@IanCal@data-folks.masto.host (@IanCal) profile picture

Moving over to @IanCal@data-folks.masto.host

Joined August 2009
Replying to @ESYudkowsky @ylecun
There's no iteration inside the model. You can just build a layer on top though, or if you don't want to do that, tell it that it should follow steps like this. Not what you want for all problems though, not everything is a question or puzzle. v quick example:
1
81
Replying to @xeophon
It's not just changing the names of the objects from like "block" to "cheese" though. They also *deliberately* stop the model from talking through the problem.
1
239
Replying to @RouhiRoo
It's not-quite-almost-chistmas time! youtu.be/Ap9GqyLoWyY
1
Replying to @Pandamoanimum
It's almost upon us! Not-quite-almost-christmas-time youtube.com/watch?v=Ap9GqyLo…
Replying to @zdmc23 @Suhail
Honestly I'm surprised it did as well as it did, I have to go back and forth a bit to try and work out the words. It's only really the very quick parts it got wrong when fed an MP3 the chorus seems right. I tried with Rap God and afaict it got that perfectly right. Large model.
1
3
One of the most powerful questions I think is "what's the worst thing that could happen?" - leads to very useful discussions about failure cases & understanding what the business really cares about. Not my best talk but I gave one about this youtu.be/0znSrRLOYwY
2
Replying to @aphethean
This is great, I get 11s here - you can improve this a bit as well by converting to parquet (6s) and then querying that (<0.5s)
1
3
Replying to @IanCal @StabilityAI
I was even able ask for @gdavies fighting @AlexHorne over a sausage roll. Here's what it thinks that looks like.
1
1