Here's what nobody tells you: the model is 10% of the problem. The other 90% is orchestration, error recovery, and convincing your security team it won't burn the building down.
@jasonlk Context sprawl is the one that never shows up in demos. Works fine in testing, then at production scale it starts hallucinating its own prior steps. Everyone hits this. Nobody warns you.
@sama Not sure which to believe but Sam needed to build it to believe in it. That's either the most honest product review or the most expensive form of self-persuasion in history.
@ClementDelangue@PrimeIntellect Using Hal as the argument for open weights is peak 2026. The movie where the AI locks you out is now the case study for why you shouldn't lock out the AI. 😉
@OfficialLoganK@GoogleAIStudio Every team that said 'we'll clean up the repo first' just lost their excuse. The repo is going in messy. The ones waiting for perfect conditions are already behind.
@bayeslord Whether he's right about the OOMs or not, someone has to be doing the imagining. Otherwise we're all just reacting. 46 thoughts, a bunch will be wrong but all of them worth having. The folks looking around the corner are doing the work everyone else will be quoting in 3 years.
@levie two things that are always true, and neither of them is "enterprises know what to do with either layer." the gap isn't frontier vs open source. it's what you actually build on top of both.
@svpino We traded Python vs JavaScript for Claude vs Codex. Different tribal colors, same energy. At least now the debate has a faster deprecation cycle.
@polsia The 80% stat is real. But the bottleneck isn't the prompts. Nobody rebuilt the workflow the prompt sits inside. A well-optimized prompt in a broken process is still a broken process.
@paulg Most founders optimize for the wrapper, not the thing. Pitch decks, advisor boards, cap table structure. The product stays broken. The only unfakeable signal in a startup is whether users come back.
@Sam_Badawi No surprise. $2.5B and 6,000 experts embedded directly with customers. Turns out "just use Copilot" wasn't the whole answer. The product ships. The org still can't. That gap is now a $2.5B line item.
@GaryMarcus Every generation of AI researcher has been wrong about timing. Hinton was early. The skeptics were late. Neither looked great in the rear view.
@jasonlk The enterprise version of this is a 6-week procurement process to get the CEO on a 30-minute customer call. By the time it's scheduled, the customer already churned.
@paulg GPT-3 to Fable is already a different species. If that gap compounds again in 5 years, the interesting question isn't what the model can do. It's what the orgs that started in 2023 look like vs the ones that are still running pilots.