oso
@osoleve
Followers
613
Following
59K
Media
1K
Statuses
22K
data scientist (derogatory) //this is stupid i'm tired i want a pretzel
he/him
Joined May 2008
👾 Glitchlings Dedicated Thread 👾 This will serve as my central explainer for Glitchlings, and I'll use it to hold future updates as well. Anyways, if you've never been held hostage by one of my explainers before, buckle up! 1/? https://t.co/1N9uk6Fl0w
github.com
Enemies for your LLM. Contribute to osoleve/glitchlings development by creating an account on GitHub.
1
1
4
👾 Glitchlings v0.7.0 👾 Been optimizing the hell out of the Rust pipeline. Most Glitchlings now >=3x faster! (measured @ 40k char length) Plus: - New experimental interface, Auggie - Fully merge Adjax and Reduple into Rushmore - Remove requirement of Python/Rust parity
0
0
0
0
1
4
It's so much fun that I apparently just lose the ability to meaningfully introspect for 24-48 hours after hitting a tipping point
0
0
0
Very cool paper. Would be interested to see if domain randomization helps push models towards the domain-adversarial goal implicitly, which I feel intuitively would make sense.
When you train a model on one dataset, it usually performs poorly on data from a different source - even if the task is the same. This paper shows how to train models that automatically learn features that work across different datasets by forcing the network to be unable to
0
0
0
The White House, famously mosquito-ridden
Tornyol (@tornyolsystems) is building micro-drones that kill mosquitoes. They use smartphone microphones, car park assist sensors, and some clever DSP and control to transform 40-gram toy drones into mosquito killers.
0
0
0
RIP wimpy you would've loved klarna
33
677
6K
I hid myself for years thinking I'd exist in public when I was "worth" being seen. Funny how much faster I healed and grew after I finally decided to just show up anyway.
putting your creation out in the world is the best way to expose the parts that need to be healed, and then healing has a purpose of expanding your ability to be creative.
1
2
56
if you want the tweet version and not the 10min video version: this is now all it takes to train with prime-rl after installing verifiers
verifiers v0.1.7 is released 🚀 this one's all about making RL training and experimentation waaaay easier: - single-command installation for prime-rl - single-command training w/ unified configs - overhauled vf.RLTrainer for hacking on new algorithms quick demo + links below :)
6
7
75
K2, K2-Thinking, and Edward Nigma 😌
0
0
1
label all you can while the companies are willing to burn for growth 🙌🏼
I am grateful to live in the $2.99 Uber rides era of LLMs
0
2
37
Not liking your tweet and only liking the Sridhar reply with his own better version of a joke on that subject from three years ago
9
42
2K
How it started (2011) How it's going (2025)
0
0
0
Learning from the Kimi details that my Prism harness is effectively [small-model]-heavy with a prompt library for specializing each parallel response
0
0
0
I hope one day tech folks learn how to factor in the negative impacts they have and how that affects the overall value. Never seen a field so thoroughly Goodharted.
i hope that one day creating trillions in economic value for the world will be seen as heroic, as wealth creation is the only thing we know to increase life standards within the nation & globally
0
0
0
Interesting turn of events: With the given plan (consolidated by GPT-5) and the instruction to begin milestone 1 - Codex works for 10 minutes and gives up - Claude Code dives right in, but checks in at random spots in the middle of milestones
0
0
1