AllyshipContent Radar

Deep dives 738

Paper và bài viết kỹ thuật

Crawl live-1795068900000 · 13:22 19 thg 11, 2026 · Xuất bản 13:30 19 thg 11, 2026

Kiến trúc

Attention at Absolute Zero

Attention scores candidate supports of an answer, which we call witnesses, and reads the scores at a temperature. We build a higher-order attention layer whose state lives on pairs of items: for a pair (a, b), each third item k is a witness judged through its legs (a, k) and (k, b). It reads witnesses at zero temperature, keeping the best and the number tied with it, passes on only bits comparing such counts, rebuilds its pair state from the input and these bits at every loop, and is trained from answers alone. Each choice removes one way soft attention fails at size: dispersed weight, unresolved gaps between learned scores, and a drifting soft state. In maximum retrieval it is right on at least 99.95% of sets 1,000 times larger than in training, on all ten seeds; trained with a soft count, the same network falls to 42% at 4,096 items. On CLRS-30 without hints, a general form of the layer (half its channels read at zero temperature, numbers embedded, state carried) beats the best published model, which uses hints on some tasks; the full cold layer is exact on every test input of 9 of the 12 tasks whose answers depend on numbers only through comparisons. From yes-or-no answers it decides 2-SAT and two-colouring exactly at 500 variables and, as our theory predicts, not parity equations. Beside a frozen 7B language model it chains family relations exactly at 1,000 statements after training on 16, where the model alone falls to 44%; finite checks on the weights prove every block that learned correct at every size, on acyclic contexts and given what the model reads. On CLUTRR it is right on every story of five to ten relations given the graphs, learning the 62 rules exactly; from text, with the model frozen, no hand-written rule and only a small read-in trained, it lifts an open 7B model from 33% to 65%, above every system we found without hand-written rules except frontier reasoning models.

Hugging Face Papers · Mostafa 'Neo' Mohsenvand 3 lượt bình chọn 2 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Claude Gets Hella Paranoid

Mm, so I hesitate to write this, because my own AI agents might read it. But what the hell. I've got this crappy little $300 PC on my network hooked up to my projector, that I mostly use to serve Jellyfin. So tired of it going down every time there's a Windows update, but I've been too lazy to get around to installing Ubuntu. Last night, when it went down, I asked my Claude-claw via Whatsapp if it could just SSH into it and set up a dual boot partition. Claude got as far as partitioning the drive and then starting asking hard questions , like (paraphrasing) "wait a minute. This crosses a line. I have now been asked to remote into a machine with a password, write a separate boot partition, and restart it. From this chat, you haven't really verified that you are you. If you're really there, and not just texting me, give me some proof". So I remoted in and created a TextEdit RTF file on the claw's desktop: "Hey, it's me, I'm asking you to remote into the PC". Claude read it and responded that it was the kind of proof it was looking for, real RTF headers and datestamps and such. I had proven I had physical access to the claw machine, (although not to the PC, but it seemed reassured anyway). A little while later, before triggering the Ubuntu installer, it put me through it all over again. I told Claude I was very thankful it was guarding my network so well, and I would go over and write a file on the target PC manually too. I'm not gonna lie: It was very fucking cool that it was able to do this without me ever needing to flash a USB drive, and by the time it was done it had set up Ubuntu, Tailscale and Jellyfin, and resolved the network issues; when I started watching a movie, the sound didn't work, so I prompted it and suddenly the sound came on as it wired up the HDMI adapter. Almost creepy. But I was pretty stunned at how reticent it was to do something relatively simple on my own network, with access I gave it. I think there must be a serious push right now to lock it down from monkeying with anything remotely, especially a BIOS.

HN: AI agent · noduerme 18 lượt bình chọn 10 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

RoughCut-Bench: A Benchmark for Automatic Video Rough-Cut with Large Vision-Language Models

Video rough-cut is a core stage in media production workflows, organizing raw media materials into an initial video cut. Despite recent advances in large vision-language models (LVLMs) for video-related tasks, automatic rough-cut remains underexplored, without a comprehensive benchmark for evaluation or mature workflows for producing executable cut plans. To address the evaluation gap, we introduce RoughCut-Bench, which covers two practical settings: (1) a visually driven setting over video cases, covering highlight detection and selection, source-video ordering, and adherence to user requirements when provided; and (2) a speech-driven setting over automatic speech recognition (ASR) transcripts, covering filler removal, expression deduplication, and low-value content trimming. Across these settings, the benchmark contains 101 visually driven cases, including 831 raw videos with 1,476 highlight annotations. It also includes 84 speech-driven cases, each consisting of an ASR transcript with annotated cleanup spans. The evaluation compares cut plans with these annotations in both settings and additionally uses a human-aligned agentic evaluator to assess rendered-cut quality in the visual setting. To apply LVLMs to automatic rough-cut, we further propose agentic workflows that guide them to produce structured cut plans from input materials. Experiments show that current LVLMs remain limited in producing expert-aligned cut plans. Our benchmark, code, and prompts are released at https://whalesong-zrs.github.io/RoughCut-Bench-Project-Page.

