Rahul Atlury

@Prince_Canuma They say int4 per-group affine quantization is more robust across architectures, while angular methods like PolarQuant can fail badly when attention scaling is non-standard. It reports PolarQuant works on Llama 70B but fails on Gemma 31B...FYI

✅

Coming to MLX 🚀

@Prince_Canuma @Prince_Canuma no intention of hijacking this thread but have you looked at Open-TQ-Metal arxiv.org/abs/2604.16957 They apparently compute attention directly from compressed cache.

@jimkxa @tenstorrent Please do consider releasing them standalone, achieving scale via massive, centralized data centers is just unsustainable for nature. Local AI allows for distributed, sustainable growth, thats the scale moat companies are missing...

@atlury @tenstorrent We need it the P300 cards in the QuietBox2, rethinking releasing stand alone

Don’t worry, we’ll make this fast. About 3 weeks. Lots in the pipeline. @tenstorrent

@DeepComputingio @jimkxa @tenstorrent This memory crunch is killing local AI from taking off, what we have been doing is going to back writing specialized simd kernels in assembly, testing in old generation DDR rams. Aah painful but needed at the place I live...waiting for better

@jimkxa @atlury @tenstorrent Soon RISCV HOST QB2

@brk0v I think the baggage of having to maintain compatibility, especially with the weird use cases is the biggest inertia for lightweight Linux systems to take off...more modernization is needed...

🦀Rust: Rewriting GNU coreutils in Ubuntu. FOSDEM talk: Memory safety matters, but Rust is not magic. The hard part is compatibility, weird Unix edge cases, test suites, distro integration, rollback paths, and users finding real bugs. fosdem.org/2026/schedule/… #rust #rustlang

@Prince_Canuma In terms of accuracy how close was it? usable?

You can now run DeepSeek4-Flash on 256GB Mac. Next up speed 🚀 PR: github.com/ml-explore/mlx…

@jimkxa @austinlyons @tenstorrent Love the work Tenstorrent is doing, I wish they existed in M.2 form factor as well for edge devices, and I wish we had access to your hardware, especially in this era of unbelievable level of vibe coding, we could just translate the whole hugging repos to TT. :-)

@austinlyons Jensen said Groq runs 25% of agents AI. I doubt specialized hardware is going to win. AI is too diverse and changing too fast. @Tenstorrent runs 100%. Hardcore benchmarks coming.

Is Groq the only inference chip for agentic AI? Or just the first of many? Same for Vera and CPUs.

@steeve @sluongng This is very interesting, i recently bootstrapped a non-systemd, openrc, non gcc, glibc minimal os from scratch using nothing but zig with zig based package manager i wrote, the traditional LFS stages collapsed to just 3, zig is wonderful, libgcc shims n all...

Coming from @sluongng, it means a whole lot

@anil_nal @AAI_Official @RamMNK Sad!

An immigration officer slapped a passenger (Eerlam Anjaiah) who was arguing after being sent to the end of the Immigration clearance line at 8 pm on 24 Nov at Hyderabad Airport. Written complaint lodged to check CCTV cameras. Please take action @AAI_Official @RamMNK

@Prince_Canuma Indeed, sad to see this behavior.

I feel sorry for people like him! I lived in Gujarat for 5 half a decade, and I loved it because of how amazing, kind and respectful people are. Don’t know what’s happening in his life for him to speak to others this way.

@m7server your support@monibuca.com is bouncing. Are you still working on commercial version?

New Interactive special effects

@Prince_Canuma aaah nice!

@atlury sure I will take you up on it too Btw, I used to live in India till 2022 🇮🇳

Work so hard the community wants to feed you😎🙌🏽

@skalskip92 How did you convert to a vector image?

okey... I jumped on the bandwagon

@sarahcstock this is very painful but correct picture. Its the same within India as well, I feel the same within India itself. Everyone within here just wants to build acres and acres of concrete jungle. They have just destroyed nature. I hope you guys rebuild and survive.

If next generation Intel GPUs (Battlemage) can ship in 24gb and upwards then its game on. Can easily combine 4 or 6 of them and run 70b models. :-D

@JustinWolfers excellent talk! amazing....

Here's a overview talk I recently gave on how AI will change the world of work. My goal in this talk was to put together the existing literature, hopefully in a way that makes the whole more interesting than the sum of the parts: youtube.com/watch?v=jIL5pV…

@Prince_Canuma Congrats! We are getting there….

Florence-2 is next level good! 🔥 On-device OCR will never be the same. Btw, I will be speaking about MLX at Data Science Summit in a couple weeks. Date: 22 of November 2024 Location: PGE Narodowy Stadium, Warsaw Poland 🇵🇱 20% Discount Code: DSS24SP20 See you there!

@ElecNotes This is excellent vintage stuff, please give references, it will help us explore further. And thank you!!!

Valve / Vacuum Tube Reflex Amplifier One of the techniques widely used in early radios was that of what was termed a reflex amplifier. The curious typical of one used for this technique. The IF signal is applied to the control grid. Then the amplified signal is passed to the diode rectifier elements of the valve from the IF transformer T2. On passing through the filter formed from C1, R1, and C2, the AF signal appears across the volume control R2. It is then passed back through the secondary winding of T1 which has no effect at audio and into the control grid again. The audio is amplified by the valve and this time appears across the anode load resistor R5, after which it is passed onto to further audio stages.