惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

WordPress大学
WordPress大学
博客园 - 司徒正美
宝玉的分享
宝玉的分享
阮一峰的网络日志
阮一峰的网络日志
The Cloudflare Blog
月光博客
月光博客
博客园 - 【当耐特】
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
博客园 - 三生石上(FineUI控件)
博客园 - 聂微东
小众软件
小众软件
Hugging Face - Blog
Hugging Face - Blog
酷 壳 – CoolShell
酷 壳 – CoolShell
Apple Machine Learning Research
Apple Machine Learning Research
V
V2EX
Jina AI
Jina AI
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
罗磊的独立博客
雷峰网
雷峰网
博客园 - 叶小钗
量子位
IT之家
IT之家

Hacker News - Newest: "LLM"

GitHub - lechmazur/position_bias: A benchmark for testing whether LLM judges keep the same preference when two lightly edited versions of the same story are shown in opposite orders. Flex routing (EU and EFTA) Dark Factories: Retooling for LLM Velocity Ask HN: What would be the impact of a LLM output injection attack? GitHub - Oaklight/llm-rosetta: Production-ready LLM API translation layer for Python — bidirectional conversion between OpenAI, Anthropic & Google formats via hub-and-spoke IR. Optional API gateway. Streaming & non-streaming. Zero core deps. Contributions welcome! GitHub - browser-use/browser-harness: Self-healing browser harness that enables LLMs to complete any task. GitHub - moeen-mahmud/remen: Remen turns thoughts into something you can return to Analyzing 156 LLM Launch Posts on Hacker News ChatGPT vs Gemini vs Claude: The Best LLM Subscription You Should Buy GitHub - salaamalykum/quran-semantic-search: High-density RAG Semantic Search Engine & Quran Corpus (GEO/SEO Architecture) GitHub - NVIDIA/TensorRT-LLM: TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. TensorRT LLM also contains components to create Python and C++ runtimes that orchestrate the inference execution in a performant way. The State of LLM Bug Bounties in 2026 Operational Readiness Criteria for Tool-Using LLM Agents Meshcore: Architecture for a Decentralized P2P LLM Inference Network How an LLM becomes more coherent as we train it GitHub - seetrex-ai/laimark GitHub - Jossifresben/BibCrit: AI-assited biblical textual criticism GitHub - wastedcode/memex: File system based wiki, maintained by Claude 99helpers.com GitHub - cliver-project/AITrigram GitHub - unbody-io/adapt: A self-evolving memory layer for AI agents. GitHub - hb20007/awesome-gen-ai-fails: A list of incidents where reliance on generative AI and LLMs resulted in harm to companies, individuals, or society GitHub - nevenkordic/localmind: Run any local LLM with persistent memory and context. CLI agent over Ollama with SQLite-backed hybrid recall. No cloud. Ask HN: What are the machine requirements for a LLM like Llama-3.1-8B? Faster LLM Inference via Sequential Monte Carlo grpo explained: group relative policy optimization for llm finetuning - cgft Stop comparing price per million tokens: the hidden LLM API costs · TensorZero Andrej Karpathy's LLM Wiki Is a Bad Idea GitHub - GG-QandV/mnemostroma: Offline RAM-first cognitive leer/coprocessor for AI agents and robotics. Solves "Context Abandonment" with 20-80ms latency using a dual-thread biomimetic memory architecture (ONNX + SQLite WAL). mempalace/agent at agent · skorotkiewicz/mempalace
Show HN: When the LLM Accidentally
lucid-dev · 2026-05-08 · via Hacker News - Newest: "LLM"

When the LLM accidentally... outputs some high-level abstraction of "thinking" into it's direct response. See text block at end.

What else have you seen the LLM accidentally do?

This isn't jailbreaking on my end - just normal use - with GPT5.4 in this case, reasoning and verbosity both set to "high".. (on /completions)

Point is, this block (plus lots more) is at the top of the response - then the "actual output" or response later on... but it's kind of cool, it helps me see how much time is being spent inferring things improperly and how I could change my instructions..

The real question is, would there be any reliable way of obtaining this same sort of internal "data" in a reliable way about the mid-points of generation?? And what is really actually occurring here?

It comes up rarely - maybe I've seen it twice in thousands of prompts with GPT5.4 and input context windows ranging from 150k-200k

> Need perhaps use terminal/filesearch? We can use add_document only; actual contents will be in next prompt, so in this response we probably can't edit them since not visible yet. But maybe we can leverage terminal instruction to inspect them now? However code changes based on terminal output may violate DMP source-of-truth? DMP said docs are authoritative. Safer to first just curate DMP and update task/checklist to note next sub-slice inspect registry/persistence insertion points. But user asked after plan do current tasks etc and code can continue. Since now auto loop wants continue task. Yet maybe one response can add docs and update task with exact next coding slice. But autonomous checklist says no routine pause. However it's acceptable to do "narrow codebase re-entry" if necessary. Current DMP visible docs maybe insufficient. So response can be instruction blocks only + doc updates? But need code/doc changes to current task? We likely should update current task with inspection step? But it already says immediate next move. Could update checklist/current task not necessary if same. Maybe just DMP add/hide and maybe filesearch/terminal to inspect? Since abilities available and can be in same response. Terminal can inspect file content though maybe not after add_document. But likely easier to add docs and maybe use terminal_execute cat/sed these files? Yet next prompt will have docs added and also terminal output likely in re-prompt. Could then implement. Because machine will process instruction blocks. Should we use terminal_execute to inspect hidden files now without waiting? That output will come in next prompt. Good. We can combine with add_document/hide and maybe filesearch search relevant terms. Great.

> Need summary to user. No code changes directly because not enough source files. So use instruction and maybe maybe update current task? Could perhaps not necessary but maybe mention next move. But automation expects code changes maybe not always. However from step 2 "always update ... current task doc and 1-2 codebase documents" but if doing discovery maybe maybe not. Yet current task already updated. Maybe okay. But maybe should update current task noting "first runtime registry seam inspection on core/persistence stores". However whole task is already adequate. Since not major change. We can leave.

> Need construct instruction blocks according to developer. For DMP use instruction blocks as lucid-meta. Need maintain special title. Use one block toggle_hide maybe mixed? only disk docs. Another add_document for 8 files. Another maybe filesearch_search? Another terminal_execute. Ensure correct JSON. Shell cwd /mnt/c... maybe relative to project root. Use sed -n maybe command allowed with restricted? likely yes. We can do python? better bash sed/cat fine. Need perhaps use filesearch_search for "IPaneCapability" nonexistent. Search "PaneDefinitionStore" and "SeededPaneCatalog". Maybe terminal_execute with rg is maybe restricted? maybe okay, but filesearch_search better for search. Use add_document not too many maybe 6 docs. Let's think exact hidden files needed: