frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Open in hackernews

Show HN: When the LLM Accidentally

2•lucid-dev•1h ago
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:

The room, not the words in it. My AI harness

https://blog.imjackofitall.com/posts/2026-04-28/
1•imjackofitall•1m ago•0 comments

Why I'm giving up on Facebook (2017)

http://blog.afandian.com/2017/01/why-i-am-giving-up-on-facebook/
2•chistev•1m ago•1 comments

Tesla Recalls Cybertruck Because Wheels May Fall Off

https://www.kbb.com/car-news/tesla-recalls-cybertruck-because-wheels-may-fall-off/
1•bjhess•3m ago•0 comments

Pitchfork: A devilishly good process manager for developers

https://pitchfork.en.dev/
1•thunderbong•9m ago•0 comments

AI-Native Hedge Funds Are Possible and Profitable Just Not the Next Unicorn

https://podium-finance.com/blog/ai-native-hedge-funds-and-yc
1•nlpnerd•10m ago•0 comments

Four sites. Four ways money was leaving the table

https://noemica.io/blog/four-ways-money-leaves-the-table
1•SebastianSosa•10m ago•0 comments

Auth Proxy Injection for LLMs

https://www.grepular.com/Auth_Proxy_Injection_for_LLMs
1•mike-cardwell•10m ago•0 comments

A Simple Ld_preload Tutorial

https://catonmat.net/simple-ld-preload-tutorial
1•gurjeet•14m ago•0 comments

Why the Future of AI Depends on Smaller Models

https://jeevan.life/posts/why-the-future-of-ai-depends-on-smaller-models.html
1•lroe•16m ago•0 comments

The agent principal-agent problem

https://crawshaw.io/blog/agent-principal-agent
1•bumbledraven•18m ago•0 comments

Show HN: Sfera – sliding puzzle in curved space (open-sourced)

https://github.com/bntre/Sfera
1•bntr•18m ago•0 comments

SubQ: Sub-quadratic LLM built for 12M-token reasoning

https://subq.ai/
1•anujbans•21m ago•0 comments

Camera

https://online-camera.com
2•ent101•25m ago•0 comments

KDE Plasma 6.7 to Provide a Better Experience for CPU-Based Rendering

https://www.phoronix.com/news/KDE-Plasma-Faster-WL-SHM
2•doener•27m ago•0 comments

Live Global Hantavirus Outbreak Tracker

https://hantavirusmap.com
3•N19PEDL2•27m ago•0 comments

Interviewing/Applying to YC

1•chanlee514•31m ago•0 comments

I tried the viral 'jaw release' fascial massage

https://www.thetimes.com/life-style/health-fitness/article/jaw-release-facial-massage-v7rmgqsb9
3•petethomas•32m ago•0 comments

Ask HN: Who Uses a todo.txt?

3•mikenikles•32m ago•1 comments

Why does AI like goblins and Japan so much?

https://english.elpais.com/technology/2026-05-07/why-does-ai-like-goblins-and-japan-so-much.html
4•geox•33m ago•0 comments

Dirty Frag: Universal Linux LPE

https://github.com/V4bel/dirtyfrag
1•danaris•36m ago•0 comments

Root-based Android app backup tool powered by Restic

https://github.com/hddq/restoid
1•anotherevan•40m ago•0 comments

A man who blew up a nuclear power station and disappeared

https://www.theguardian.com/world/2026/may/05/the-man-who-blew-up-a-nuclear-power-station-koeberg...
4•Michelangelo11•47m ago•0 comments

Pronouns in Perl (2009)

https://www.modernperlbooks.com/mt/2009/11/from-novice-to-adept-pronouns-in-perl
1•tosh•48m ago•0 comments

Discussion on Fsnotify Maintainer Removal

https://github.com/fsnotify/fsnotify/issues/757
1•mjyut•48m ago•1 comments

3 Women Who Fundamentally Shaped Modern Software [Reddit!!!]

https://old.reddit.com/r/compsci/comments/1t6jo79/3_women_who_fundamentally_shaped_modern_software/
1•signa11•50m ago•0 comments

Eating eggs could cut Alzheimer's risk by 27%

https://www.sciencedaily.com/releases/2026/05/260506225214.htm
3•y1n0•50m ago•2 comments

Can houseplants purify the air in your home?

https://phys.org/news/2026-05-houseplants-purify-air-home-science.html
2•y1n0•51m ago•0 comments

AI giveth and AI taketh CPU

https://stackoverflow.blog/2026/05/08/ai-giveth-and-ai-taketh-cpu/
1•y1n0•52m ago•0 comments

When a Meta CTO and an AWS ProServe Principal Converge on the Same Philosophy

https://www.outcomeops.ai/blogs/the-o16g-manifesto-validates-what-weve-been-building-since-july
2•linsys•55m ago•0 comments

The layoffs will continue till we learn to use AI

https://twitter.com/championswimmer/status/2051807284691612099
4•s4i•57m ago•0 comments