Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Claude Code can debug low-level cryptography

words.filippo.io

Post reply on HN
Print view View original

211–217 of 217 posts

Re: Claude Code can debug low-level cryptography

#211

Post by Razengan »

theshrike79 wrote: post #164
Razengan wrote: post #139

Earlier quoted context omitted.

I did ask the AI first, about some things that I already knew how to do. It gave me horribly inefficient or long-winded ways of doing it. In the time it took for "prompt tuning" I could have just written the damn code myself. It decreased the confidence for anything else it suggested about things I didn't already know about. Claude still sometimes insists that iOS 26 isn't out yet. sigh.. I suppose I just have to tre…

You can write the "prompt tuning" down in AGENTS.md and then you only need to do it once. This is why you need to keep working with different ones to get the feeling what they're good at and how you can steer them closer to your style and preferences without having to reiterate from scratch every time. I personally have a git submodule built specifically for shared instructions like that, it contains the assumptions…

> You can write the "prompt tuning" down in AGENTS.md and then you only need to do it once.

Yeah, I just mean: I know how to "fix" the AI for things that I already know about.

But how would I know if it's wrong or right about the stuff I DON"T know?? I'd have to go Google shit anyway to verify it.

This is me asking ChatGPT 5 about ChatGPT 5: https://i.imgur.com/aT8C3qs.png

Asking about Nintendo Switch 2: https://i.imgur.com/OqmB9jG.png

Imagine if AI was somebody's first stop for asking about those things. They'd be led to believe they weren't out when they in fact were!

Re: Claude Code can debug low-level cryptography

#212

Post by simonw »

Razengan wrote: post #210
simonw wrote: post #191

Earlier quoted context omitted.

Unfortunately LLMs mostly suck at Swift and SwiftUI from what I've heard - they still change pretty often and as a result there aren't enough fresh examples in the training data. As primarily a Python/JavaScript programmer I don't have that problem!

They're terrible at anything new, including knowing about THEMSELVES and their latest versions. This is me asking ChatGPT 5 about ChatGPT 5: https://i.imgur.com/aT8C3qs.png Asking about Nintendo Switch 2: https://i.imgur.com/OqmB9jG.png This could be solved and LLMs could be a lot more useful if they could be a wrapper around live web search: Just search for this shit, scrap the top few results, and summarize the inf…

Claude Code has a neat fix for that - it knows to look at its own documentation if you ask it questions about itself: https://simonwillison.net/2025/Oct/24/claude-code-docs-map/

I've had great results from ChatGPT running the "GPT-5 Thinking" model since that almost always opts to run a search before it attempts to answer a question.

Here's what I got from that for your Switch 2 question: https://chatgpt.com/share/69089028-db8c-8006-b238-1d6946e791...

Screenshot of the searches it ran here: https://gist.github.com/simonw/048ffb895dd6b94419f0b4e066143...

Re: Claude Code can debug low-level cryptography

#213

Post by Razengan »

simonw wrote: post #212
Razengan wrote: post #210

Earlier quoted context omitted.

They're terrible at anything new, including knowing about THEMSELVES and their latest versions. This is me asking ChatGPT 5 about ChatGPT 5: https://i.imgur.com/aT8C3qs.png Asking about Nintendo Switch 2: https://i.imgur.com/OqmB9jG.png This could be solved and LLMs could be a lot more useful if they could be a wrapper around live web search: Just search for this shit, scrap the top few results, and summarize the inf…

Claude Code has a neat fix for that - it knows to look at its own documentation if you ask it questions about itself: https://simonwillison.net/2025/Oct/24/claude-code-docs-map/ I've had great results from ChatGPT running the "GPT-5 Thinking" model since that almost always opts to run a search before it attempts to answer a question. Here's what I got from that for your Switch 2 question: https://chatgpt.com/share/69…

A month ago when I asked Claude (on the website) about its privacy options and stuff, it always pointed me to the Antrhopic website to look it up myself.

Another annoying example: I thought Google's Gemini would be search-first since, well, they're Google.

I asked Gemini to search for Airbnb rooms in an area and give me a summarized list.

It told me it can't and I could do it myself.

I told it again.

Again it told me it can't, but here's how I could do it myself.

I told it it sucks and that ChatGPT etc. can do it for me.

Then it went and I don't know, scrapped Airbnb or used a previous search it must have had, to pull up rooms with an Airbnb link to each.

