@BernardMarr The gap we're seeing isn't technical adoption, it's organizational. Most companies are still structuring teams like it's 2015 while buying 2026 AI tools. The infrastructure is ready but the operating models aren't.
@KanuGulati@recraftai@AnnaVDorogush@khoslaventures Small teams with a clear technical thesis tend to move faster than anyone expects. Recraft's focus on controllability over pure prompt-following seems like the right bet — that's what actually matters for production use.
@edsim The scanning problem gets worse when every team ships their own bespoke code. We've seen orgs with 600+ internal apps — each one a unique security surface. The real unlock isn't better scanning, it's constraining how things get built in the first place.
@saammotamedi What's your take on evals as the bottleneck vs. the primitives themselves? We've seen teams get religious about measurement but still struggle when the underlying components aren't production-hardened.
@BernardMarr The gap we're seeing isn't technical adoption, it's organizational. Most companies are still structuring teams like it's 2015 while buying 2026 AI tools. The infrastructure is ready but the operating models aren't.
@paulg The hard part isn't getting CEOs to build with AI, it's getting them to ship the messy v1 instead of endlessly prompt engineering. I've watched more exec demos than production deploys. Building matters, but shipping teaches you what actually breaks.
Mind blown -- @garrytan's #gbrain is INSANE! All of those things I had forgotten, couldn't find, gone forever, are browsable, searchable, queryable!!! #musthave
@stephzhan@karpathy@sequoia The production bottleneck isn't agent capability, it's human judgment bandwidth. The teams winning aren't chasing autonomy, they're optimizing how fast humans can feed judgment back in. That's what raises the ceiling.
@eladgil The gap tightens for teams shipping weekly. SOTA helps, but iterating with users matters more. Quarter-old models beat cutting-edge in retention—we see it constantly. The real edge: execution speed, not model timing.