2026-07-11

TL;DR: Quantization has a marked impact on agentic performance but little effect on knowledge. I manage a small HPC cluster at a university, and we have recently begun running common benchmarks to help our users understand the effects of quantization. We have just completed the runs on the Qwen 3.6 quantizations and posted the results on our website: The results are consistent with what most people would expect: knowledge, as measured by GPQA Diamond, varies very little across quantizations. GPQ

tl;dr pi-coding-agent (bash for everything/minimum tools) is up to 2x cheaper and even has higher pass rate according to their own benchmarks across the board. GLM 5.2 is above GPT 5.5 high and xhigh, on par with Opus 4.8 high. This is yet another "in our use case"-type benchmark but it comes from DBRX who actually trained a sizeable LLM in the past, and I think they know what they do. I think their analysis makes sense, and GLM 5.2 genuinely do feel on par with Opus 4.6/4.8 for most coding task

I've been running some systematic tests on a few models comparing FP16 vs various GGUF quant levels, and instead of looking at one aggregate benchmark score, I broke it down by capability: math (GSM8K), code (HumanEval), reasoning (ARC-Challenge), and knowledge recall (MMLU-Pro). The results are way more nuanced than "Q4 loses X% quality." For example on one 27B model, Q4_K_M barely moved the needle on conversational/knowledge tasks (under 2% degradation) but dropped multi step math accuracy by

Apple has sued OpenAI, alleging that engineers stole Apple secrets to advance the AI startup's hardware plans. In its complaint, Apple says it uncovered "a pattern of theft of Apple's trade secrets by OpenAI employees who were formerly at Apple." In addition to OpenAI, the lawsuit also names IO Products - Jony Ive's hardware startup, […]

Apple alleges the misconduct was directed by OpenAi's senior leadership, including a long-time former employee.

How Deutsche Telekom is becoming an AI-native telco with OpenAI-transforming customer service, employee workflows, network operations, and the future of voice.

Hi everyone, A year ago I began pre-training language models exclusively on 1800’s London data. Recently I have completed my largest dataset ever, containing 40B tokens or 160GB of 1800-1875 english data from England and the United States. I will soon train a 2B parameter model on it, but for now I’ve trained a 500M parameter evaluation model on a 5B token sample. I have also fine tuned the eval model on 1800’s Q&A pairs (using synthetic questions and answers pulled straight from the dataset), s

Hello, I'm continuously working on OpenFox (MIT-licensed - no business model whatsoever), which is a harness dedicated to local AI, mostly for coding but well you know, this can do anything. I'm using it every day with my 2x Spark cluster, mostly with DS4 Flash these days. I noticed a small opportunity for improvement, nothing revolutionary but it kinda clicked at some point. When you create a new session and start typing your prompt, there is this time where your local rig does nothing. Then yo

Here's what I'm thinking about: A USB thumb drive that you can plug into any PC or laptop, and immediately get a usable knowledge base powered by an LLM, without requiring an Internet connection. I believe the technology for this should be ready. Rough architecture: llama.cpp binaries for CPU-only inference, for Windows, macOS, and Linux (each for all major architectures, a few hundred MB total) Qwen3.5 35B-A3B @ Q4_K_M (22 GB, for systems with >= 32 GB RAM) Gemma 4 E4B @ Q4_K_M (5 GB, for syste

Been testing out the ONNX models to see how far I can push the CPU to take on ASR and TTS, so the GPU is completely free for running the LLM. The video attached shows me testing latency on a 2022 Macbook M2 and an AMD Ryzen 9 7900. This is just running the regex fast commands, so most of the latency (apart from grabbing the spotify music) should be from the ASR and TTS. These ONNX models are really great. The M2 is mostly usable, the Ryzen 9 is blazing fast. The two models I am running are: Daum

Author here, sharing something I built. If you run agents on a local model, feeding a whole page as raw HTML burns your context fast. barebrowse turns a URL into a pruned ARIA snapshot — the semantic tree with nav/ads/boilerplate stripped — so each page is a fraction of the tokens. Useful when the context window is your bottleneck. No Playwright / bundled Chromium — it drives a browser you already have (Chrome/Brave/Edge/Chromium) directly over CDP. Pruned ARIA snapshot instead of raw HTML/DOM —

HiLS-Attention is a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling loss, enabling native sparse training for efficient long-context modeling. This repository hosts the 7B checkpoint continued-trained on top of an OLMo3-style backbone. Model introduced in the paper Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling . Model Description Naive block sparse attention selects top-k chunks by their exact chunk mass, but c

