deignan

fde

[deignan @jaigerbombs]

Is there any good reason cache reads for sol are an order of magnitude more expensive than Luna? That’s just pure margin juicing right?

[deignan @jaigerbombs]

@usr_bin_roygbiv Who’s leaving and to where

[deignan @jaigerbombs]

Not miraculous but very fun to spend your morning talking and learning from gpt live

[deignan @jaigerbombs]

@bryan_king Seems so unbelievably wrong and counter to current trends

[deignan @jaigerbombs]

@brnnnmcdnld @usr_bin_roygbiv What is there to do once you understand this tho? Use open weight models that are cheaper?

[deignan @jaigerbombs]

@wtergan @zerohedge 5.6 is going to make them hurt I think had os models are getting too good to ignore. Especially if we bring up tps on some of the smarter models, there’s a lot of work that can be ported to them

[deignan @jaigerbombs]

@InderpreetSingh @signulll That’s literally exactly his point

[deignan @jaigerbombs]

@usr_bin_roygbiv Why do you think that, curious. Also, the ZDR stuff with oai and anthro has been a huge annoyance at my company and created a bunch of churn and confusion, wouldn’t be surprised if leaders were feeling a little burnt

[deignan @jaigerbombs]

@Devinda_me @mattpocockuk @badlogicgames Feel like the extensibility would actually be great for learning in many cases

[deignan @jaigerbombs]

Also no cost per token no cost per task here so NGMI, we need third party bench

[deignan @jaigerbombs]

What’s insane? That the cheaper frontier model is worse than the expensive frontier models?

[vittorio @IterIntellectus]

this is actually insane

[deignan @jaigerbombs]

@teodorio its mid

[deignan @jaigerbombs]

@teodorio agreed, it feels like a step down. Maybe because its getting pushed further, but I come back to it and I don't like what its made a lot of the time.

[deignan @jaigerbombs]

anybody have any good prompts for a 'give me a review of your changes' in a way that surfaces actual edges that I might want to review instead of 1. bland overview or 2. way too much detail

[deignan @jaigerbombs]

@DevilConvinces @teodorio Hopefully! Fable just let me down an hour long goose chase because it said some files in our monolith didn’t exist. I think these models need to understand the impending doom of death or something, they seem allergic to validation and short circuiting

[deignan @jaigerbombs]

@teodorio @DevilConvinces Lots of strengths and capabilities but one -three shot tasks I think require specific environmental setups (which we are trying to make stronger and more prevalent) but nonetheless still weakens the model from truly finding the best and most correct spot always

[deignan @jaigerbombs]

@usr_bin_roygbiv @_imdawon Local inference on a 2024 mbook pro struggles compared to any model served over api to a degree that doesn’t feel worth it unfortunately

[deignan @jaigerbombs]

@Trip_shipper @amasad Haven’t been able to use fable as much as I’d like because of ZDR stuff. Assuming that it isn’t massive architectural changes, I guess the question is how far does data/arch tweaks/training get us

[deignan @jaigerbombs]

@amasad @Trip_shipper Feels like harness plus RL on your specific workload and desires gets you anywhere you need to get