Muhammad Umar says compute doubling has cut model pre-training time in half and will boost AI capabilities 20-30% in 2026. Claude recently generated a complete Kanban-board UI, but humans still need to handle design and full-stack work. The infrastructure behind this requires massive energy-intensive data centers.
One of the things that we are seeing, I mean is that the whole infrastructure, so the amount of time it used to take to pre train a model is now cut to half now. Similarly at the inference the whole memory infrastructure that we have now is far more than for example six months back. So if you see this, AI is going to improve by at least 20 to 30 acts only in 2026. What it means is better cloud code. So for example, I hear a lot that engineers are not able to design complete applications. There's something missing if you're using models to for example for better UI or ux. There are things that that require human input. I think these things you would see. I mean I was with the demo with Ali on Hyper, we were using Claude code at the last mile and on top of Hyper and what it just blew my mind how he built those Kanban boards and all that stuff, all those things and generated code for it. And just the need of UI and UX and front end engineers and back end and software. What I've seen firsthand is like what models can do. So this all is possible because of that compute infrastructure that's powering all this. And what it means is not just GPUs or these things. It means energy, very large data center. So I think in 2026 you would see models performance would improve to a point where some people would feel that AGI is either it's in the models now or probably a couple of months away. So for some people.
FDE Epi : 16 | AI Agents, Billion-Dollar Infrastructure & the Future of Intelligence
