Search: “chat”

TL;DR - yeah it works, no wild skills or patching required. 37-50 tg/s when generating text, spiking over 60tg/s when writing code and mtp hits hard. Your results can be better if you have a better mo

Banger paper from Stanford on efficient test-time scaling. If you run agents that think for a long time, this one is worth your time. (bookmark it) Long reasoning keeps the entire trace in memory thro

Interesting paper on prompt optimization. They claim that a single-lineage prompt optimizer just matched GEPA on a smaller rollout budget. Prompt optimization has been drifting toward heavier machiner

Interesting technical work from Microsoft. Provides a better understanding on SFT and how to leverage it better for RL. Microsoft researchers asked whether a standard SFT pipeline actually produces th

I was bored and decided to test some of the image generation. You can see from the images that I got *very* consistent patterns. These were only a few, but I was approaching 20 generations before I st

// Your LLM judge disagrees with the experts // LLM Judges can be tricky to build. Here is an interesting showcasing why: There propose a reference-full benchmark of hundreds of complete human-to-huma

Banger paper from Apple. If you build MCP servers, this can help you turn your specification into an evaluation suite. (bookmark it) It's actually a really neat idea that's easy to implement. And it s

Finally, a paper testing whether hiding your agent skill files actually protects them. The short answer is no. That's concerning. Worth reading if you sell access to a skill or share one across teams.

Introducing Hy4 Preview New open weight text input (no vision) LLM from Chinese company Tencent today: 770B total parameters, 49B active parameters, 1M token context window, 1.56TB on Hugging Face. Th

// Memory is what breaks long-horizon agents // It's undeniable how important memory/recall is for long-horizon tasks. If you are building for long-horizon tasks, this is a great read. (bookmark it) T

Below are the benchmark results for running Qwen3.8-Flash-Next on Strix Halo using the Vulkan backend of llama.cpp, combined with MTP model. Hardware Item Details CPU AMD Ryzen AI MAX+ 395 (16C/32T) G

Hey, An LLM writing into durable memory is an unreliable author. It invents, mislabels, restates, and contradicts. Once one of those mistakes is saved, the agent can carry it into every conversation t

I got early access to SpaceXAI’s new Grok @Bot template sharing feature, which lets you create custom Bot templates that other people can download and use. I created a Home Robots template that lets y

arXiv:2608.26650v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) models route each token to a subset of expert networks, increasing capacity while keeping per-token computation sparse. In many

Things getting wild in automated systems engineering. This new work claims to build an AI system that designed and deployed a frontier accelerator. Two human architects wrote the hardware specificatio

arXiv:2608.25022v1 Announce Type: new Abstract: As people adopt transformer-based language models (e.g., ChatGPT and Gemini) for an increasing number of use-cases, it is important to know how such mod

// Automata from agent traces // How much of your agent's behavior comes from the model, and how much from the harness you wrapped around it? New work collapses an entire corpus of agent traces into a

model: add Qwen3.8-Flash-Next (qwen4exp) (#27742) gguf: add qwen4exp (Qwen3.8-Flash-Next) arch and converter Adds the GGUF-side plumbing for HF model_type qwen4_exp: MODEL_ARCH.QWEN4EXP plus tensors f

If you maintain a skill library for long-horizon agents, this one is worth your time. (bookmark it) It discusses one of most common topics I get asked about these days. It shares some good ideas on ho

Important read if you build with agent skills. Shared skill libraries are treated as a safe way for coding agents to reuse each other's work. New research shows they propagate malware. EvoMal plants a

Got the new Qwen3.8-Flash-Next (176B total, 6B active MoE) running locally and figured I'd share for anyone else wondering if their setup can handle it. **My hardware:** 192GB DDR5 RAM, RTX 5070 Ti (1

I hit my token limit three times a day on my max subscription - got sick of that and designed this MCP setup to shift some of the coding load to my local Qwen3.8-27B model. I've been iterating on it n

arXiv:2608.24920v1 Announce Type: new Abstract: This study examines whether LLM-generated replies remain semantically consistent when the underlying LLM changes. Using messages from real collaborative

