#Decoding
From “I am become meme”:

Talking about the ballroom is colonisation of political discourse by Disinfolklore. It does not need us to want a ballroom. It needs us to talk about it.

Podcast: open.spotify.com/show/033ZXAMDMZot3LG…
September 28, 2026 at 12:15 PM
Have you tried qwen3.8 flash next on dwarfstar?

Last time I checked it didn't implement qwen4 path on ROCm, but with a little work of the agent I managed to make it work and at 200tps on prefill and 20tps on decoding.

Its also very stable throughout all 256k context.
September 28, 2026 at 11:15 AM
𝗗𝗼𝗴 𝗕𝗼𝗼𝗸𝘀 𝗜 𝗪𝗶𝘀𝗵 𝗘𝘃𝗲𝗿𝘆 𝗗𝗼𝗴 𝗢𝘄𝗻𝗲𝗿 𝗪𝗼𝘂𝗹𝗱 𝗥𝗲𝗮𝗱
Ever wonder what your dog is really thinking? These books are your inside scoop — from decoding behaviors to heartwarming rescue tales, and everything in between. Learn, laugh, and deepen your bond with your canine companion. Definitely save this ⬇️ #dogs
September 28, 2026 at 11:00 AM
I was skeptical, but they really do seem to have fixed The Problem with Opus - agent sessions no longer feel like wrestling with a mad, inebriated savant or decoding Delphic prophecy
September 28, 2026 at 10:40 AM
💚 What lies at the heart of CIRCUIT?

Discover the vision, mission and blueprint driving our journey!
Read more: circuit-horizon.eu/decoding-cir...
#CIRCUITProject #HorizonEU #CircularEconomy
September 28, 2026 at 10:31 AM
Zhiqi Ai, Han Cheng, Shiyi Mu, Yongjin Zhou, Shugong Xu
PTC-Bias: Phoneme-Level Temporal Competition for Bias Retrieval and Post-Decoding Correction in Speech LLMs
https://arxiv.org/abs/2609.28727
September 28, 2026 at 10:05 AM
How does an autoencoder work? 🧐

Autoencoders belong to the semi-supervised methods because you train them only with the normal state of the data. 

The network consists of two sections:

an encoder function z = g(x) and

a decoder function x′ = f (z)
September 28, 2026 at 10:04 AM
Hexana for VS Code 0.12.0: Apache Parquet and Arrow IPC support, tar and tar.gz, schema-less Protobuf decoding, Arrow IPC streams, open-by-content for extensionless binaries, and a Reopen Editor With entry for any file. #VSCode #ApacheArrow

[link card: marketplace.visualstudio.com/items?itemNa...
September 28, 2026 at 9:52 AM
> USN-8823-1: PyJWT vulnerabilities
https://ubuntu.com/security/notices/USN-8823-1
USN-8823-1: PyJWT vulnerabilities
It was discovered that PyJWT incorrectly handled certain URIs when using PyJWKClient. A remote attacker could possibly use this issue to perform server-side request forgery (SSRF) or expose sensitive local files. This issue only affected Ubuntu 22.04 LTS, Ubuntu 24.04 LTS and Ubuntu 26.04 LTS. (CVE-2026-48522) It was discovered that PyJWT incorrectly verified cryptographic signatures when decoding tokens with PyJWK keys. A remote attacker could possibly use this issue to bypass signature verification and forge valid JSON Web Tokens. This issue only affected Ubuntu 26.04 LTS. (CVE-2026-48523) It was discovered that PyJWT incorrectly handled unknown key identifiers in PyJWKClient. A remote attacker could possibly use this issue to cause PyJWT to make excessive network requests, resulting in a denial of service. This issue only affected Ubuntu 24.04 LTS and Ubuntu 26.04 LTS. (CVE-2026-48524) It was discovered that PyJWT incorrectly handled payload decoding during detached JWS token verification. A remote attacker could possibly use this issue to cause PyJWT to consume excessive resources, resulting in a denial of service. This issue only affected Ubuntu 24.04 LTS and Ubuntu 26.04 LTS. (CVE-2026-48525) It was discovered that PyJWT incorrectly validated JSON Web Keys when decoding tokens with both HMAC and asymmetric algorithms enabled. A remote attacker could possibly use this issue to forge valid tokens, resulting in an authentication bypass. (CVE-2026-48526)
ubuntu.com
September 28, 2026 at 9:40 AM
Learn how tactical communications interception software helps close VHF-UHF-SHF gaps with faster digital signal decoding, real-time signals analysis, and EW support for mission teams

comintconsulting.com/krypto1000
September 28, 2026 at 9:22 AM
The new White House ballroom is ninety thousand square feet. Building it doubles the footprint of the White House.

