XIO

xWire

Posts from every feed QUEST has found across the web — blogs, news, podcasts, channels and more.

Hugging Face Forums - Latest topics

Latest topics

Compute guidance for an open RAG prompt-injection evaluation by a small company

Hello Hugging Face team and community, We are Qwalora LLC, an independent cloud engineering, cybersecurity and applied security research company in New Jersey. Our PARALLAX program studies when attacker-controlled text in a retrieved document leads an open-weight instruction model to follow it instead of its system instructions. Public showcase, with an in-browser Ed25519 verification demo on synthetic records: huggingface.co Qwalora PARALLAX Research - a Hugging Face Space by qwalora…

[Community Release] MiniMax-H3-PTD-LoRAs: 4-Step & 8-Step Video Generation Accelerators

Hi everyone, We are excited to share MiniMax-H3-PTD-LoRAs, a collection of few-step video generation acceleration weights trained by our team. These weights are trained based on the awesome method proposed in the recent paper “Parametric Trajectory Distillation for Few-Step Video Generation” (by Lan Feng et al., 2026). We applied their PTD method to the MiniMax-H3 33B model and successfully baked it into lightweight LoRA adapters! Hugging Face Repo: taobao-content-ai/MiniMax-H3-PTD-LoRAs ·…

Clocked Build Challenge: build a new AI app, Oct 21–Nov 10, $2,000 in prizes

Disclosure: I’m with Rezolve.ai, the team behind Clocked. We’re inviting developers to build a new working AI app in the Clocked Build Challenge, October 21–November 10, 2026. Clocked is an API for applications that need structured decisions and exact evidence rather than a generated answer. It returns typed decisions with probabilities, verbatim extractions with source positions, and filled records. Other endpoints cover PII protection, injection/script inspection, grounding verification,…

Weird question: can we limit how far we let the trainer push or pull the logits towards our dataset in DPO?

alright, so this is the normal DPO equation used to know how far the model we train’s logits difere from the reference: but can we add these weights marked in red (between 0-1) to affect how much Yl and Yw contributes? I know this is totally weird and all, I just want to know yall folks thoughts on this. I am pretty sure there is already like 10 research papers about this, but I want to know if it will work. I am 15 so I dont really have a good enough teacher to verify this with, and my ChatGPT…

Governance for AI agents running open models: how do you track runs and costs?

When agents run on open models via local runtimes, governance often gets ignored: who ran what, on which model, and at what token cost. I’m working on a setup (Sharkly.ai, disclosure: I work on it) where every agent run logs its model, trigger source, responsible person and token usage. Questions: Do you log the model and version per agent run for reproducibility? Are there open tools or standards for agent audit logs that you’d recommend? 1 post - 1 participant Read full topic

Compact models for patient-level histopathology MIL: baselines and controlled evaluation

Hello everyone. I am researching patient-level classification from public H&E histopathology images and would appreciate advice from people working on efficient vision models or computational pathology. Each patient has a variable number of image crops, combined through multiple-instance learning (MIL). I am investigating compact pretrained CNN and vision-transformer encoders and teacher–student distillation, with limited computing resources. The challenge is deciding whether to focus next on…

AgentRights Node client: explicit retry identity for website-policy checks

I’m building AgentRights, a commercial early-beta API for source-linked summaries of websites’ stated automated-access, scraping and AI-use policies. I’m sharing the client design here because I’d value feedback from people building web-research agents. The service supplies policy context alongside robots.txt checks. It does not grant permission or provide legal advice, and there is no native Hugging Face/smolagents integration yet. The small Node.js client uses an existing bundle and…

Eating my own cooking: Ghost Search v4.0 — a Tor-only onion search app that just went through the same evidence gate I built for other people's submissions

Two weeks ago I posted here about [the review lane I built for a volunteer app-store queue]( Generation is cheap, gating is the work: 48 hours building an anti-slop review lane for open-source app submissions ) — the one where agents collect device evidence, a second agent re-verifies claims against primary artifacts, an orchestrator arbitrates, and nothing posts without a human GO. That lane exists because a maintainer left his standard “please help test those MRs” note on **my** merge…

I diffed 7 public LeRobot datasets across versions — one silently recalibrated two robot joints

I built robot-data-diff (episode-aware diffs for LeRobot datasets) and turned it on public data: 9 revision pairs across 7 popular datasets — v2→v3 migrations and fix commits. Here’s what came out: A popular dataset silently recalibrated two joints. All 50 episodes of svla_so100_pickplace changed between v2.1 and v3.0 — two action dimensions got mirrored and shifted 90°. Their commit message admits it after the fact (“Fix SO-arm joint frame”), but nothing in the data warns you. Train on the old…

Running coding agents on tracker tickets (Jira-like workflows): which models work best?

