turboderp

[turboderp @turboderp_]

@noctus91 I plan to build it out into a more cohesive benchmark with more scenarios, and it probably needs better physics. I suspect some of the lesser models fail because they can't deal with the table having infinite friction.

[turboderp @turboderp_]

soon...
🤖🧠⚡️🚀🏙️🔥🏃‍♀️😱☠️🔚

[turboderp @turboderp_]

@noctus91 It's hastily vibe coded using MuJoCo, since I needed something pretty custom. The interface the agent sees are just three unfamiliar CLI tools; it has to actively request screenshots and figure out how to work the arm without knowing in advance how any of it works.

[turboderp @turboderp_]

@max_paperclips Worse still, all transformers can loop: https://t.co/tT5vzWr1D4

So you see, we were doomed all along anyway. 🙄

[turboderp @turboderp_]

@superalesha Also, you're comparing an INT4 format to QTIP quantization. It's not exactly apples to apples.

[turboderp @turboderp_]

@CosmicMonad I only have so many characters, you know. But check out the dev branch for a preview. It's a huge number of optimizations and fixes (many of them focused on Ampere), no more FA2 dependency... so much. It's wild. Woo.

[turboderp @turboderp_]

I'm having a great time with GPT5.6 so far. I'm still not sure what it's for, but I feel very safe.

[turboderp @turboderp_]

@CosmicMonad It does. Also, big big big update coming soon.

[turboderp @turboderp_]

@SciumoInc I was asking it to add a test automation feature to a simple chat client, which I guess I could see why it might seem dangerous...? It's a stretch, though. But hard to say what really triggered it.

[Eric Hartford @QuixiAI]

https://t.co/oxTnWJZN2v

The *only* reason I have been using Android for the last 8 years is the fact that I have control of my device, I can hack on it and write software for it and sideload it without begging anyone for the privilege. As soon as Google's "Close Android" initiative goes into effect, I will buy an iPhone and I won't develop on Android again.

@Google - Don't do this. Turn around. Cancel. Change your mind. Realize that you just made a stupid mistake after one too many beers, and you didn't actually mean it. And you are sorry for the confusion and chaos. Then this will just be a bad memory that fades away - instead of the permanent destruction you are about to inflict to your own platform, and millions of customers. And the major boon you are about to hand to Apple.

[turboderp @turboderp_]

@Teknium Slow and steady wins the race. But who does agentic inference at temp 0?

[turboderp @turboderp_]

@CosmicMonad @TheAhmadOsman Or, actually maybe I was thinking about 35B (:

Yeah, 27B is dense and might benefit a bit more from TP. I do plan on adding it soon though, just currently doing the Gemma4 side quest.

[turboderp @turboderp_]

@CosmicMonad @TheAhmadOsman The reason is time, really. Priorities, you know. It has 3B active parameters so TP can only do so much given the sync overhead it adds. The model is rarely compute-bound and, being mostly linear attn, that stays true at very long context lengths as well.

[David @dnhkng]

1/n I topped the HuggingFace Open LLM Leaderboard without changing a single weight.
No training. No merging. No gradient descent.
I duplicated 7 middle layers of Qwen2-72B and stitched it back together.
This is the story of LLM Neuroanatomy 🧵

[turboderp @turboderp_]

@Pirat_Nation @BubbaGde No, that's the paper the blog post cites. It sounds like a different method because they make a bunch of new claims not mentioned in the paper (about it being 8x faster, etc.) but they don't cite any other sources.

[turboderp @turboderp_]

@GoogleResearch I wish the paper would elaborate a bit on the "zero overhead" claim. Storage overhead, sure, but what are the specific circumstances where adding a whole extra d×d rotation and codebook quant has no impact on performance, such that you get speedup equal to the bandwidth savings?

[turboderp @turboderp_]

@QuixiAI @sarlev_ @XiaomiMiMo Sadly not supported yet. We're waiting for Tri Dao to review/merge PR #1819 for attn sinks in FA2. One of these days, I'm sure (🙏)

Other than that, 96 GB would still be a tight fit. Could be usable at a little over 2 bpw, though.

[turboderp @turboderp_]

@SchmidhuberAI Neural networks don't beep like they used to. :|

[turboderp @turboderp_]

@elonmusk Open weights would be an improvement. Just sayin'. 🐈