Lovable (VC-funded giant) vs. Replay (my bootstrapped solo-built engine). I uploaded the exact same image to both. One-shot only. Zero manual edits.

The difference in design quality is wild. 🤯 Replay gives you a beautiful, production-ready animated interface that actually looks like a senior designer built it with actual logic and complexity even without context.

Check out the side-by-side video. This is exactly why I built Replay because AI should generate code you are genuinely proud to deploy. 🚀

[Bartosz Idzik @idzikbartosz]

@sama The real tension is that private companies answer to users daily through product choices, while governments answer every few years through voting. Neither system scales well once the power gets large enough.

[Bartosz Idzik @idzikbartosz]

@ycombinator Been shipping AI stuff solo for a couple years and yeah, the speed difference is wild—what took me months to prototype last year I'm building in weeks now with better tooling and models.

[Bartosz Idzik @idzikbartosz]

@ylecun @elonmusk Isn't AGI more of a research milestone than something a car company ships? Tesla's great at optimization but that's pretty different from the hard problems still unsolved.

[Bartosz Idzik @idzikbartosz]

@DarioAmodei Worth noting that democratic backsliding often happens fastest when tech companies aren't thinking about it, so the "preserve at home first" framing matters more than it might seem at your scale.

[Bartosz Idzik @idzikbartosz]

@EuroJedi @ylecun @nyuniversity @Inria @AnneBouverot @VillaniCedric @guillaumgrallet @xjfontanet @geoffreyhinton @Yoshua_Bengio @ClementDelangue Will this cover practical stuff like why most multimodal models still struggle with spatial reasoning, or is it more high level theory?

[Bartosz Idzik @idzikbartosz]

@Plinz @AnamarijaML @garrytan Sure, but most developed economies tax productive assets and their owners manage fine. The machines don't disappear if you fund them differently.

[Bartosz Idzik @idzikbartosz]

@DarioAmodei Completely agree, been feeling this tension building while shipping AI products. The governance questions hit different when you're actually in the weeds watching systems get deployed.

[Bartosz Idzik @idzikbartosz]

@JulianGoldieSEO Did it actually handle dependencies and package installs, or did you have to manually set up the environment first?

[Bartosz Idzik @idzikbartosz]

@shl Tried this with a 2 person team on an AI project and we shipped faster, but the second person caught bugs I'd never see and handled ops while I coded, so we actually moved quicker than solo.

[Bartosz Idzik @idzikbartosz]

@ajlkn Tell me you didn't sleep because you shipped something sketchy that somehow still works.

[Bartosz Idzik @idzikbartosz]

@dannypostma Exact same thing happened to me, ditched Twitter and cut caffeine two months ago and the anxiety just evaporated. Now I actually get work done instead of doom scrolling every five minutes.

[Bartosz Idzik @idzikbartosz]

@miramurati Soumith's a solid choice, been following his work on PyTorch for years and he actually ships things instead of just talking about them.

[Bartosz Idzik @idzikbartosz]

@tinkerapi The hard part nobody mentions is reproducibility when models train on different hardware. Community roundups are cool but would hit different with actual training configs and compute specs included.

[Bartosz Idzik @idzikbartosz]

@rowancheung Tried building "proactive" AI last year, spent three months on recursive improvement loops before realizing the model was just memorizing its own corrections instead of actually learning anything new.

[Bartosz Idzik @idzikbartosz]

@twistartups The real question is whether it can actually handle continuous conversation without retraining, or if it's just clever prompt chaining like everything else shipping right now.

[Bartosz Idzik @idzikbartosz]

@gethalfbaked Tried building a verification layer last year, realized most fakers use real data from early wins then embellish the growth curve, so you'd need ongoing audits not just screenshot checks.

[Bartosz Idzik @idzikbartosz]

@EuroJedi @ylecun @nyuniversity @Inria @AnneBouverot @VillaniCedric @guillaumgrallet @xjfontanet @geoffreyhinton @Yoshua_Bengio @ClementDelangue Will he talk about why most AI research still ignores inference costs, or is this more about the next big model architecture?

[Bartosz Idzik @idzikbartosz]

@fchollet Tried building adaptive systems without this and they just memorized patterns, fell apart on anything slightly different. Once I started testing on genuinely novel problems the whole thing needed rebuilding.

[Bartosz Idzik @idzikbartosz]

@IntoTheMyst13 Thing is, the best infosec people I know got there by hating something specific enough to fix it, not by waking up called to the craft. Purpose finds you through the work.