The AI chip boom just produced its biggest Wall Street moment yet. Now SK Hynix and Samsung are being asked to build U.S. factories.

arXiv:2607.06760v1 Announce Type: new Abstract: Autonomous systems under partial observability act on beliefs, not raw sensor events. QANTIS treats the quantum processor as a calibrated belief-update service in that loop: it receives a prior and an observation model, estimates the rare-event evidence term, and returns an ordinary posterior to a classical planner. This paper asks whether that service can be reused across a sequential Tiger POMDP horizon on present IBM Heron hardware without corru

More Fable, GPT-5.6 and new ChatGPT voice

Devthropology is a passion project built on top of GitHub pull data. The name is a play on developer anthropology. Pull request data can be cut a lot of ways. The functionality has been built out of curiosity as I want to see different insights into codebases that I work on. Some of the data is typical and other parts I haven't seen elsewhere. I think of this as an improved GitHub Insights page, with faster performance, more detail, and a focus on how work moves through a codebase. The main enti

Hello, Locallama! "long-time" member here, from the days when the max context window doubled from 2K to a whooping 4K! Now I feel like I am living in a dream world, doing things I could not have imagined back then. This is Belochka, or Bella, running on a Mac and projected to an ESP 32 w-10 board, the "Rune Board" because runes :-). In the background is Qwen 3.6 35B 3AB q4_k_m gguf on LM Studio. She kinda mentioned her capabilities, but pretty much she is like a personal assistant on a Mac. She

Disclosure: No AI was used to write this My specs are: RTX 4090 128 GB DDR5 5600 MT/s Intel Core Ultra 7 270k Running nvidia-595 on ubuntu 26.04 with latest llama.cpp build (pulled and rebuilt this morning). Tried a lot of things, ended up running unsloth's UD-Q2_K_XL quant with command: taskset -c 0-7 /home/kevin/ai/llama.cpp/build/bin/llama-server -lv 4 -m /home/kevin/ai/models/DeepSeek-V4-Flash-UD-Q2_K_XL-00001-of-00003.gguf --temp 1.0 --top-p 1.0 --min-p 0.0 -t 8 -fitc 64000 -fa off -np 1 Sp

As part of our ongoing efforts to strengthen our safeguards for advanced AI capabilities in biology, we’re evolving our Bio Bug Bounty into an ongoing private program, known as the OpenAI Bio Bug Bounty program and doubling rewards to $50K. We’re inviting researchers with experience in AI red teaming, security, or biosecurity to try to find a universal jailbreak that can defeat our predefined biosafety challenge against OpenAI’s frontier models.

As the AI boom boosts demand for RAM, SK Hynix - one of the world's biggest suppliers of memory chips - launched on Wall Street Friday. The South Korean chipmaker opened at $170 per share and raised $26.5 billion, surpassing Alibaba's record as the largest debut of a foreign company, according to reports from The […]

Digital Services Act may force Meta to make big changes on its platforms.

The massive round was co-led by ARCH Venture Partners, Spark Capital, and Khosla Ventures.

SensorFM, a wearable health foundation model from Google Research, Google DeepMind, and university collaborators. We walk through its ViT-1D masked-autoencoder backbone, pretrained on more than one trillion minutes of unlabeled sensor signals from 5,000,000 consented participants. We examine the co-scaling results across four model sizes and four data volumes, including the case where capacity outruns data. We show how frozen embeddings plus a PCA-50 linear probe beat feature-engineered baseline

arXiv:2607.06925v1 Announce Type: new Abstract: Compact world models that condition on a language goal promise to ground relations such as ``put the red block left of the blue block'' using a sparse set of explicit \emph{reference anchors}. We ask when such references actually ground a relation, and identify a trap: a goal-conditioned predictor reaches a striking $0.90$ relation-readout accuracy, yet this is \emph{instruction transcription}, not perception. Withholding the goal collapses it to c

A few months ago, I introduced you my project Homelable, a design tool intended to make it easier to import and document our self-hosted infrastructure via diagram. And theses few month was damn busy. The response was honestly way bigger than I expected (we're past 2k stars now, still feels a bit unreal), and most of what's in this release came straight out of your issues and PRs. So, thanks for that. So, after 2.2k stars, +100 issues (mostly opened by you) and 33 releases : Homelable v3.0.0 is