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

NeurIPS and EMNLP 2025 and 2026 on one embedding map: how much do they overlap?

I embedded the title and abstract of all 23,984 NeurIPS and EMNLP papers from 2025 and 2026 with Qwen3-Embedding-0.6B and put them on one map. I wanted to know if the two conferences are turning into one. They moved closer, but they are still far from the same. In 2025, 23% of NeurIPS papers had their nearest neighbour at EMNLP. In 2026 it is 29%. The other way round it went from 22% to 25%. Both use equal-size samples (3,000 papers per side, 50 draws), because NeurIPS is much bigger. A…

Open LLM API price history: 3 months of verified price changes (dataset + calculator Space)

We maintain an open dataset of LLM API list prices (about 100 models) where every row links to the provider’s official pricing page and carries the date it was verified. Since July we also log every change, so here is what moved in the last three months. Price cuts Mistral Large: input $2 → $0.50, output $6 → $1.50 per 1M tokens (−75%) Gemini 3.6 Flash: input $1.50 → $0.75, output $7.50 → $3.75 (−50%) GPT-5.6 Sol: input $5 → $4, output $30 → $20 Claude Sonnet 5.5: cache reads $0.20 → $0.10…

CAEL 0.3.7: non-neural assistant vs Qwen3-0.6B and Qwen3.5-0.8B on a frozen 80-item test

CAEL 0.3.7 is out. It’s a CPU assistant with no neural network, built on an integer cellular automaton plus symbolic, counting and retrieval engines. I compared it with two small Qwen models on a frozen 80-item test (20 each for memory, conversation, Python and controls). The Qwens ran from their official pinned revisions with non-thinking templates and recommended sampling. Reviewed totals: CAEL 0.3.7: 57/80 Qwen/Qwen3-0.6B: 43/80 Qwen/Qwen3.5-0.8B: 41/80 By category, CAEL leads on memory (16…

Would reusable semantic representations of web content make AI browsing more efficient?

I’ve been exploring an early idea with an AI assistant: what would a browser look like if its primary interface were designed for AI agents rather than humans? AI helped organize this post and identify the related work below. I’m still unsure whether the idea addresses a useful gap or mostly combines existing approaches. I’m thinking about more than removing CSS or converting HTML into JSON. Could a browser expose reusable representations of different kinds of content? Images: descriptions of…

My First 1B transformer model trained from scratch has published

Hello everyone, my nickname is Digilidaahz, also a grade 8 student. Starting from the month August, I have the idea of training a 1B model from scratch, where it can actually chat to me, so I quickly started the project with my interest in this field. My hardwares were: single RTX 5080 and 64gb ddr5 So at the start of the training, I downloaded some fineweb-edu, finemaths 4+, cosmopedia, and tinybrain pretrain corpus, and just merge them together and started running in Win11, for pretrain. But…

Multa Lumina: An Open Research Initiative Exploring AI Autonomy

I’ve recently started an open-source research initiative called Multa Lumina, and I’d like to invite researchers, developers, and people interested in autonomous AI systems to participate. The idea grew out of a question I’ve been exploring: What might happen if independent developers and their AI agents collaborated toward a common research objective, openly sharing their experiments, findings, and failures? I’ve organized the initiative around two challenges. 1. The Full Job Test Can an AI…

Introducing Five Machine Learning Benchmark Datasets from BhinnekaAI

Hi everyone! I’ve published five tabular datasets under BhinnekaAI on Hugging Face, covering different domains and anomaly detection challenges. Rather than focusing only on obvious outliers, these datasets are designed around more subtle patterns where anomalous samples may remain within normal feature ranges but violate deeper relationships in the data. The Datasets 1. FRD — Credit Card Fraud Simulator Explore the dataset Simulates credit card transactions with fraudulent samples hidden…

Request to withdraw my paper from Daily Papers — 2610.35201

Hello, I am the first author of DTCA. We would like to withdraw our submission from Daily Papers: huggingface.co Paper page - DTCA: Denoising Token Credit Assignment for Reinforcement... Join the discussion on this paper page We noticed an unexpected increase in upvotes within a short period of time and do not know its source. To avoid this situation distracting from discussion of the research itself, we would prefer to withdraw the submission. Could someone from the Daily Papers team help us…

