Shreyas

Tech. I represent only my own opinions.

[Shreyas @shreyasOO7]

@cHHillee Isn’t the more general abstraction just dynamic scheduling(batch,chunk) and placement(same GPU,cross gpu,T/E-P,kv transfer)? Why not let the runtime choose it all based on load/TTFT/TPOT/Throughput targets. Is this something people don’t already have one shotted with gpt/claude?

[Shreyas @shreyasOO7]

So they sit there while the more articulate person explains it to them.

A remarkable amount of perceived intelligence is just being better at talking through your thoughts.

[Shreyas @shreyasOO7]

The person who can explain an idea most fluently is very easily mistaken for the person who understands it best.

Someone else in the room may already understand the point, the exceptions, the consequences, and where the discussion is going. None of that is particularly visible.

[Shreyas @shreyasOO7]

@Windsofchange72 You have started a new Yuga

[Shreyas @shreyasOO7]

With seemingly infinite resources we can brute force the brute force and eventually reach the paths of interest. It’s not about how intelligent you are, but how sustainable and enduring your peak is. Everything already exists, can be approximated if it doesn’t, later made real

[Yanco @the_yanco]

Where was Terence Tao when Artists fell?
Where was Terence Tao when Writers fell?
Where was Terence Tao when Programmers fell?

Terence Tao was very pro-AI until recently.

Most even very smart ppl will not pay attention until AI punches them right in their own face.

[Shreyas @shreyasOO7]

@Fintech03 I have always thought if not more, Bharat’s AI landscape will be quantization and making AI extremely efficient. Someone give me this whitepill by the end of this year.

Also I haven’t done any benchmarks on PrismML Ternary bit model and I have my own doubts.

[Shreyas @shreyasOO7]

@UHC_Media Disgraceful & ungrateful people, no good ones anywhere, left, right or center.

[Shreyas @shreyasOO7]

@henrylhtsang Unfortunately, in a broad sense, this becomes part of understanding why models work today. To avoid the effort involved in understanding this (not that it’s easily done), we have to rely on e2e evaluations and a simpler metric like KLD/PPL for output coherence.

[SRA VJTI @SRAVJTI]

What does it take to build the worlds cheapest lithography system

@4rynv @hackerfabindia are building it using a BeagleBone Black paired with a miniature TI DLP board
<$300 in optics.

[Shreyas @shreyasOO7]

@viditchess Did you try the Qwen3.8 27B models quantized by Unsloth?

[Shreyas @shreyasOO7]

@superalesha These are great and interesting results! Did I miss the model link? Very helpful contributions to the community btw, thanks a lot!

[Shreyas @shreyasOO7]

@oneandonlyhusam Is 128MiB RAM low?

[Saurabh Kumar @drummatick]

Computer Science decided to use Kernels for everything

>Kernel (GPU Programming): A parallel-executed function on a GPU device.
>Kernel (OS): core OS component managing hardware and processes.
>Kernel (ML): Similarity measure in transformed feature space.
>Kernel (Computer Vision): small matrix for image convolution filters.
>Kernel (Linear Algebra): Set of vectors mapping to zero.
>Kernel (Functional Analysis): function defining inner products in spaces.
>Kernel (Statistics): Weighting function for density estimation.
>Kernel (NLP): Similarity function for structured data comparison.
>Kernel (Cryptography): A keyspace/function concept in some cryptographic constructions.
>Kernel (Compilers): Minimal/core representation or execution component in some compiler architectures.

[Shreyas @shreyasOO7]

@pupposandro @luceboxai Strix Halo 128 GB + 32 GB GPU 🫣 Interesting

[Shreyas @shreyasOO7]

@svembu @daeroplate_v2 Print money, create jobs, 101

[Shreyas @shreyasOO7]

@rmotafreitas Well atleast it’s useful for some people, what did the rest do?

[Shreyas @shreyasOO7]

@__tensorcore__ Isn’t this only to learn? it’s CuTe

[Shreyas @shreyasOO7]

@_WydAria Matcha too

[Shreyas @shreyasOO7]

@HarveenChadha Can Sarvam not deploy GLM5.2, or even some low param Qwen3.6 seem pretty good, on its unused compute to start a service for cheap? Instead of using Claude/Codex I mean

[Shreyas @shreyasOO7]

@svembu @ImperiumHindu People wanted accountability for the paper leak and irregularities. If we’re going to talk about fairness in education, then we should be willing to discuss all forms of unfair advantage, including reservation,rather than treating some topics as untouchable (no pun intended).

[Shreyas @shreyasOO7]

@svembu @ImperiumHindu Resignation of Shri Dharmendra Pradhan achieved nothing and nobody asked for the students that cheated and nobody demanded against reservation which is a cheating too 😂 There’s no hope, the people are simply extraordinarily shitty 😞