@Machinedelusion 1.15% WER after GRPO? That's insanely low, like, big lab quality low, and at only 100k hours? That's insane.
I've been looking for something specifically like this all week. Will definitely be trying it out, thanks for the release!
And this isn't a glaze post for OpenAI either, they've had a lot of blunders along the way.
But something is weird with all of this, it's a complete break in the pattern. Given all the NVIDIA shit going on, + Anthropic? Something REALLLL fishy is going on here... 🧵3/3
@_pedrohro@kusurichan_ Only 20 tok/s? Jeeze. I was hoping for more out of a small model. Gonna have to test it out myself and see if I can get more out of it.
@JoeGuglielmucci@jc50000000 I’m not saying it to shit on you, but you should really dig into the actual eval and the “training data” used for the model before burning more of your time.
It might really open your eyes to how incredibly inaccurate their “report” was. I used to like Nous. What a shame.
@HeLi8866 11.9GB if you use Q2_K which is essentially braindead. Especially for a 31B model. And given it's stats vs Qwen 3.8 27B, I wouldn't call it "top tier" lmao
@jeffboudier I was literally up all night thinking about MoEfication after my last CMoE project failed. Thank you guys so fucking much, this is literally a gift, and I think a sign.
Time to get back in the fucking seat and start tinkering with models again. Thank you so so so very much.
@MiaAI_lab Too bad 75% of pruning is incredibly bad. @cerebras published papers on how pruning experts holds up. I've used it on many projects. Up to 50% is okay depending on the calibration data.
This is uncalibrated pruning with high sparsity. I'd be surprised if it did anything useful.
@0xFitz The route just needs to be re-optimized. It has unnecessary pathing, and doesn't group the beginning quests together as well as it could.
It kept pushing routes instead of telling me of the trainer skills I should get (BoM, Judgement) which are crucial for early pally leveling.
Retweet to get your friends & family aware, protected & an invite to use Underdog
Safety is the priority. Free, safe, & smart AI for everyone https://t.co/HXqMWoya9U
Underdog is the only personal AI that is capable and SAFE
It runs local, not in a Linux VPS, not at the mercy of any cloud attacks. Your data stays on your own device
We hire the world's best hackers to keep Underdog safe: the team who hacked OpenAI, 3-time world #1 competitive hacking team, auditors that secured $36B+ & patched $1B+ in vulnerabilities
read more: https://t.co/nysp6zx8AZ
@cleeeeeeeeement@OpenBot_ Looks fantastic! Quick question on the claude plan option...
Since anthropic now only allows subscription sign-in through the unmodified claude code binary under their commercial terms, can you confirm that’s how rerun handles it?
Sorry for being a pain in the ass.
@volkdude85@sudoingX Oh right I don't disagree with you about the quality of the models, you're right on that by far.
I just meant for those who want to learn the actual basics of how LLMs work and want to tinker with them.
@volkdude85@sudoingX It's like buying your first car. Most people don't go out and buy luxury cars to learn to drive in. Same goes for local AI. You don't need the best. Just what works for you.
That being said, I'm not disagreeing with you that it's MADLY expensive right now. But there are ways in.