Friday, 2 October 2026

40 stories · 3 from multiple sources

Gufo performance .... 70tps Qwen 3.8 27b but you need to read the fine print.

So everyone has been yelling about how I should be using Gufo instead of halogen because it's open source and it's "just as good or better". Checking in on their GitHub (GitHub.com/gufo-org/gufo) got me immediately .. "Qwen 27B Q4: 70.56 tok/s single user, 123 tok/s with 8 users" on a strix halo device? Yes please ... So I broke down and tried it today ... Setup: gufo 0.4.0 from their podman…

Friday, 2 October 2026

Gufo performance .... 70tps Qwen 3.8 27b but you need to read the fine print.

Also leading

[AINews] Pi 1.0, Pi Durable, and AIE NYC

Claude-shaped science

More stories

Show HN: Made an open-source Lego AI generator

From the creator of Redis; run LLM locally with ds4

I've ended up with an AI lab in a public community college. What should we actually be teaching?

With most information hidden, the game Stratego had stumped AI–until now

Decision models like Jev don't beat LLM-as-a-judge or traditional classifiers

Buying RTX 5090 At Micro Center Reportedly Now Requires Paperwork, Including A No-Export Declaration

Also today

microsoft/FrogNano-4B-2609 · Hugging Face

Jeff-Qwen3.5-0.8B v1.2 + 9 LoRA adapters: put it in front of Qwen3.8-27B for 38× faster decisions and +8.7 points accuracy, for under 2 GB extra memory

Self-hosting AI does not save money, and I do it anyway

CUDA: fuse shared experts into MMVQ by am17an · Pull Request #29184 · ggml-org/llama.cpp

New Architecture from Percepta: Spotlight — separating intelligence from memory, allowing knowledge and skills to grow without changing the model's weights.

I made my iPhone a second GPU for my 24 GB MacBook: Qwen 3.8 27B prefills 29–44% faster & my holds part of the CTX window.

Hierarchical Continuous Diffusion Language Models

From Knowledge Access to Source Learning: Developing Source-Specific Competence

Finetuning with Sampling: SFT Learns Better Than You Think

Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes

Scalable, Transferable Meta-network for Data Selection Requires a Different Loss (and Why the Obvious Choice is Problematic)

Sean Parker is rebuilding Stability AI around music

Meta open sources code to let you make Muse AI gadgets

Introducing AstaBrief 8B, an open model that turns complex research questions + literature excerpts into cited reports. Run it locally on your own hardware, wit

Greg Kroah-Hartman – Security in the LLM Age [video]

Apple will limit Mac disk access as AI agents ‘substantially’ increase risk

Apple says it’s tightening macOS ‘Full Disk Access’ controls due to new risks from AI agents

OpenAI’s Dot agent is enterprise software that can also order your dinner

Call it AI, call it Super Intelligence, only 2% of consumers are buying it

It’s not AI anymore, it’s ‘super intelligence’ (according to the White House)

Breaking up (with Elon Musk) is hard to do

Circuit Breaker Labs hopes to make AI safer for your kids (and you)

Sites in ChatGPT

Show HN: Graphene – Data analysis toolkit for your coding agent

A model guide for the GPT-6 family

One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars

KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards

ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research