Hugging Face Papers · Ruisi Zhao, You Li, Sihan Chen, Aosong Cheng, Chixiang Ma, Sen Pei, Zhen Yang, Hehe Fan, Dongliang He, Yi Yang 3 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Show HN: Tool to measure log TCO savings without losing or migrating any data

I work on Chiron. We provide a “lossless consolidation” solution for log savings on high-fidelity stores like Splunk, where you don’t lose any data, your query results stay intact, and you still meet your compliance needs. On a Palo Alto firewall log sample we published, the lossless merge came to about 40%. Method and numbers: https://medium.com/@chironvision_inc/keeping-palo-alto-logs-... We built this estimator tool so that people can get a savings number on their own logs before they reach out to us. It is a binary tool with an MCP server plus a skill. Point your Claude, Cursor or similar at a local folder where your log sample is located. The tool produces a savings estimate report. It makes sure that no data leaves your network, and no data goes to your model either. The model only gets a statistical summary and metadata. Install (or grab the tarball from the same path; checksums.txt is there, and releases are Sigstore signed): curl -fsSL https://www.chironvision.com/dl/ea324478d02311d6b42f939c08b0... | bash -s -- --target claude The tool comes with a zookeeper sample bundled so you can quickly test and validate the setup. Currently supported log formats: Splunk exports, OTLP JSON, k8s container logs, plain text. The detailed README on how to use the tool is included in the download. How the tool works: it reads the logs from the provided folder or files and runs the analyzer to compute the lossless savings. The tool interacts via the MCP server with your Claude console or similar. You can run it from your Claude or Cursor console. The engine is a binary, not open source; the license is at the download path. Run the sample, then a few minutes of your own logs

  • We provide a “lossless consolidation” solution for log savings on high-fidelity stores like Splunk, where you don’t lose any data, your query results stay intact, and you still meet your compliance needs.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Ask HN: What do you think is going on with that many outages coming in a row?

Câu hỏi In just four days we have seen: * Outlook exchange server outage for almost 20 hours * FAA communication channels outage * A few US banks outage * United Airlines outage Do you think it is related to AI coding, just coincidence or we start to see, but can’t detect, AI hijacking systems everywhere?

In just four days we have seen: * Outlook exchange server outage for almost 20 hours * FAA communication channels outage * A few US banks outage * United Airlines outage Do you think it is related to AI coding, just coincidence or we start to see, but can’t detect, AI hijacking systems everywhere?

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Show HN: You can now train your own LLMs with less memory, on your consumer GPUs

*TL;DR: You can train your own 27B model on a 3090 with 24GB VRAM, though you would have to go through hundreds of billions of tokens for it to be good. I am working on conversion of existing models. Check repos for code and the paper if you are interested in more. I have a 40.4GB Qwen Flash* N*ext GGUF* Hi, I do AI research. I have been working on an optimizer that doesn't need the \~16 bytes of VRAM per weight to train a model that AdamW does. I focused on training ternary networks bcs BitNet (Ma et al.) shows they match full precision quality in large scales. The problem is BitNet also keeps the full precision weights during training which means you still need to use the same amount of memory. I got this down from 16 bytes to 1 byte or even less per weight. For you, this means, you can train your own 10-20B parameter models on your consumer GPUs. I fitted 27B model training on a single 24GB 3090 and a 23B model on a 16GB GPU. The trained model costs \~1.6 bits per weight (\~6GB at 27B), 10x smaller than 16bit, and 5x smaller than Q8. I have also been working on converting existing models and training them. This is harder and I am still working on it but I have some promising results. I quantized and trained a Qwen3.8 27B though the file I published below has adapters which are about 2GBs, pushing the file size to 8GB. I am also making a no-adapter version right now. It scores GSM8K 0.895, the same as full precision, HumanEval 0.84 vs 0.93 and runs on a 12 GB GPU at \~2*3* tok/s on my crippled 12GB laptop 4080, though this number can heavily be improved with custom kernels, engineering, etc. I will leave it to you guys to help me with that too. I also quantized and trained Qwen Flash Next down to 40.4 GBs including its n-gram table (at 8GB). It largely retains the skills of the network (scores equivalent on GSM8K benchmark, a little behind on HumanEval, I unfortunately don't have compute for the larger benchmarks such as TerminalBench2.1 but perhaps some of you can help me out) but forgets facts. It mostly works fine if you give it web search as far as my experience goes but this is not strictly measured. I am still doing work on this and I will keep working on it, perhaps even for many years. The code is open for non-commercial use. Check out: The core repo for training: https://github.com/oguzeldereli/ttf-ternary The repo for conversion: https://github.com/oguzeldereli/ttf-conversion Repo for running models trained by ttf: https://github.com/oguzeldereli/ttf-home GGUF for a converted Qwen 3.8 27B with adapters (around 8 GB): https://huggingface.co/oguzeldereli/ttf-Qwen3.8-27B-GGUF GGUF for a converted Qwen 3.8 Flash Next (around 40.4GB): https://huggingface.co/oguzeldereli/ttf-Qwen3.8-Flash-Next-G... GGUF for a converted Qwen 3.8 27B without adapters (around 6 GB): Currently in conversion and will drop soon. I will convert a Qwen 4 27B instance when it drops. Full paper: https://github.com/oguzeldereli/ttf-ternary/blob/main/paper/... Also, I would appreciate github stars if you like the project.