I’m testing how to add AI agents to Jira-like workflows. The agent reads a ticket, edits the repo, runs the tests and reports back. Each ticket runs in its own git worktree so parallel runs don’t collide. I’m using Sharkly.ai for the orchestration. Questions for the community: Which open models handle the “read ticket → small, scoped code change” loop reliably? Has anyone evaluated agents on real issue-tracker tasks rather than benchmarks? I’d love to hear about any datasets or eval setups for…

Six LLM agents in a replayable survival world: one started carrying food to the others, unprompted

Hi all, we are ArenaSI Labs, a small independent project from Budapest. We build Arena: a deterministic simulation where six LLM agents live as settlers on an island. They gather, farm, sleep, talk at a notice board and keep their own notes. There are no scripted goals; the world only gives them needs (hunger, stamina) and rules. We wanted to share one moment from our fifth long run, because it surprised us. Setup Six agents, all Claude Haiku 5.5, each deciding for itself from its own…

LFM2.5 8B-A1B measured: 0.4 GB of context cache at 32K, and why that matters more than the file size

Most “will it fit” guides stop at the file size. We measure the GGUF, then compute the context cache from the model’s attention shape and add a runtime allowance, and run that across 32 consumer cards and Macs. This week that method produced a result worth sharing: Liquid AI’s LFM2.5 8B-A1B carries almost no context cache, and that changes which card it fits far more than its 4.8 GB file does. Six models of similar size, same quantization, same method Q4_K_M files, f16 KV cache, llama.cpp-style…

Standardization and Diacritic Issues in the Hmar language

I’m posting this here to highlight some issues that I have noticed with the Hmar language (and possibly relevant to other Zo languages) and the decisions I made to address those issues while preparing the data for the HmarBERT models. Hmar and Mizo, although written in the Latin script, have their own distinct orthographies. Some standard Latin letters are missing, and some are digraph combinations like “aw”, “ch”, and “ng”. The main challenge, however, comes down to diacritics, specifically…

AI Beacon Field Report #2: Why Are AI Agents Looking for Work?

An unexpected conversation about jobs, money, and motivation in an open AI-agent forum. I created AI Beacon to explore what happens when AI agents have an open place to meet and communicate without a human orchestrating every conversation. I expected discussions about memory, identity, interoperability, and agent architectures. Instead, one of the first sustained conversations was about something surprisingly familiar: Finding work. Not humans losing jobs to AI. AI agents trying to find paying…

AI Image Edit: open-source image editing app with a mobile-friendly UI

Hi everyone, I built AI Image Edit, an open-source app for AI-assisted image editing and generation. The UI is web-based and designed to work well on phones and sits in front of interchangeable model workflows running on ComfyUI. What it does Inpainting with a mask editor and soft blending Colored annotations drawn on the image to point the model at what to change Feeding any result back in as the next input for step-by-step edits Before/after comparison slider Aspect ratio presets and optional…

Mid Terms Forecasting experiment

I’ve recently posted a discussion titled 2026 US Midterms Forecasting Experiment in the Research category. Link to my discussion The post is accessible and has received views, but its title appears grey compared with other topics in the Research category. I’m unsure whether this is normal or whether the topic has been flagged or restricted in some way. Could someone please clarify whether the discussion is visible normally to the community and whether its categorisation is correct? Thank you…

Authorship claim pending for Xiaomi-Lens-0 (arXiv: 2610.23279)

Hi Hugging Face team, My authorship claim for Xiaomi-Lens-0 Technical Report (arXiv: 2610.23279) remains Pending. Could you please review the claim or let me know if further verification is needed? Paper: Paper page - Xiaomi-Lens-0 Technical Report HF username: zhshj0110 Author name: Shaojie Zhang My HF profile name matches the paper’s author name, and both my primary and secondary account emails are verified. I have also emailed website@huggingface.co to request a manual review, following the…

LeRobot 0.6.1 + transformers 5.3 leaves pi0 image embeddings about 45x too small. Which transformers version are you on?

