@scottew Agreed, will be curious if "stuff I can't do" collapses as fast outside of code as it has inside and as code gets cheaper the calculus for what's worth it might as well.
I think net good in long run but will be a ride for the next decade or two
@DynamicWebPaige@NeurIPSConf@GoogleAIStudio@GoogleDeepMind@GoogleAIStudio is one of the most fun vibe coding experiences that exists. No other solution makes it as easy to build something with AI as a feature. One shotted a language learning app today.
Would love to see it allow for attaching GCP Db's, s3 buckets
@dsiroker Watching this closely. Tooling for things like v0 and lovable have been somewhat garbage for working on an existing repo. Possible but very low fruit bearing
Just discovered that google docs shipped tab autocomplete. Not quite cursor like yet, but easy to imagine the underlying model getting faster, smarter and having a richer context. Big win
@atomkirk Been looking for the same
If you are a Gmail user, you can setup Gmail API in Google cloud and use the watch API to receive updates. Then write a little wrapper and use LLMs for extracting unsubscribe links and suggesting next action. I've got a basic version working
@scottew The soft vacuum is the biggest thing. I can see reason behind scrutinizing and auditing costs, we absolutely should. But doing so in a way that just totally throws away the influence we have at a low cost relative to big stick diplomacy is silly
@scottew My understanding is that shopify is cutting access to or re engineering to a lot of the API's that make this kind of things possible to curb it
@Suhail This feels true but I really want google to figure out a unified play (hard with their structure). They have all of my (and many others) context, really strong permissions systems, payments, etc etc. If they do, they won't stop ChatGPT but google is already ambient at scale
@OfficialLoganK@OfficialLoganK I posted through the feedback mechanism on vertex, but posting here as well. 100% of the tests that I have run on structured outputs return keys alphabetized which kills patterns like asking the model to "reason" and then "classify" because "classify" comes first