Hacker News: lootsauce

New comment by lootsauce in "Indirect Prompt Injection Attacks Against LLM Assistants"

lootsauce — Wed, 03 Sep 2025 20:02:54 +0000

The recent DEFCON talk for the referenced paper https://www.youtube.com/live/pleLhJRW9Fw

The risks in LLM powered systems seems like an opened pandora's box the more I look into mitigating it.

Indirect Prompt Injection Attacks Against LLM Assistants

lootsauce — Wed, 03 Sep 2025 19:20:06 +0000

Article URL: https://www.schneier.com/blog/archives/2025/09/indirect-prompt-injection-attacks-against-llm-assistants.html

Comments URL: https://news.ycombinator.com/item?id=45119502

Points: 6

# Comments: 2

New comment by lootsauce in "See how a dollar would have grown over the past 94 years [pdf]"

lootsauce — Fri, 06 Jun 2025 16:59:15 +0000

now do purchasing power of the dollar

https://fred.stlouisfed.org/series/CUUR0000SA0R

New comment by lootsauce in "[dead]"

lootsauce — Wed, 19 Feb 2025 17:42:35 +0000

I totally agree with the sentiments in the video that the current state of agent architectures feels not there yet and we need an interoperable composable standard.

Rebos: repeatability for any Linux distribution

lootsauce — Tue, 13 Aug 2024 02:56:32 +0000

Article URL: https://gitlab.com/Oglo12/rebos

Comments URL: https://news.ycombinator.com/item?id=41231748

Points: 1

# Comments: 0

New comment by lootsauce in "Show HN: Zerox – Document OCR with GPT-mini"

lootsauce — Wed, 24 Jul 2024 02:43:27 +0000

In my own experiments I have had major failures where much of the text is fabricated by the LLM to the point where I just find it hard to trust even with great prompt engineering. What I have been very impressed with is it’s ability to take medium quality ocr from acrobat with poor formatting, lots of errors and punctuation problems and render 100% accurate and properly formatted output by simply asking it to correct the ocr output. This approach using traditional cheap ocr for grounding might be a really robust and cheap option.

New comment by lootsauce in "What I think about when I edit (2019)"

lootsauce — Sat, 06 Apr 2024 10:16:09 +0000

My brain must be warped by using LLMs because I could not help but think that it would be an amazing prompt, complete with multiple examples of each principal. Perhaps good prompting is good writing?

New comment by lootsauce in "Money bubble"

lootsauce — Thu, 29 Feb 2024 19:48:26 +0000

Logical yet, how overvalued did Tesla get?

New comment by lootsauce in "100 Years Ago, IBM Was Born"

lootsauce — Wed, 14 Feb 2024 12:09:42 +0000

To the contrary does it not imply any well organized group should have an easier time of accomplishing their aims while the rest are mired in the chaos?

New comment by lootsauce in "What the Goddamn Hell Is Going on in the Tech Industry?"

lootsauce — Wed, 01 Nov 2023 12:11:52 +0000

You think the waste is high in tech saving a few mill here and there with db settings an ec2 optimizations? Imagine all the low sophistication marketing managers and CDOs, CMOS literally with 10 to 100+ million budgets much of which is completely wasted. Ten years in this space I have 2nd hand knowledge of specific cases that make your face melt.

New comment by lootsauce in "Approximate Nearest Neighbor Oh Yeah (Annoy)"

lootsauce — Tue, 31 Oct 2023 02:39:14 +0000

Found code for it https://github.com/microsoft/DiskANN

New comment by lootsauce in "Approximate Nearest Neighbor Oh Yeah (Annoy)"

lootsauce — Tue, 31 Oct 2023 01:36:39 +0000

This popped up in the new feed yesterday and seems significant in the ANN space https://news.ycombinator.com/item?id=38054537

New comment by lootsauce in "DiskANN: Fast Accurate Billion-Point Nearest Neighbor Search on a Single Node [pdf]"

lootsauce — Sun, 29 Oct 2023 01:47:33 +0000

This looks to be a big deal for vector database hosting costs. The pattern of moving away from hot ram towards algorithms that efficiently leverage ssd storage is a winner every time.