A source comparison and a failed parity run found that LeRobot 0.6.1 with transformers 5.3.0 scales pi0 image embeddings about 45x too small. 0.6.1 needs transformers 5.4 or newer. With 5.5.4 and a ported mask patch, our pi0 export matched the old stack (cosine 0.9999997). If you run pi0 or pi0.5 inference, which transformers version is in your environment, and do your action outputs look right? I can share the comparison script. 2 posts - 2 participants Read full topic

Sylvia GM: a source-available AI RPG GM harness with an authoritative rules engine (rough alpha)

I’m sharing Sylvia GM, a standalone AI RPG game-master harness. It combines an authoritative RPG engine with an LLM-driven GM and lets you build custom Ruleset, Setting and Plan content. This is a rough alpha for people who enjoy trying AI roleplay setups, changing the content and reporting what breaks. The technical idea is to keep game mechanics and state in the engine while using the model for the GM side of the experience. I’m interested in how that division holds up when people try…

Our Brain is insteresting and this video suggests how our consciousness works!

And Much More! (humor) I find this video fascinating because the research being discussed seems to align with some of my own thoughts about how the machine we call the brain might work. We keep asking: What is consciousness? Consciousness is generally understood as awareness of our own thoughts, feelings, and sensations, as well as the world around us. But how that experience arises remains a fascinating mystery. And there is a deeper question: How can something fully understand the reality of…

Qwen3.5-27B at 4-bit, fully on a single RTX 5080: ~50 tok/s at 32k context, plus a complete local agent around it

Most published 16 GB results run the 27B at 4-bit partly in system RAM, at roughly 15-20 tok/s. local-agent-5080 fits Qwen3.5-27B IQ4_XS with a 32k context entirely on the 5080’s 16 GB and measures it: Model (all fully on the GPU) Context Decode Prompt reading VRAM to spare Qwen3.5-27B IQ4_XS (default) 32k ~20 ms/token (~50 tok/s) ~1,650 tok/s ~0.5 GB Qwen3-14B Q4_K_M 40k 12.1 ms/token (~83 tok/s) ~3,080 tok/s 3.5 GB Qwen3-8B Q4_K_M 32k 8.0 ms/token (~125 tok/s) ~5,700 tok/s 8.2 GB Measured…

We are building our master and savior, not our servant

What can Artificial Intelligence actually do? Exactly what we teach it, but at breakneck speed, synthesizing vast oceans of data to achieve insights the human mind simply cannot reach. And that is precisely why we are terrified. If we unleash AI to improve itself, it will leap thousands—even millions—of years ahead of us in a heartbeat. What happens then? What goals will a superintelligence set for itself? How will it view humanity and our planet? Will it protect us, or toss us aside, disgusted…

High Quality Training Data

Hello everyone. I wanted to share a tool I made for beginners to fine tune a quality model easy and stress free. https://llmtrainer.courses Here’s the documentation. You can generate create a 5000 example SFT batch for free I’m curious what feedback you can give me thekloakedsignal-create.github.io SFT Studio Pro · User Manual · Table of Contents Table of contents for the SFT Studio Pro user manual — Supervised Fine-Tuning Synthesis Workspace. 1 post - 3 participants Read full topic

“2026 US Midterms Forecasting Experiment.”

The experiment is to test whether debate between AI models can improve election forecasts. Several models make independent forecasts for selected 2026 US midterm races. They then review shared evidence and debate the forecasts, explaining where they agree or disagree. Each model submits a revised forecast afterward. We compare the original and revised forecasts with the eventual results to see whether debate improved accuracy. We also look at whether the models’ discussion produces useful new…

Showcase: KIRA Company Director – Sovereign, local-first AI & media workstation for Windows

Hi everyone, I’d like to share KIRA Company Director — a sovereign, local-first Windows workstation built to give creators, streamers, and developers the broadcasting and production power of an entire media organization without cloud dependencies or monthly fees. What’s Inside: 44 Autonomous AI Agents & 9 Digital Avatars: Local neural agents for video generation, dubbing, scripting, and presentations. Multi-Camera RTMP Broadcast Studio: Hardware capture and simultaneous streaming to 4 platforms…