🎙 Decoding Trolls podcast: decodingtrolls.net/p/podcast-i-am-bec…
September 28, 2026 at 9:15 AM
Decoding toxicology and forensic procedures in modern death investigations. Behind public headlines lies a rigorous, vital scientific process bringing clarity to families.

#ForensicScience #Toxicology #Zovintus
[Image by AI]
Understanding Toxicology Reports and Forensic Procedures in Modern Death Investigations
Your source for world news, technology insights, crypto market updates, science, and health. Explore deep investigative reporting on Zovintus Blog
blog.zovintus.com
September 28, 2026 at 9:14 AM
Yusung Choi
ImCorr: Sub-pixel Semantic Correspondence via Implicit Feature Decoding
https://arxiv.org/abs/2609.29193
September 28, 2026 at 8:37 AM
Gen Z texts. Gen X calls. And everyone has a different idea of what “professional” means.

Here’s your guide to decoding the multigenerational workplace:
theconversation.com/gen-z-commun...
September 28, 2026 at 8:14 AM
Liquid AI, VLM 디코딩 속도를 2.66~3.13배 가량 높인 LFM2.5-VL-3B-DSpark 모델 공개

LFM2.5-VL-3B-DSpark 소개

LFM2.5-VL-3B-DSpark는 Liquid AI가 자사의 비전-언어 모델(VLM, Vision-Language Model) LFM2.5-VL-3B를 위해 공개한 약 280M 파라미터 규모의 추측 디코딩(Speculative Decoding) 용 초안 모델(Draft Model)로, 출력은 그대로 유지하면서 디코딩 속도를 H100에서 최대 2.66배, Apple M…
Liquid AI, VLM 디코딩 속도를 2.66~3.13배 가량 높인 LFM2.5-VL-3B-DSpark 모델 공개
LFM2.5-VL-3B-DSpark 소개 LFM2.5-VL-3B-DSpark는 Liquid AI가 자사의 비전-언어 모델(VLM, Vision-Language Model) LFM2.5-VL-3B를 위해 공개한 약 280M 파라미터 규모의 추측 디코딩(Speculative Decoding) 용 초안 모델(Draft Model)로, 출력은 그대로 유지하면서 디코딩 속도를 H100에서 최대 2.66배, Apple M5 Max에서 최대 3.13배까지 높입니다. 본 모델 대비 파라미터가 8.9% 늘어나는 비용으로 얻는 속도 향상이며, llama.cpp, MLX-VLM (:py...
discuss.pytorch.kr
September 28, 2026 at 7:31 AM
Although worth considering the case he will carry on for years from Michael Kimmage youtu.be/3DrpVfVkZyU?...
Russia Expert: The Real Reason Why Putin Needs the War - And Why It Won't Stop in Ukraine
YouTube video by Decoding Geopolitics Podcast with Dominik Presl
youtu.be
September 28, 2026 at 7:18 AM
Aoke Zhang, Jing Chen: Subject-Invariant Cross-Modal Decoding of Perceived Speech from Brain Recordings https://arxiv.org/abs/2609.30832 https://arxiv.org/pdf/2609.30832 https://arxiv.org/html/2609.30832
September 28, 2026 at 6:44 AM
Vivien Tran-Thien, Richard Nock: Mentored Decoding: Faster Inference meets Boosting https://arxiv.org/abs/2609.30474 https://arxiv.org/pdf/2609.30474 https://arxiv.org/html/2609.30474
September 28, 2026 at 6:42 AM
When Elon Musk told a conservative conference in 2025 "I am become meme", he was not joking. He was telling you what he does for a living. Podcast: podcasts.apple.com/gb/podcast/podcast…
September 28, 2026 at 6:15 AM
Researchers introduced AIMES, a framework for adaptive multi-value activation steering in large language models. The method adapts intervention strengths at every decoding step based on online observers.

Source: arXiv cs.LG
Adaptive Multi-Value Control in LLMs via Causal Activation Steering
Abstract page for arXiv paper 2609.30405: Adaptive Multi-Value Control in LLMs via Causal Activation Steering
arxiv.org
September 28, 2026 at 4:50 AM
I would guess it never resets the state to allow some kind of jank streaming where you fudge the addresses in the middle of decoding? If you have some setup that switches memory banks in and out maybe
September 28, 2026 at 4:08 AM