This could actually be THE absolute killer app for a lot of people, if AI could plan your trip from a single sentence: "I'm free next week. I'd like to go to A, B, or C for a couple days. What's a cheap flight and a room within this budget near X area?" and if it could also go and make a booking through your accounts it would be orgasmic. Finally we would have what people in the 1960s thought computers would be doing in 2000 :')

But as it is, in their current state you have to wade through quite a bit of dumbassery.

Re: Claude Code can debug low-level cryptography

#214

Post by jerf »

This is one of the things I've mentioned before, I think it's just hidden a bit and hard to see, but this is basically the LLM doing style transfer, which they're really good at. There was a specification for the code (which looks like it was already trained into the LLM since it didn't have to go fetch it but it also had intimate knowledge of), there was an implementation, and it's really good at extracting out the style difference between code and spec. Anything that looks like style transfer is a good use for LLMs.

As another example, I think things like "write unit tests for this code" are usually similar sort of style transfer as well, based on how it writes the tests. It definitely has a good idea as to how to sort of ensure that all the functionality gets tested, I find it is less likely to produce "creative" ways that bugs may come out, but hey, it's a good start.

This isn't a criticism, it's intended to be a further exploration and understanding of when these tools can be better than you might intuitively think.

Re: Claude Code can debug low-level cryptography

#215

Post by theshrike79 »

Razengan wrote: post #211
theshrike79 wrote: post #164

Earlier quoted context omitted.

You can write the "prompt tuning" down in AGENTS.md and then you only need to do it once. This is why you need to keep working with different ones to get the feeling what they're good at and how you can steer them closer to your style and preferences without having to reiterate from scratch every time. I personally have a git submodule built specifically for shared instructions like that, it contains the assumptions…

> You can write the "prompt tuning" down in AGENTS.md and then you only need to do it once. Yeah, I just mean: I know how to "fix" the AI for things that I already know about. But how would I know if it's wrong or right about the stuff I DON"T know?? I'd have to go Google shit anyway to verify it. This is me asking ChatGPT 5 about ChatGPT 5: https://i.imgur.com/aT8C3qs.png Asking about Nintendo Switch 2: https://i.im…

There's your problem right there.

Don't use it as a knowledge machine, use it as a tool.

Agentic LLMs are the ones that work. The ones that "use tools in a loop to achieve a goal"[0]. I just asked Claude to "add a release action that releases the project as a binary for every supported Go platform" to one of my Github projects. I can see it worked because the binaries appeared as a release. It didn't "hallucinate" anything nor was it a "stohastic parrot". It applied a well known pattern to a situation perfectly. (OK, it didn't use a build matrix, but that's jsut me nitpicking)

In your cases the LLM should've seen that you're asking about current events or news and used a tool that fetches information about it. Now it just defaulted to whatever built-in training data was in its context and failed spectacularly

AIs have a branding issue, because AI != AI which isn't AI. There are so many flavours that it's hard to figure out what people are talking about when they say "AI slop is crap" when I can see every day how "AI" makes my life easier by automating away the mundane crap.

[0] https://simonwillison.net/2025/Sep/18/agents/

Re: Claude Code can debug low-level cryptography

#216

Post by braebo »

cmrdporcupine wrote: post #97
reachableceo wrote: post #85

Earlier quoted context omitted.

Have you tried to ask the agents to work with you in the way you want? I’ve found that using some high level direction / language and sharing my wants / preferences for workflow and interaction works very well. I don’t think that you can find an off the shelf system todo what you want. I think you have to customize it to your own needs as you go. Kind of like how you customize emacs as it’s running to your desires. I…

I have, but the problem is in part the tool itself and the way it works. It's just not written with an interactive prompting style in mind. CC is like "Accept/Ask For Changes/Reject" for often big giant diffs, and it's like... no, the UI should be: here's an editor let's work on this together, oh I see what you did there, etc...

This is why I still prefer Cursors workflow to the CLIs!

Re: Claude Code can debug low-level cryptography

#217

Post by AbstractH24 »

monocasa wrote: post #207
AbstractH24 wrote: post #188

Earlier quoted context omitted.

So we can all only succeed if we know how CPUs handle individual instructions?

There's generally a pretty quick falloff of how much help knowledge of each layer under you generally provides as you go deeper. That being said, if you're writing in C, having a pretty good idea of how a cpu generally executes instructions is pretty key to success I'd say.

Agreed, also depends on the scale you are working at.

If you are a tiny startup, the marginal gains from these optimizations matter a lot less than if you are Netflix.

Post reply on HN