I've been building a small tool opensource project that sits between applications and local LLM servers (starting with Ollama) to diagnose silent inference failures instead of just showing logs. Thing

arXiv:2608.24076v1 Announce Type: new Abstract: Evaluation of agentic information retrieval remains limited to scripted interactions with uniform users, missing both natural personality diversity and

I discovered something today that I suspect a lot of AI users may not realise. I’ve been pressing 👍 on ChatGPT responses because I thought it meant: 'Yes, this is a good response, this is the register

The Artist: Qwen3.8-27B-UD-Q3K_XL, q8_0 caches, xhigh, temp 1.0, image-min-tokens 1024, froggeric template) I was screwing around with different Qwen3.8-27B quants and thought of this very simplistic

Introducing GLM-5.3-Flash - Leading capabilities at a highly competitive price - Natively multimodal with a 1M-token context window - A 320B-A18B model released under the MIT License - Previously prev

arXiv:2312.17535v2 Announce Type: replace Abstract: In the past two years, the outstanding performance of ChatGPT in multilingual and multitasking has led to large language models (LLMs) attracting wi

In the past, I have use llama.cpp, but I read that the exl3 quantization format should give better precision, so I have tried exllamav3/tabbyAPI. It was able to write the shown simple HTML game withou

Presented by Tata Communications Enterprises are deploying AI agents, voice AI, and automation across messaging, voice, and digital channels faster than the architecture meant to support it. Most of t

ollama ps settles this in one line, and almost nobody runs it before changing their quant. The PROCESSOR column reports exactly one of three things: 100% GPU - every layer on the GPU 100% CPU - the GP

I ran 11 uncensored variants of Gemma 4 12B that I grabbed from huggingface, sorting by downloads. 10 full abliterations plus 2 LoRA adapters which were requested to be added in the comparison, agains

With hopes of a Qwen3.8-35B-A3B release now mostly dashed, many people including myself are looking at fine-tunes and other variants of Qwen3.6-35B-A3B to run on VRAM-limited hardware. I decided to tr

arXiv:2608.22761v1 Announce Type: cross Abstract: Large Language Models generate text autoregressively, but open-ended generation is prone to verbatim looping, in which models repeat spans already pre

I’ve been going back and forth on this for a week and I can’t settle it, so I’m hoping someone here has hands-on numbers. The two configs (German prices, dealer quote, incl. VAT): Config Price Mac Stu

I’ve been experimenting with a custom inference/runtime research project called CRANE V2, mostly because I wanted to answer a stupid question: How far can you push absurdly large MoE models on an ordi

This summer, we brought in two first-year Waterloo interns (‘rising sophomores’ in 🇺🇸) and tasked them with shipping complex production features that help push the frontier of our document processing

What do I want in AI? Here are just a few things. I want voice-enabled everything. I want things to be more proactive. I want to manage AI’s escalations (versus delegation). I want those escalations t

arXiv:2608.20622v1 Announce Type: new Abstract: Frontier models have collapsed the cost of writing custom code: a niche problem a specialist sees in their own domain now costs an afternoon. The cost o

Hi everybody! Every now and then these days, we’re seeing really huge open-weight models popping up. But since not everybody has a DGX Station at home, I’m interested in really small models. It’s incr

I have deleted tons and tons of older models to make space since I can't afford storage anymore. Easily 10TB... Anyways, I have been considering deleting DeepSeekV3.2 but decide to run it one more tim

Hey everyone! I’ve been working on a 100% local cognitive memory engine designed specifically to give multi-agent systems and NPC societies human-like memory. It's called Eternal Thread. Instead of ju

arXiv:2608.20743v1 Announce Type: new Abstract: Speculative decoding accelerates autoregressive generation by allowing a lightweight drafter to propose future tokens while a target model verifies them

I commented on another Qwen 3.8 27B post that I was frustrated getting anything to work. You all gave some great comments. I nuked openwebui and straightened out my llama.cpp docker config. 1 hour of

Thanks to Unsloth Dynamic 3.0 quants coming in slightly leaner and better preserved, I settled on Qwen 3.8 27B (`UD_Q4_K_XL`) at 100K context as my daily driver for Hermes Agent and OpenCode. On an RT