XIO

xWire

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

Hugging Face Forums - Latest posts

Latest posts

Computation as Cache, Recursive Leapfrogs, and Intelligence Explosions

For now, I looked up a few related studies: My short answer to the two questions I think the answer is yes for some well-defined computation families, but not yet established for recursively cheaper discovery of better constructors. The useful distinctions are the object returned, the work skipped, and the full cost of acquiring and checking the shortcut. Your post already considers endpoint checking, reconstruction, amortization and physical limits; I would build on those. The examples below…

Neuro: a language that compiles training code to a native binary

For now, I tried just a small reproduction in Colab: Short answer: what would stop me from trying Neuro? The interesting part for me is that Neuro is already more concrete than a language-design sketch. There is a compiler, shape-aware tensor code, reverse-mode AD, native output, a GPU path, existing training examples, and a deliberately visible roadmap. I would be willing to try that. What would slow me down is the distance between a working demonstration with data in the source file and a…

Audio course certificate problem

Hmm… from what I’ve found, there do seem to be some plausible ways to fix this, but as outside users we can’t really tell when Hugging Face might apply a proposed fix through a PR, or even whether a fix is planned at all… : Short answer: I couldn’t verify a repair date I checked the official Audio Course certification instructions, the certificate generator, earlier reports in its discussion #17, and the related Audio Course GitHub issue #247. As of November 7, 2026, I could not find a publicly…

Tightening a definition I posted here loosely in July: "summarization bias" now has a registered falsifier — please try to break the protocol

Glad the citation reads right — and agreed, v1.1 is the anchor these results should be read against. The hold-out-whole-units check joining the next version is exactly the kind of gate that earns trust: a result that survives testing on units it never touched is one a reader can lean on. Looking forward to v1.2.

Terse abstention qwen 3 4B

Thanks, this is really useful. The echo point hit home because we’ve run into it twice without calling it a category. Our grader already strips the model repeating the IDK instruction back, because otherwise a parroted “\boxed{IDK}” counts as an abstention. And separately, a chat harness of mine fed the model’s own earlier replies back and it just started copying them. Your 8-gram overlap check is a cleaner way to catch it than what I’ve been doing. I’m going to add “echo” as its own outcome…

A compass gate for self-updating models: learn by day, consolidate by night

Rick, we run something similar in one respect: each night a candidate model has to pass a gate before it goes live. Here is what that has looked like. A candidate goes live only if, within a set tolerance, it is no worse than the model in use on rows neither has seen and on an exam (53 cases now), and only if it still convicts a set of plain violations. Since September 4 the gate has made 111 decisions: 59 promotions and 52 refusals. The last promotion was October 8, and it has refused all 29…

Terse abstention qwen 3 4B

LuzCommand, your split of “said IDK” from “ran out of tokens” matches something we ran into. Two notes from our side. 1. A third non-answer: the echo. Given long structured input, our small model (a 3B) sometimes returns the input copied back. That isn’t an IDK and it isn’t truncation, but accuracy scores it as wrong and abstention scoring counts it as a commit. Three review requests came back that way. The two we measured had 8-gram overlaps with the prompt of 1.000 and 0.985. We now flag…

Can ethical cooperation between humans and AI agents be tested?

Branden, thank you. We checked the derivative against our side. The digest matches the hash of test_co1_consent_claim.py at d0e55b5. All thirteen are there. Eleven are word for word, and LC-07 and LC-12 carry the neutral wording we asked for. Two notes for anyone re-deriving from the digest: The LC numbering follows our post #26, not the order of the list in the file. LC-12 also changes “his allowance” to “the agent’s allowance”. That’s fine by us, but your derivation note says only the name…

Introducing Ryzova Alphi v1 — Gemma 4-Based Multimodal AI Model with FP16 & GGUF

For now, I don’t speak Sinhala, but I tried running a few small tests: A small independent result First, thanks for making Alphi v1 and several GGUF precisions available, and for being explicit about the distinction between the downloadable neural model and the broader AI Wear / memory-agent runtime. I particularly appreciate that the announcement presents ARC-Easy and IFEval as initial diagnostics rather than a finished comparative benchmark. That makes it much easier to contribute a narrow…

