oso (@osoleve) header
oso (@osoleve) profile picture

oso

@osoleve

data scientist

 he/him
github.com/osoleve
Joined May 2008
  • Tweets29.5K
  • Following670
  • Followers817
Pinned Tweet
oso (@osoleve) profile picture
oso
@osoleveSep 9
Looking For Group! Data Scientist (Neutral Good) Multiclassed from Linguistics Ships Open Source Stats: 12+ YoE in data ops in regulated domains 2x DGX Spark homelab Proudest achievement: nominated by the food bank for the governor's volunteer service award (for data work)
oso (@osoleve) profile pictureoso
@osoleveSep 3
Career update: the world is big and scary but I've decided it's time for my next adventure If you know anyone who needs a data scientist/linguist with over a decade's experience in data mining and curation, who is weirdly into text distance metrics, hit me up 🤓 cc @max_spero_
1
2
1
14
46,356
Hot damn Christmas came early! 65B base model! Thank you @DZhang50 🙏
Dan Zhang @ Ricursive (@DZhang50) profile pictureDan Zhang @ Ricursive@DZhang502h
Replying to @osoleve @RicursiveAI
here you go: huggingface.co/ricursiveAI/R…
ricursiveAI/Ripple-65B-Base-Preview · Hugging Face

ricursiveAI/Ripple-65B-Base-Preview · Hugging Face

Machine learning model with 65B parameters. Download, explore files and run it on Hugging Face.

huggingface.co
1
73
Man goes to doctor. Says he’s depressed. Doctor says: “Treatment is simple. Famous clown, Pagliacci, is performing tonight. I’ll pick you up at 7. Wear that dress I like.
12
240
5
5,240
83,945
Just got a phone call from an AI. It was almost natural feeling and fairly productive. What the fuck
1
86
@boardyai I was unfamiliar with your game
1
2
17
Just hummus... For you and meeeee
brent 🌹🗽 (@BrentInQueens) profile picturebrent 🌹🗽
@BrentInQueensNov 6
This right-wing AI slop was somehow meant to be racist but it goes so unbelievably hard. I have been singing “Rajesh says no beef upon the grill” all day.
39
How do we turn this into a roguelike
Jameson Hodge (@jamesonhodge) profile pictureJameson Hodge
@jamesonhodgeNov 7
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
1
18
1,005
JK that's the whole benchtop computer for codex to flash devices from. Hell yea, it works
Image from oso (@osoleve)
oso (@osoleve) profile pictureoso
@osoleveNov 8
Deeply unserious bench camera
Image from oso (@osoleve)
1
2
80
Deeply unserious bench camera
Image from oso (@osoleve)
2
1
3
150
If it's stupid and it works...
1
18
oso retweeted
mr. joshua (@pants) profile picture
mr. joshua
@pantsNov 8
Image from mr. joshua (@pants)
24
736
11
12,984
144,719
This stupid fucking tweet just changed my life
joyful critter (@masochistmartyr) profile picturejoyful critter@masochistmartyrNov 7
so was spongebob :)
19
1,768
7
40,244
725,278
If you want to see how brittle your own benchmarks are, glitchlings has you covered 👾 github.com/osoleve/glitchlin…
Justin Lo (@loresjpg) profile pictureJustin Lo
@loresjpgNov 7
Nice DeepSWE score! Does it survive a lowercase filename? Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement. Models clustered around 74% on DeepSWE end up 23 pp apart. Good at coding, or good at DeepSWE? 🧵
Image from Justin Lo (@loresjpg)
2
148
This 97 year old horse still replyguys the old fashioned way.
Max Spero (@max_spero_) profile pictureMax Spero
@max_spero_Nov 6
Replying to @paularambles
Lately I’ve had to block 5-10 replybots per post
49
actually testing kind of a neat idea here: can we selectively apply recurrent depth to tokens at syntactic boundaries during reasoning and then coast at depth 1 between? will report back
oso (@osoleve) profile pictureoso
@osoleveNov 7
it's important that you constantly invent new notation for your agent to parse so it doesn't get too comfortable or good at what it's doing
Image from oso (@osoleve)
1
2
171
Update: task I chose was too hard for the model under normal conditions, oops
1
22
Just keep swimming
15
it's important that you constantly invent new notation for your agent to parse so it doesn't get too comfortable or good at what it's doing
Image from oso (@osoleve)
1
2
209
Gemini will occasionally have the sauce. Not often, but sometimes.
Image from oso (@osoleve)
1
51
You can get 2.5x throughput with Dory 1.0 4B with this One Simple Trick* *might be lobotomized, we'll see
4
228
I use the word "fauxperiment" once and now Codex uses it for everything
1
5
181
you're babysitting 6 hours of grid rollouts don't call it that now i'm worried you're not taking it seriously
8
Load more