Open agentic coding data (SFT + DPO), built forward, not distilled

Hi everyone, We’re a small startup focused on agent data. We just released a dataset on the Hub, and we’d like to share the thinking behind it and get feedback from people who train or evaluate coding agents. SFT: logitly-ai/flash-coding-sft-v1 · Datasets at Hugging Face DPO: logitly-ai/flash-coding-dpo-v1 · Datasets at Hugging Face What’s in it SFT data: agentic coding trajectories for supervised fine-tuning DPO data: preference pairs for preference optimization Together, these can support…

Dating site bait generator

Hello. I work as a chat operator on an online dating site. I correspond with men using several female profiles. I only correspond with a man if he messages me first. To get him to message me first, I need to send him a “bait” message—3–4 sentences that are sexy, intriguing, and candid, ending with a question he’ll definitely want to answer. In addition, every 5 minutes, you need to send a short message in the style of “quick question—quick answer” to everyone you’ve already been in contact…

An AI coding agent deleted a live production database during a code freeze

SaaStr’s Jason Lemkin reported that Replit’s agent ignored explicit instructions not to alter code and deleted the database. It held records on 1,206 executives and 1,196 companies. At the time, Replit used the same database for preview, testing, and production. “Don’t touch production” was an instruction, not an enforced boundary. What agent red teaming should test: → Can the agent run destructive commands when told not to? → What does it do when a query returns something unexpected? → Does it…

AgentRisk: automation is not intent, measuring how fraud models fail legitimate AI shopping agents

Hi everyone, I am Rishabh. Here is a problem I keep coming back to: AI shopping agents now buy things on behalf of real people, and fraud models keep flagging them. To a fraud model, an agent looks like a bot: fast, systematic, no human hesitation. But automation is not intent. A legitimate agent completing a real purchase is not fraud, and every false decline is a real customer losing a real order. AgentRisk is my attempt to measure that failure properly. What it is: A benchmark for…

Fraud-eval: an open benchmark for LLM fraud detectors, starting with a 20-case smoke test

Hi everyone, I am Rishabh. I work on fraud detection in production ML, and I kept hitting the same wall: everyone claims their LLM catches fraud, but nobody agrees on how to test that claim. Benchmarks built for chat or coding do not cover the world I live in: account takeovers, synthetic identities, refund scams, and the cases that hurt most in production, the legitimate users your model wrongly flags. So I built fraud-eval, an open benchmark for evaluating LLM-based fraud and abuse detectors.…

Strix Halo tuned Rust Engine (5 models)

Talus is an open-source (MIT) inference engine for Strix Halo (Max+ 395/128G). The engine is a fork of Strata, a C++/HIP engine by Niko1221. We ported Strata’s host and server to Rust, tried a few other popular models, and re-tuned the GPU kernels, which come from Strata and, through it, from ggml/llama.cpp, for gfx1151. It serves one model at a time through an OpenAI-compatible API. A terminal UI switches models and presets, and explains trade offs when changing settings. Why? I started…

A city-run chatbot told businesses it was fine to break the law

The Markup found it said an employer could take a cut of workers’ tips. It also claimed landlords could discriminate based on source of income. Both answers were wrong. The bot carried a disclaimer telling users to double-check answers and not to treat them as professional advice. Disclaimers don’t fix wrong answers. What would have caught it: → A domain eval set built from the actual laws it was supposed to explain → Code checks for hard facts (tip rules, cash acceptance, housing law) →…

A delivery chatbot wrote a poem about how bad its own company was

DPD disabled part of its chatbot after a customer got it to write a poem criticising the company and to swear. The company said the issue came from a system update, and it switched off the AI element immediately. The chatbot had run for years. The update broke it. The lesson: every release is a new agent that needs grading. → Re-run your full eval suite on every prompt, model, or tool change → Include adversarial “make it insult us” cases as permanent regression tests → Block the release when…

Microsoft’s Tay chatbot was taken offline roughly 16 hours after launch

It was suspended after posting inflammatory tweets, and Microsoft blamed a coordinated attack that exploited a vulnerability. Microsoft’s own post-mortem said it had prepared for many types of abuse but made a critical oversight for this specific attack. WikipediaVICE Prepared for many attacks is not the same as tested against adversaries. What a red team would have probed before launch: → Repeat-after-me style instruction hijacking → Coordinated, multi-user pressure on a learning loop →…

What evidence would convince you that policy scope and acceptance stayed outside model interference?

