←
sankalp48,410 posts
sankalp avatar

sankalp

@dejavucoder

looking for beauty in frontier ai and culture into ai engineering, benchmarks, evals, rl envs, post-training and such rl resident @primeintellect

◷ Joined October 2021▧ 10,054 Photos and videos
sankalp avatar
sankalp@dejavucoder·54m

aging like milk

sankalp@dejavucoder17 Oct 2025

oh husbant... karpathy sensei said AGI is a decade away... we have 10 years to escape the permanent underclass now!

Post attachment
sankalp avatar
sankalp@dejavucoder·5h

can google drop a model better than fable 5.2 one day before anthropic ipo. stay tuned.

sankalp avatar
sankalp@dejavucoder·13h

they are engrossed in the details of their life. often dealing with their own shit. i find listening and asking them more questions (especially *new* people) pretty entertaining. you can just go on and on. i tend to learn new stuff plus it's also a good break from my own mind.

sankalp avatar
sankalp@dejavucoder·13h

i find it interesting that when i people meet irl or like just chat, people tend to yap more about themselves and their lives. sometimes you would expect them to ask questions or stuff you are not very comfortable to talk about but they won't.

sankalp avatar
sankalp@dejavucoder·13h

remember anon

sankalp@dejavucoder28 Jun 2022

most validation(v) in already established interactions are pretty useless other than providing a temporary sense of excitement (which I like). Although the v provided by folks who can directly influence your reality and reality itself should be taken seriously. h/t Taleb

sankalp avatar
sankalp@dejavucoder·Nov 17

where is thinky buying their data from is the question

You don't need many edits in weight space to fix a model In the WSP blog, GPT-6 Astra fixed a broken car policy by changing 6k weights and cut Qwen3.5-0.8B's math looping by changing 130k Inkling 2 Small did the same with just 10 and 8k edits: ~600x and ~16x fewer, with no degradation in performance on either task

3:08
sankalp avatar
sankalp@dejavucoder·Nov 17

as i said, every model release these days is an advertisement for opus 5.5 and fable 5.2. anthropic cooked very hard.

sankalp avatar
sankalp@dejavucoder·Nov 17

based on benchmarks, astra 6.2 is fable tier or better in agentic work spanning various knowledge domains + science and cheaper. somewhere around or lesser than opus 5.5 in coding perf but cheaper. sol 6.2 sol looks below or equivalent opus 5.5 and cheaper