New comment by lootsauce in "I'm fed up with it, so I'm writing a browser"

lootsauce — Fri, 22 Sep 2023 21:26:38 +0000

A thought, write a very constrained sane rendering/layout engine as a compile target for an llm to convert html/css to rather than support all css or any of it for that matter. Could there be some kind of llvm for html/css?

New comment by lootsauce in "The new science of meditation"

lootsauce — Tue, 22 Aug 2023 19:54:54 +0000

I picked up the Tao Te Ching in middle school and would reread it frequently. The personal insights would evolve as I grew up and faced different issues in my life. I found an abiding peace in those personal insights and they have held with me since. One of those insights is something like having instead of seeking as an identity. I don’t feel that I want to seek to find myself or some great answer, it’s already there, I already have it. YMMV bits a wonderful ancient text. I recommend the S. Mitchel translation.

I could never get into the sitting and breathing kind of meditation, but I do find different times I am able to practice very meditative activities such as walking in the woods, aimless wandering, puttering about with no intention just taking in the world. I used to do zen inspired landscape photography and that whole process felt very meditative to me.

New comment by lootsauce in "Supreme Court Opinion on Compelled Speech and Protected Class [pdf]"

lootsauce — Sat, 01 Jul 2023 03:24:04 +0000

“Consistent with the First Amendment, the Nation’s answer is tolerance, not coercion. The First Amendment envisions the United States as a rich and complex place where all persons are free to think and speak as they wish, not as the government demands.”

New comment by lootsauce in "The Bookmarking Data Model Is Wrong for Highlighting"

lootsauce — Wed, 17 May 2023 12:19:08 +0000

Part of the challenge of firehoses of links or hilights being less valuable to others, is that they are the residue of an particular persons thought process and that thought process was clear to them but obviously not us. A description of the list might help, but there is something more. We have these thoughts for a reason and my motivations will be very particular to me and you probably won’t care so much about them. So yes knowledge can be social but it is first hyper personal and then social if you can find people with similar motivations and thoughts. But to most, that residue of thought will just feel like a pile of stuff they don’t connect with. It takes a lot of effort to recast it for a broader audience and that’s more blogging than note taking. Finding your cohort is an altogether different project as well, you need to find an existing nexus of the right people if it exists or become the maintainer of one and attract those people, this is more like being an influencer or party planner. The idea that any note taking platform will be the nexus is unrealistic, so I think we will always have these hyper personal digital gardens and out of them we will harvest things to share with others, sometimes you put out a stand of free carrots at the end of the driveway, sometimes you bring a basket of raspberries to a friend you know who loves them, most likely you take it all to the farmers market, lastly maybe after much planning and work you host a garden party but it won’t be in the actual garden beds but in the adjacent lawn. :-)

New comment by lootsauce in "You can't tell people anything (2004)"

lootsauce — Fri, 24 Mar 2023 17:28:06 +0000

I always say it is non-trivial to achieve a meeting of the minds. We often act adversarially to new ideas to challenge them as they in turn challenge our conceptions. Often this looks like arguing and many people ain’t got time for that, it can be emotional, you’re not listening, just give it 5 mins, no you’re wrong, etc. This is not a failing of individuals so much as the impedance to aligning two different cognitive patterns. One does not simply walk into a new cognitive state you have to do battle with your own and another’s conceptions and in the face of imperfect communication.

New comment by lootsauce in "Introducing ChatGPT and Whisper APIs"

lootsauce — Thu, 02 Mar 2023 15:58:47 +0000

Wow this is the best content I have seen on prompt engineering, it feels much more like computer science than the usual prompt engineering recipes. This formalism of analytic augmentation and higher order augmentations seems like it is ripe for some kind of library or dsl to build and evaluate and monitor these processes. But somehow I feel like these capabilities will become so standard that they will be incorporated into the official apis.

[edit] Just noticed it looks like you are working on just that. I will keep an eye on this! https://github.com/williamcotton/transynthetical-engine

New comment by lootsauce in "Against Method"

lootsauce — Thu, 22 Dec 2022 20:08:35 +0000

YES! "Science is ultimately cultural and there's no way around that." I think that many people recoil at this idea and just don't want to believe this, but it is the truth. To those individuals I would say do not confuse the map with the territory or perhaps they map making and surveying they also are not the territory.