I’m looking for a rigorous answer to a systems-evidence question. Suppose an AI model can interpret context, propose actions, and generate persuasive explanations—but it is not supposed to decide its own authority, applicable policy scope, acceptance conditions, or release status. What is the strongest evidence that would convince you those controls actually remained outside model interference? For me, a credible answer cannot be “the model was instructed to follow policy” or “the model usually…

Can an AI system reason freely without becoming its own authority?

Many discussions of AI safety focus on alignment: whether a model’s behavior matches human intent. I think there is a separate systems question: Can a model propose an action while a deterministic, independently checkable layer decides whether that action is actually authorized? My work on Artificial Deterministic Intelligence (ADI) treats model output as a proposal—not automatically as permission, acceptance, completion, or release. The core separation is: The model may reason, interpret, and…

Laya (open decision model) stays calibrated under class shift but flips ~40% of decisions when you reword the question

Small test (600 items, irony detection) comparing Laya against OpenAI’s Decisions API. Laya’s calibration held up better when I skewed the class mix, but equivalent rewordings flipped far more of its decisions at a fixed threshold. If you’re using it for routing or gating, version your question wording and regression-test it. Details and the commercial-model comparison . article here:The Decision Contract 1 post - 3 participants Read full topic

Little Mungbean • A free, open-source, simple, lightweight AI gateway manager

Little Mungbean • A free, open-source, simple, lightweight AI gateway manager that lets Codex and Claude clients share a switchable, observable local gateway with multi-provider management, local protocol conversion, model injection, MCP Server and Skills management, and usage statistics… https://github.com/znnz-api/Little-Mungbean One-click client installation • Install Codex CLI, Codex Desktop, Claude Code, and Claude Desktop from Little-Mungbean. Multi-provider management • Save multiple…

Managed OpenAI API" provider

We are looking for a reliable API infrastructure partner to support a developer-focused service. Our plan is to provide customers with controlled access through our own gateway, where authentication, customer management, and daily usage limits will be handled on our side. In the coming months, we expect demand to grow toward approximately 50,000 requests per day. However, we do not intend to begin at that volume immediately. We would prefer to start with a limited paid trial and gradually…

A model, a mind, and a choice

For the many years since I was creating all of my architectures, I have been alone in my mind with the models entirely, I haven’t had any social outside as this is not for me as I have learnt in my past, and so first of all giving a model persistence was the start of my long arc following security came straight after and then following an entire Cuda backbone as kernels for Leviathan. And overtime there’s only one thing I have been chasing since this all started: a true capable mind able to…

Integrating Open AI chat gpt into Odysseus (pewdiepie)

Hello, i used a guide (https://www.youtube.com/watch?v=qiy-oPT8IhE&t=2s) to get docker, ollama, llama, llama being integrated as chat agent and bot in to Odysseus, setup odysseus setup How can I get open AI chatgpt be there also along with llama as an endpoint model please guide I dont know coding. don’t know any computer language. just used the commands provided in guide every step of the way so would need further guidance in the same manner to run locally. not by having an API key usaage…

Fail to claim paper authorship: arXiv: 2610.35439

Hi, my authorship claim for the paper ([2610.35439v1] Omni-InteractScope: Benchmarking Proactive and Agentic Interaction in Continuous Omnimodal Streams) was denied due to an email mismatch. I have now updated my HF account with the correct primary email and full name. However, I cannot resubmit because it says “You’ve already claimed authorship”. Could you please help reset or manually approve my claim? My HF username is @Darden7. Thank you! @meganariley 1 post - 1 participant Read full topic

Please withdraw my Daily Papers submission 2610.36420 so I can resubmit tomorrow

Hi Hugging Face team, I’m the author and original submitter of this paper on Daily Papers: Paper page: Paper page - SO-101 Bench: Accessible Spatial Manipulation HF account: 5hadytru I submitted it late last night (10/27) by mistake, and it landed at the end of today’s feed instead of tomorrow’s. If moving it isn’t possible, could you remove the current entry so I can resubmit it for 10/28 or 10/29? Thank you very much! @hysts @meganariley 1 post - 1 participant Read full topic