lsm_

@StefanoErmon I love diffusion models. AR models getting 500 TPS on 8xB300 is sota while diffusion models are casually crushing 1000 TPS.

New real time experiences powered by Mercury 2, our flagship diffusion LLM

@kwangmoo_yi Wish there was a blog somewhere guiding through paper writing and submission processes

For example, allows me to point out critical experimental flaws that could be a sole reason to reject the paper despite the paper having a strong accept rating. I can confidently point this out NOW, to be rebutted by the authors in case I am wrong.

I'm gonna regret saying this, but I like this year's NeurIPS decision to make me write initial metareviews. Gives me an early enough assessment and feedback to **both** authors and reviewers. (1/2)

@jasondeanlee Yes

Was there just a usage reset for codex?

@clankrmedia @grok how long does the battery last?

Researchers built a soft floating robot for indoor interaction. It uses helium and flapping fins instead of propellers. The result is quiet, lightweight, and safe to touch. It can follow people, give reminders, and act as a study buddy. Published at ACM DIS 2026.

@praneetdutta Man I just read hellopomo's "m" as "rn" (RN) and was very confused

Something's cooking, stay tuned.

@durov You guys need to get in VR

Today’s Telegram update brings Rich Messages, Communities, Ephemeral Messages, and private GIF search. telegram.org/blog/communiti…

@thomasfbloom @jdlichtman And I thought this level of maths would be the last to get automated

@jdlichtman Update - have now read this proof and yes it is correct (and formalised!), so the site has been updated.

Beautiful result on 2-primitive sets! GPT-5.6 just solved another 50+ year old problem (Erdős #793)

@matthewjgunton 😅 "when op is mem bound try to use less mem".. any other insight I missed ?

Inferencing with lower precision datatypes moves less data across your bandwidth and gives you access to low-precision-specific hardware For example, an H100 SXM has approximately 3,350 GB/s of HBM bandwidth Imagine that each decoding step requires the GPU to read 5 GB of model weights and other data from HBM. At the theoretical maximum bandwidth, that movement has a lower bound of: 5 GB / 3,350 GB/s = ~1.5 ms FP32 uses four bytes per value. FP16 uses two This drops the amount of data to move through memory to 2.5 GB, resulting in a floor of: 2.5 GB / 3,350 GB/s = ~0.75 ms Then when you are doing your matmuls and other calculations during the forward pass you can sometimes access parts of the hardware only setup for certain precisions (eg FP8 tensor core instructions in H100s) I need to call out: this does not mean inference will always speed up by exactly 2×. Inference kernels can accumulate in higher precision, unpack quantized values, apply scaling factors, or simply upcast data unexpectedly But the rule of thumb remains useful: when decode is memory-bound, reducing the number of bytes moved per token directly improves performance

@j4orz not saying it will happen 100% but there's a possibility (maybe Yann LeCun's team or SSI discovers something). Remember seeing posts on twitter early on, saying we won't cross the 4k-10k context length anytime soon.. which we did. so just a hunch.

@thisispiyushK what makes you think this is going to change quickly?

i've been hammering on both fable and sol for sitp's intermezzo to formalize (and better understand) the tight vs leaky distinction, and what exactly are the measurable spaces language models are defined on from cotterell's (eth zurich) course notes. these are a great complement to jurafsky, eisenstein, and manning. fable and sol seems to continue the (slight) divergence of opus and chat5.5's rl stacks? x.com/willdepue/stat… i'd really like something in between the two post-trainings. because at times fable is just way too much and wants to explain the world to me. this is amazing alien-like intelligence (i LOVE the compression) but it feels like a rocketship for the mind rather than a motorcycle. on the other hand sol's (whether intentional or not) simpleness is what is needed. using opus before fable is also a nice combo. and beyond the post-training, there are so many times where i wish the model could tell me, "based off your questioning you are completely conceptualizing/attacking/inquiring about this the wrong way, this is how you should go about it. definition 1.1, definition 1.2, lemma 1.3, theorem 1.4, etc." a lot of times i feel like these models don't have a good theory of mind (world model?) for humans (i.e thomas nagel what it's like to be a bat), and entropy collapse is more likely when learning with claude/chat rather than coding with cc/codex. it is really up to the learner to reset the line of questioning if confusion starts arising. talking to fable is starting to feel more and more like browsing the collective intelligence of stackoverflow or wikipedia. this is akin to some of the stuff that @3blue1brown and @dwarkesh_sp were riffing about recently. also for the lesswrong asi pilled folks that believe alien-like compression which is out of sync with humans in the space of minds doesn't really matter given that humans will be soon out of the loop, my priors and the way i've been updating tell me the opposite. i expect this century is going to look much more like jevons paradox, ai-human hybrids, rather than some ai 2027 (and most recently ai 2040) ¯\_(ツ)_/¯

@kohya_tech good stuff

bmaltais氏のmusubi-tuner-guiがひさびさに更新されてKrea 2にも対応した。 github.com/bmaltais/musub…

@sgl_project been going through sglang lately, adding this to the list

In two weeks since the model launch, we pushed GLM-5.2 NVFP4 on SGLang to 500+ tok/s/user at bs=1 on 8xB300. Thanks to our new TopK-V2 kernel, interactivity stays essentially flat from 80K all the way out to 1M-token context. We attacked it on two fronts: cutting overhead (zero-bubble scheduling, sync removal, kernel fusion) and rebuilding the hot kernels themselves (TopK-V2, CuTe DSL GEMM). Check the blog for the full technical details and repro commands 👇 lmsys.org/blog/2026-07-1…

@fujikanaeda Hit the iron when it's hot

sam smells blood

@vikhyatk What's your plan after that?

waiting for ai to automate software engineering

@AlpinDale @ShitpostRock Never got its appeal

@ShitpostRock Astro Bot

Okay but like geniuenly what the fuck is there to play on a PS5 emulator outside of a downgraded version of Demon Souls?

@KyeGomezB That's how you get prime intellect and personal universes

"a quadrillion parameter model" 😩

@matei_zaharia @jefrankle Seems like a magical place

@jefrankle Sweater in the summer is also required in SF.

Catching up on the summer fashion trends.

@norpadon Yeah looks vibecoded tbh (no disrespect to the author)

@thisispiyushK This code shouldn’t exist really

Vibecoding monkey bastards made GitHub totally fucking unusable...

@norpadon No idea what is this code about but somehow I keep thinking the formula should be 0.1 * math.log(max(1, self.scaling_factor)) + 1.0

This is not a joke btw, I am literally sending people code screenshots

@SamJWasserman @u1tra_instinct I just need 4 more 🥲

@u1tra_instinct doing incredible work. Got GLM 5.2 at 30tok/sec 128 context running on 4x Dgx Sparks - Fully Uncensored. Wild. Now I just need 2x more Dgx Sparks? Accepting donations! 🤣