Jackson Kernion

Jackson Kernion

@jacksonkernion.bsky.social

Now: LLM finetuning, human feedback at Anthropic. Before: MIT postdoc, UC Berkeley philosophy PhD. Built http://deductivelogic.org

1,038 Followers 134 Following 16 Posts Joined Aug 2023
1 year ago
Post image

Ocean Beach in February

1 0 0 0
1 year ago

For this kind of thing, we already have superhuman AGI. Or, whatever is possible is possible with today's LLMs. I can imagine that the UX gets better than claude.ai projects, but in terms of machine intelligence, higher scores on SWE-bench don't really help.

0 0 0 0
1 year ago

Someone told me about a day they spent talking with Claude using a project with hundreds of daily journals, doing whatever Claude recommended to have a good day. It sounded like it was a real back and forth, ultimately a pretty interesting day in the park and doing deep work.

0 0 1 0
1 year ago
Post image

SF in February is so pleasant

1 0 0 0
1 year ago
Post image

SF in January is so pleasant

2 0 0 0
1 year ago
Post image

RL'd LLMs can learn *anything*, so long as you can reliably evaluate 'correct' behavior and have enough examples/compute.

I think the most important question now is: how does performance on arbitrary cognitive tasks correlate with RL'd cognition?

1 0 0 0
1 year ago

Indeed it is, everyone should live here!

0 0 1 0
1 year ago
Post image

SF in December is so pleasant

1 0 1 0
1 year ago

A priori metaphysics of mind is what allows you to make inferences from empirical work. It's more picked-over and there's maybe less progress to make, but it's seen as more 'core' because it's more important.

1 1 1 0
1 year ago

The world is all that's in the prompt.

0 0 0 0
1 year ago

I remember doing a FB post about which direction people stand in the shower and people went mad

2 0 0 0
1 year ago

Do you have something new and different you'd be excited to work on? Or, alternatively, some engaging extracurricular to look forward to over the holidays?

0 0 1 0
1 year ago

2017 Facebook philosophy was at least a ton of fun. It approximated a global salon, with (too) little inhibition. It seems now an impossible dynamic to recapture.

4 0 0 0
1 year ago

Are you on the 18.2 beta? The preview images I saw looked...not enough like Midjourney for my tastes

0 0 1 0
1 year ago

I was talking to my therapist today about a stressful situation and mentioned how having access to Claude made me a lot more confident and comfortable in knowing how to respond.

Having an always-available, immediately responsive counselor is a real benefit!

4 0 0 0
2 years ago

I still don't understand. This is the sort of thing I used to use Twitter for. I found myself search for videos on TikTok.

0 0 0 0