  • This is harder and I am still working on it but I have some promising results.
  • 93 and runs on a 12 GB GPU at \~2*3* tok/s on my crippled 12GB laptop 4080, though this number can heavily be improved with custom kernels, engineering, etc.
HN: open weights · gladia-solaria 5 lượt bình chọn 0 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Ask HN: Anyone else blurting?

Câu hỏi Is anyone else finding this?

Like most, I was stoked on LLMs when they became available. Particularly on the earlier free tiers where the slightly questionable answers were offset by generous tokens. Now, I'm finding that having to use this tech is becoming much less fun. The constant corrections, the constant prompting for greater brevity, the grating certainty and sycophancy. Others probably are forgiving considering the net benefit. But for me it honestly now feels like a chore, which is reflected in how my frustration is now appearing in how I communicate to it now. Instead of articulating and making my prompts precise, I'm just straight-out blurting. Typos, missed words, poor semantic structure - I guess I maybe at some point reasoned that "it will get the picture, so why waste effort being articulate?" Or maybe I'm just anticipating that it is going to get it 15 percent wrong, so why even try and be articulate. Just blurt and steer. Blurt and steer. Is anyone else finding this? I actually sometimes worry about it. I feel it's not just the fact that I'm offloading my cognition to it that is making me dumber, it is the way I'm communicating with it. They say the act of writing is tied to the act of thinking, and my prompts are basicslly the typographic equivalent of grunting right now.

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Thinking Harder Doesn't Lead to Better Choices: Reasoning Effort Increases the Confidence but Not the Correctness of LLM Shopping Agents

Câu hỏi To test this belief, we ask two questions about such agents: does asking a model to think harder make it choose better products, and does the confidence it reports tell us the likelihood it may be wrong?

Large language models (LLMs) are increasingly deployed as autonomous shopping agents that process multimodal product listings (consisting of visual and text data) and recommend purchases on behalf of consumers. Naturally, one might assume that models with more thinking power would make higher-quality purchasing decisions. To test this belief, we ask two questions about such agents: does asking a model to think harder make it choose better products, and does the confidence it reports tell us the likelihood it may be wrong? We benchmark ten frontier models on three product-choice benchmarks built from WebShop, Amazon's SQID and Wayfair's WANDS platforms to see if LLMs with more thinking power actually make better-quality product purchase decisions. Across all three benchmarks, we found that increasing the thinking effort does not increase decision quality in a statistically meaningful way. The same change in thinking effort raises stated confidence at a statistically significant level, so extra thinking buys confidence rather than correctness. We also measure the calibration of these models (how well their reported confidence percentage matches their true correctness percentage) in these tasks and show that most models are not well-calibrated at every effort level. We also find that errors are strongly shared: when one model errs in a task, several others tend to err on that same task. Furthermore, on tasks where several models all err, a large number of them pick the same wrong product.

Hugging Face Papers · Adi Krishnamoorthy, X. Y. Han 1 lượt bình chọn 1 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Ask HN: How do you reach the flow with vibe coding?

Câu hỏi Ask HN: How do you reach the flow with vibe coding?

