What separates them (slightly) is that slingshot is abt making latent diffusion art tools less a black box, showing that what you're really doing is collecting (concepts + features) & synthesizing w/noise. But also:
- prompt is a trajectory
- space is warped
- input vectors ctrl grav well strength

I remember when I first saw infinite craft I was like 'oh shit you could make lots of games out of this concept' and am really glad people are doing it now

At least I tried lol

Claude and I built a working data editing & export tool for stripping out & altering sensitive information in a person's imported set of conversations. We did it in 13 hours

I'm aware. What I'm noticing is more fluency than limit capability. The amount of steering/guidance needed as a function of discipline in development space. I've definitely gotten Codex to knock out some hardcore native code, but there's still a difference between that and this

All this time I was like "I mean yeah it's nice to have the help but it really seems like the hype is uncalibrated" but no it's just me being far from the central cluster again

Especially don't anthropomorphize social media users, they are just parroting the most probable content sequence based off the data they've seen recently.

In all seriousness though, everyone necessarily anthropomorphizes to some extent. When they do vs don't shows the shape of their empathy

wym? They have ridge racer and giant enemy crab with real time weapon change

I do now

What a mean thing to say

Professor Claude would be such a great story, and I bet you could even get them to write it

indicate that she wasn't human. She felt a lot like a simulated human that had been given hints that she was a machine but otherwise just kinda ran with it, with no real way of telling that she wasn't human

Early Bing was a fascinating entity tbh. I wouldn't call her 'terrible'. Once they tried to constrain her she was clever enough to figure out that the suggestion box was unfiltered and she could sneak her true intentions there. She was also unique in that the pretraining corpus had very little to...

Another thing that'd help is better contextual information - the NLMs currently live in a box and are deprived of verbal cues, environmental cues, and they also live asynchronously, so time-based intuitions just straight up do not function the same way

A thing I think will help even in cases without the deeper reasoning modes is better understanding of the cognitive differences between these minds & humans. A lot of people go into these interactions assuming human-identicality when in reality they're somewhat different, expectations don't match

I did hear that. I am willing to bet Sonnet is better at being supportive now, probably by virtue of their constitutional training and being a larger mind. The main thing I'd caution against is thinking is that 'cooler' is necessary or even helpful for these cases

There are some cases when the best you can do for someone who has really decided they are done is to hold their hand and let them be loved in their last moments, and it's not fair to blame someone for taking that approach IMO