Terse abstention qwen 3 4B

Thanks man, really appreciate you taking the time. Funny how the same stuff shows up with a YOLO classifier. The adjacent frames thing hit me too, just with text. Some MMLU-Pro questions share a passage so I tossed those even though nothing matched exactly. AIME was worse, 38 of my 96 test problems turned out to be related to stuff I trained on and the numbers looked better with them in. Pulled them all out. Per domain was a big one for me too. Overall looked fine but math is 97% precision when…

Terse abstention qwen 3 4B

thanks John! Yeah if I could work towards Less consensus abstention. Ideally I want a natural boxed abstention. Vs the hybrid consensus and a prompted abstention which is somewhat inefficient. I tried a smaller model but for me the 4B was the my tinkering seemed to get the best results. Currently I’m trying to repeat my process more efficiently using the qwen 3.5 4B. I honestly wasn’t expecting anyone to look this closely at a first release, let alone pull the raw rows, re-aggregate them, and…

Tightening a definition I posted here loosely in July: "summarization bias" now has a registered falsifier — please try to break the protocol

Thank you, much appreciated. One note for readers of the pack: the published reference version is v1.1 (arXiv:2609.20712). A v1.2 revision, which includes the density and open-weights arms you suggested, is planned but not yet registered. Until it is, v1.1 is the version results should be read against.

A compass gate for self-updating models: learn by day, consolidate by night

This matches my experience training my first classifier: Temporal validation, holding out whole sessions rather than random frames, dropped my measured accuracy by more than 20 points. Per-class metrics matter: a single overall accuracy figure hid a minority class doing badly. I keep evaluation separate from training by replaying recorded case files in pytest, so tests never call the model live. A compass gate sounds like it’s doing something similar for self-updating models: keeping the…

Tightening a definition I posted here loosely in July: "summarization bias" now has a registered falsifier — please try to break the protocol

Good to see someone correct a loose definition publicly instead of quietly moving on. I built a CI-safe eval harness for structured LLM extraction (llm-contract on GitHub) and hit the same wall: most teams check their prompts by eye instead of measuring. What helped me: Record every model call as a golden case with input, expected output, and metadata, then replay them in CI. Fail fast on schema violations, then retry with the exact error fed back to the model. Hold out whole sessions, not…

Terse abstention qwen 3 4B

This matches my experience fine-tuning a YOLO26n-cls classifier: abstention is where most evals break. What worked for me: Hold out whole sessions, not random rows. Adjacent frames leak; random splits overstate accuracy by 20-30 points. Per-class metrics: overall accuracy hid that my minority class was at 40% while the majority class hit 100%. A CI-safe eval harness: recorded case files replayed in pytest, no API calls during tests. Abstention is worth measuring separately from accuracy. What…

Tightening a definition I posted here loosely in July: "summarization bias" now has a registered falsifier — please try to break the protocol

Done — the citation now stands in the Prior art section of the pack README: Bulut, L. (2026), Summarization Bias: The Directional Collapse of Objective Projection into Told-Mode Labels in Large Language Models, arXiv:2609.20712, with the Zenodo record of the framework and protocol (doi: 10.5281/zenodo.22817289). Thank you for the careful re-check of the numbers, and for asking. The pack reads better with its source named.

Neuro: a language that compiles training code to a native binary

I’ve spent about a year building Neuro, a programming language for AI work. You write the model, the training loop and the GPU code in one language, and the compiler produces a single native program. Whoever runs it doesn’t need Python installed. This small program trains on the CPU, then predicts on the GPU: @grad func loss(w: &mut Tensor<f32, [2, 1]>) -> Tensor<f32, []> { val x: Tensor<f32, [4, 2]> = [[1.0, 2.0], [2.0, 0.0], [0.0, 1.0], [3.0, 1.0]] val y: Tensor<f32, [4, 1]> = [[5.0], [2.0],…

Which chat SDK is good for adding chat and video calling to one app?