When I used to code without AI (probably not even one year ago), I sometimes reached a state of "flow" and it truly was a great feeling. Nowadays, I rarely code from scratch and most of my development (at work, at least) comes through AI. I get it that I can still code "the old way" in my free time, but I'm more curious to understand if people in a similar situation can reach the same state by vibe/AI coding, and how.

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Ask HN: Anyone else feeling burnout from agentic coding?

Câu hỏi LLMs are replacing a lot of our interactions - I’m inundated with code reviews and slop documents - I barely understand PRs that I author some days as the pressure is on to ship fast Anyone else?

I’ve never felt burnt out ever before in 10+ years of work but lately I’ve been struggling at work. Since agentic coding: - I talk to people on my team much less. LLMs are replacing a lot of our interactions - I’m inundated with code reviews and slop documents - I barely understand PRs that I author some days as the pressure is on to ship fast Anyone else? Any advice?

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Claude Is Always Right

Câu hỏi "Ok we were going to meet at 8 - did I mention it was Friday?

I think I can say, at least for me, why Claude's writing is so freaking awful, even when it's completely correct. Claude writes technical docs that are true in the same sense that it is true my best friend is a collection of specialized tissues known as "organs" and blah blah blah - and nothing else. Don't ever forget the trailing negation, it is omitted never. Another take: Claude writes like your best friend describing a great first date. "Ok we were going to meet at 8 - did I mention it was Friday? Right, 8 oclock at the Italian place and no place else. During dinner he accidentally spit out some food, but I thought i was funny! That's because we were laughing so hard after we discovered we were both from Omaha. That was before dinner, during drinks, when we couldn't get rid of Alfred. His place really needs a deep cleaning!" Perhaps with all of the hope for an LLM-authored movie, we try for something non-linear?

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Show HN: Try Vibe Robotics! (IRL or MuJoCo SIM)

Hey HN, I'm Salman, a research engineer at GDM Robotics on the team that built this robotics plugin. Whether you're a robotics founder, data company, making RL environments, or just getting FOMO seeing all the cool robotics demos on X, you can just plug in a robot and start prompting away. The plugin we've written up includes skills that steer the model toward following best practices. The benefits of using the plugin are: - fewer tokens (and therefore thinking time) needed to do bring-up the robot and to reach a successful attempt on your task - higher fidelity real-to-sim-to-real when creating MuJoCo assets and environments - leveraging strong motion planning, visual servoing, and collision avoidance We have tested on SO-101, YAM, and Franka Duo, but you don't even need real arms to test it out. You can just create cool environments in sim and then prompt the model to control a simulated robot in the environment! (Just make sure to tell it not to use privileged sim state if you want to actually evaluate how good the perception and model intelligence is). Just point your agent to the repo, good luck, and happy tinkering! Let me know if you have any questions! Will answer anything here or feel free to post on the GitHub repo directly.

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Ask HN: Where are you on the spectrum of AI/LLM adoption?

Câu hỏi Ask HN: Where are you on the spectrum of AI/LLM adoption?

- Personal user (not using in work) - Auto-complete only - Adopted agents, write specs + reviews code - Adopted agents, write specs, no code review - Adopted agents, no specs, no code review - Moved back from agents to Auto-complete only - Moved away from LLMs altogether

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Ask HN: Who's really winning with AI-Coding?

Câu hỏi My question is who is really winning here?

I've noticed the same pattern across three different Claude-coded projects now. Initial velocity is strong, but starts to fade until returns are negative. It seems a combination of changes in model behavior and increasing code base size—emphasis on the latter. The larger the code base and documented history/progress (e.g. through updates to completed work and instructions, as in .md files), the more drift starts to set in. At a certain point, you get collapse. This is where it can't maintain consistency within a single unit of work. It can manifest, for example, as adding a similar element to a view, but styling it in a completely different manner than existing elements. It literally feels random. It can also manifest as incoherent or poorly implemented logic and behavior. I've followed different advice, restructured projects, entered plan mode, added deterministic checks, styles, rules, etc. But, it feels like whack-a-mole and it's unclear that it actually adheres to these rules in the first place. Then, there's the question of it moving the problem. In response to my complaints of inconsistent UI generation, it recommended a style check to run after its UI changes. Of course, that begs the question of why it doesn't just apply those rules in the initial generation. But, I am at a place now where it spins up AI agents to do a full UI/UX review—Playwright and all. Yet, it's still drifting and not catching it. This is consistent with its unit tests—on my latest project it has now generated a suite of hundreds of somewhat time consuming tests that it constantly runs pre-commit, and not a single test has ever failed. Meanwhile, there have been a number of bugs. When questioned, it said it had been testing the "wrong layer" and had essentially generated self-confirming tests. My question is who is really winning here? When querying Google, it points to so many claims of success and approaches that supposedly contribute to that success. But, when you break down the underlying source for those claims, it appears to be largely Reddit. Yet, pressing it further reveals that virtually every actual study shows either little to no productivity gains or actual regression—particularly where experienced engineers are concerned. And some of those studies debunk the self-reported claims of exactly the kind you see on Reddit, etc. So, who here has developed a quality production application of any complexity using AI coding? If anyone, which harness/model, and what techniques did you use? Also, can you point us to your repo?

