@spotn00b@toddsaunders Claude supports controls. I can already limit what skills and connectors can access. Security and privacy controls seem like an easy thing to build. (model is way harder)
Imo it is a UX problem and the gap will get closed within months.
I see this even with lovable. Like today we use lovable to build a website. But at the end of the day I could have claude do the development of the website. The only gap for that transition today is the cloud infra (db, hosting, auth). Claude knows how to do that, the only issue is that the UI in claude is harder to use than lovable to do this. This feels like an easy problem for claude to fix soon.
Back on Opus 4.6.
After 45 mins of debugging with codex 5.3 to no avail for a nasty DOM issue. Opus took exactly 4 minutes to fix it (even threw in some clean up)
@thehegemony@EricLDaugh Except there are not enough Americans to perform these jobs. There is a reson that unemployment is as low as it is. Many of the jobs are highly specialized, you can't just take any person and say you are now a surgeon or civil engineer.
@karpathy Curious on your perspective on specificity of knowledge when it comes to this. Example When asked yo add 2 + 2 say 4 vs. When asked to add use formula x + y. I think there id a use case for each, but I tend to focus more of generics.
@hwchase17 Only if trust in them can be achieved. Currently we see users liking more the micromanagement route of approving smaller incremental steps that a single long running process with an unknown outcome.
@t_blom Will it also clone support, training, improvements, bug fixes, etc.?
The concept of cloning ia true from a pure code perspective, but the reality is that in SaaS the code is sometimes the least important part.
@asantanillaj Very analogous to hiring humans.
The part the is interesting here is how fast are the agents gonna beat each other?
Are you willing to pay a premium for something that might be out of date a couple months later?
Is winning one auction taking you out of competition in others?
What are the best image to video models out there right now? Did a quick test with @KlingAI_ and it was cool!
Animated our mascot for @VetRec Joey writing some notes.
In a world where all software is created bespoke by AI, the main challenge will likely be support, documentation and training.
Not because an LLM made it, means it will be easy to use / understand.
@pzakin Ideally the README is written by an LLM in a format that the same LLM can then read and show to the user. (IMO the IDE is simply something like Cursor or Copilot which is LLM powered.)