I'm writing an IDE specifically for LLM-aided code hyper-optimization. I'm going to write up a blog post about it and open source soon. Here's a screenshot of what I've got so far.

Terrible license. I didn't expect such onerous terms for a 300M param model, but here we are. Also, curious wording. They do not compare to Qwen3-Embedding-0.6B, because they do not beat it. With that said it's probably a good model.

Join millions who have switched to Grok.

And for good hardware accelerated low precision kernels you should really look at:.

The relevant scaling laws papers:.

I've become increasingly convinced that if you aren't training in FP8 or smaller you're a chump. Especially if you're trying to do RL, but also in any setting where you're compute constrained and not data constrained. It is just usually better to do quantized training. There's.

The number of LLM tools that seize up when they encounter <|endoftext|> in a string literal astonishes me. I thought everyone knew to be careful about this.

They are starting to become self aware.

I'm implementing a tokenizer for a project, and it astonishes me that tiktoken is generally considered "fast". Agony.

The SF health cult has convinced me, I'm getting into supplements. Starting out with the basics, plus a multivitamin. The D,L-Phenylalanine is a personal headcannon. But how does anybody swallow these things? They're huge. Can't split them, they're full of foul tasting liquid.

I'm 'boutta crash out, but I'mma crash into bed and go eep instead. nini ❤️.

The two techniques this paper introduces, JetBlock and PostNAS, are LITERALLY NOT EVEN DEFINED. There's a bunch of red flags in this paper. First, Nvidia did not release the code. The paper does not contain much information, certainly not enough to be reproducible. It's mostly.

What confuses me is that if you look at the scaling curves, it's clearly better. If you're not doing MoE + quantized training you're a schmuck, especially if you're doing RL, which you should be doing for human preference posttraining anyway, even if you don't believe in.

God this must be so embarrassing.

Please for the love of all that is holy tell me they were already doing this.

Jinja has import statements, I'm 'boutta crash out. I'm not sure why everyone is standardizing on it for tool call prompting.

Tonight I was at a hypnosis lecture, and someone asked: "When you imagine yourself on a beach, how do you experience it? Do you see it? Hear it? Feel it?". The thing is, I don't. None of the above. Or I imagine in third person. I think somewhere along the way my brain got fried.

Same form factor as V3/R1, hoping they revisited the pretraining data with the intent of making it better to do RL on. The more stuff that approximates logical reasoning in the pretraining data, the better.

I am faced with the sobering reality that writing an efficient mxfp4 kernel for gpt-oss is not possible in llama.cpp because of the memory layout. Blocks of quantized elements are not stored contiguously, so you cannot issue vector loads across mfxp4 blocks. Sadge.