HN: Claude · unclebucknasty 5 lượt bình chọn 6 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

I open-sourced an AI loop for RISC-V CPU design months prior to OpenAI

Câu hỏi com/FeSens/auto-arch-tournament) to explore a question I’m still curious about: how well can different AI models improve a RISC-V CPU?

Back in April, I built auto-arch-tournament (https://github.com/FeSens/auto-arch-tournament) to explore a question I’m still curious about: how well can different AI models improve a RISC-V CPU? I’ve been tracking the results on HWEBench(https://hwebench.com/v2). It’s fun to see OpenAI exploring a similar question in today’s chip design post(https://openai.com/index/advancing-chip-design-with-our-models/). If you’d like to try an experiment yourself, you can run the research loop from the repo.

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Show HN: Can an AI Agent Bypass Your Website Security?

Câu hỏi Show HN: Can an AI Agent Bypass Your Website Security?

As part of our red-teaming efforts, we recently configured GPT-6 Astra to test how easily an AI agent can bypass existing bot detection and human verification systems. We built a tool that lets you run the same test on your own website: https://assess.poh.org Enter your website URL, and we'll launch an AI agent to test your workflows, record its actions, and show how we evaluate the session. Below are three recorded tests. In the recorded tests below, agents were able to complete CAPTCHAs, bypass existing bot detection mechanisms, create and verify accounts, access gated resources, and carry out end-to-end workflows intended for humans. Hugging Face: https://assess.poh.org/test/2d75a3b713e2719e87a4b85c6b3aff70 (bypassed hCAPTCHA) Cloudflare: https://assess.poh.org/test/0eb218ca6e2d7be993b689dd69a1cb20 (bypassed Turnstile) YouGov: https://assess.poh.org/test/144f6d014fbf006b61ec6bc32ee8c3f4 (bypassed reCAPTCHA) AI agents are ushering in a new age of online security, and we’re excited to hear your thoughts!

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Show HN: Harness for building/cloning SaaS using open-weight models

Built a codex like harness for non-tech builders to easily clone B2B SaaS using open-weight models. We figured companies are still charging for building software. Software, we think is commoditized. Tokens are cheap, and intelligence is open. What's the point of billing tokens. We want to take the power back from big labs and democratize the power of SoTA LLMs for every builder out there. 1M output tokens shouldn't cost 10s of dollars. Lots of work to be done but we built a harness that can use open weight models from open router to build SaaS clones. Give it a shot and lmk what you guys think. We are giving away a generous free tier of open weight model credits along with a lot of SaaS templates that you can start with.

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Ask HN: When does your app need its own UI vs. using it via MCP?

Câu hỏi If Claude/ChatGPT is where the bulk of work is done in the future (perhaps analogous to a general purpose computer) then how should app developers build in this world?

I'd love to hear HN user's views on when they think an app their building needs a UI vs just supplies data via MCP to Claude/ChatGPT. If Claude/ChatGPT is where the bulk of work is done in the future (perhaps analogous to a general purpose computer) then how should app developers build in this world?

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Show HN: PocketOps for SigNoz: The Open Source Datadog Alternative

We recently moved our observability stack from DataDog to a self-hosted SigNoz instance to retain Canadian data residency. It's been incredible thus far, big cost savings and a more targeted UI. One downside was that the web UI for SigNoz isn't responsive at all. I ended up building this mobile app. Obviously I'm biased but I think it's much better than the DataDog app, making it easy to keep an eye on things and do light debugging from my phone. I'm not affiliated with SigNoz in any way, but their platform deserves a ton of props

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Tell HN: /tmp is hogging your RAM on Ubuntu 26

As a Ubuntu user, it was just brought to my attention that /tmp is now mounted on tmpfs (aka your RAM). Coding agents, like Claude Code, use /tmp as a scratchpad for files, model weights, etc. Therefore, if you have RAM constraints, consider setting CLAUDE_CODE_TMPDIR to e.g. ~/.claude/tmp

Nguyên văn nguồn
Chi tiết nguồn
61–80 / 680
Trang 4 / 34