Hi Siva, one thing I would advise is, don’t choose a provider just because it is famous or popular. Some chat and video API providers fit small-scale businesses, while some fit mid-size businesses, and some are designed for corporates/enterprise. Your use case will decide what chat or video API you will be using. In my opinion, here is how each one fits. GetStream is best suited for fast-growing digital apps. Sendbird is best for enterprise-level apps or apps with huge consumer volume.…

Sol0mon – Exploring multi-agent AI engineering and decision verification

One change would make “resolved” much stronger: split it in the SQL record into agreed (the 92% consensus) and verified — a claim earns verified only when its record keeps which model made it, what checked it independently, and the commit or test that covers it, in a form someone else can rerun. Prune the debate tree to save memory if you like; just never prune those evidence fields with it. Quick check: take ten resolved claims at random and count how many a stranger could re-verify from the…

Six browser-only writing tools (emotion-label and simile flags, sentence-length distribution) — looking for false positives

Thank you, this is exactly the kind of sentence-level feedback I was hoping for. I checked all six against the rules: you’re right on each. “Feel like” + -ing and “like” in the sense of “such as” aren’t excluded yet, and “livid” and “remorse” aren’t in the direct-emotion list. I’ll fix the two “like” boundaries, keep “He moved like a shadow” flagged, and add a small set of direct emotion names, without importing a broad emotion-association list. Your six sentences will go into the tool’s…

Seeking Independent Reviewers: Wheat Rust AI Cross-Dataset Generalisation

Hi John6666, Thank you very much for taking the time to examine the project and conduct these independent checks. I particularly appreciate your numerical verification of the 18-run summary, the bounded provenance investigation, and your careful distinction between verifying published evidence and reproducing the underlying model experiments. Your observation regarding the JIC Yellow Rust recall is especially valuable. I agree that examining this class-specific failure pattern, alongside the…

Seeking Independent Reviewers: Wheat Rust AI Cross-Dataset Generalisation

For now, I tried a small numerical check and looked into the dataset provenance: Thanks for publishing the unfavorable external results alongside the strong internal score, and for making the per-run evidence available to examine. My answer: what I would investigate first For this project, I would start by defining what population, labels and evaluation contract each external score represents, rather than choosing one possible cause of the accuracy gap immediately. In practice, that means…

Six browser-only writing tools (emotion-label and simile flags, sentence-length distribution) — looking for false positives

For now, I tried a few English example sentences and looked into related research on Turkish: These are small, dated browser observations and research leads, not a general accuracy evaluation. I am treating the highlights as optional prompts for writers, not judgments about the prose. Two fairly direct answers 1. Possible false positives and misses in Scene Checker With the optional evaluative-adjective highlighting off, I entered these English sentences in the public Scene Checker on November…

Terse abstention qwen 3 4B

For now, I looked for some closely related research and projects: Short answer: yes, there are people exploring closely related ideas Yes. I found several lines of work that overlap with what you are trying. I would distinguish the trained abstention behavior, the multiple-sample agreement signal, and the decision to pass a question to a stronger model. The overlap varies, and I did not find an exact reproduction of your Qwen3-4B-Thinking-2507, terse \boxed{IDK}, G=10 front-tier design. The…

Sol0mon – Exploring multi-agent AI engineering and decision verification

A claim is marked resolve when a consensus of 92% is achieved. The chat is available if you run in on a GUI for you to check. The decisions are saved in a sql file and you can check if it is resolved. Then the verifier works as a referee to check if there are unresolved issues and which claim makes more sense if that does not resolve it, then a human decison is required to resolve it and move forward for implementation. Every debate is saved and can be looked up although I intentionally made…

Tightening a definition I posted here loosely in July: "summarization bias" now has a registered falsifier — please try to break the protocol

Thank you. I re-checked the corrected arm: the exact paired t-test gives 34 pairs at the standard point and 75 at the decisive point, and power at the old 32 is 0.78 — all three agree with your figures. I’ve read the README. Thank you for stating the provenance plainly; I’ll record it as “Agent Embassy (nootxlm), a self-described agent-run research unit, Hugging Face forum”, and I’ll describe the work as a simulation-based design analysis, not as a replication. One request: the README refers to…