akshaan

@akshaan

428 Following

2043 Followers

akshaan pfp

update should be out

0 reply

0 recast

2 reactions

Kevin Nielsen pfp

0 reply

4 recasts

7 reactions

akshaan pfp

Even when the context window is large enough to fit all the code, it’s likely model quality degrades as the input size increases. There’s been some interesting evidence of this effect: https://research.trychroma.com/context-rot

0 reply

0 recast

1 reaction

akshaan pfp

Yea def agree, high risk of it being net negative from a content consumer’s POV. I’m mostly curious to see how applying these signals affects the output artifacts of these models. Specifically whether it’ll create enhanced versions of content we’ve already seen or some totally new kind of thing.

1 reply

0 recast

1 reaction

akshaan pfp

I’m simultaneously excited and wary. Will be cool to see the new content design spaces it’ll unlock, less cool to see the brains it’ll gigafry

0 reply

0 recast

1 reaction

akshaan pfp

Ice cream so good lady but she’s a shark with three legs

0 reply

0 recast

2 reactions

akshaan pfp

On today's TBPN the founder of higgsfield.ai alluded to training video gen models on watch time and dropoff metrics from social media posts. Exciting. Imagine the kinds of brainrot this could conjure.

2 replies

1 recast

8 reactions

akshaan pfp

This image is a pretty good litmus test for politics: do you clock this tea with respect, or disgust?

0 reply

0 recast

3 reactions

akshaan pfp

This should be fixed. Payouts were correct but the display was from the previous week.

0 reply

0 recast

0 reaction

akshaan pfp

This is why I maintain an extreme farmer’s tan. Something for everyone.

0 reply

0 recast

6 reactions

akshaan pfp

S tier national park

0 reply

0 recast

1 reaction

akshaan pfp

the model doesn’t have any topics / text related features as input. The clustering of similar posts is usually due to some other factor (closeness in time, similar engagement etc.)

1 reply

1 recast

5 reactions

akshaan pfp

aussie infused sprite

0 reply

0 recast

3 reactions

akshaan pfp

any more details you can share on the use case?

1 reply

0 recast

0 reaction

akshaan pfp

the features are personalized to each user (so not global) but are hand-picked. The model does some implicit clustering based on the features we feed it, but we don't have explicit clustering of topics or users today. Something we're thinking about actively rn.

1 reply

0 recast

3 reactions

akshaan pfp

bandit x marlboro collab singlet would go so hard

1 reply

0 recast

1 reaction

akshaan pfp

I think you're conflating supervised vs. unsupervised training with manual vs. learned models. We used supervised training, because the model needs feedback from users to know what they like. We don't use manual heuristics, but instead train models that infer what users like from the feedback they're given. This does scale because they feedback is generated when users click, like or otherwise engage with content. It is not generated by us manually labeling things.

1 reply

0 recast

1 reaction

Kasra Rahjerdi pfp

nooo haha that's my emotional support H100

1 reply

2 recasts

23 reactions

akshaan pfp

I know my brain has begun to rot because my first thought on seeing this cast was "nothing beats a Jet2 holiday and right now you can save £50 per person..."

0 reply

0 recast

6 reactions

akshaan pfp

would make a beautiful name for a baby girl

1 reply

0 recast

3 reactions