Tiberium

Can you show some examples of Pangram being "hot worthless garbage"? This reputation is true for most other detectors, but Pangram spends considerable effort to have an extremely low false positive rate. I think you just used some of them in 2023 and now think that they can't be made better.

If you cannot see that the linked text is LLM written, I've got bad news for you.

I only post it when I'm sure and I cross verified. And in this case specifically it's not boring, quite the opposite, it just shows that humans can be quite inconsistent. I really hope HN would extend its no LLM rule to articles/text, although this is too hard to enforce. When I look at HN, I want to read human text, I don't mind the use of LLMs for actual code and use it a lot myself. My issue is with people who generate whole articles with LLMs and present them as if they were written by a real human.

Not what? It's incredibly ironic that the creator is complaining about AI yet they used AI just to write a tiny text announcement. Apparently such a tiny snippet didn't deserve human writing.

A question: why are you writing this with an LLM? Can you not write with your own words, at least on HN?

The author is also replying with an LLM on HN: https://news.ycombinator.com/item?id=49180716

You're talking to an LLM, unfortunately

https://news.ycombinator.com/threads?id=runtime_lens

(enable showdead in your HN settings)

Indeed, I don't think that a decompilation is clearly licensable like that..

Honestly such decomp projects are the place where LLM agents help the most, I have a similar one for a 1999 game which started from the agent renaming all functions, globals in the IDA DB, adding types and function signatures. After the database was basically 100% annotated, it was exported, made compilable with the same compiler as the original game. Then after some fixes (basically all being IDA decompilation bugs) it became playable.

Having the same compiler helps, and I also have a binary matching workflow, but matching functions 100% is a huge token sink due to compiler optimizations, so I just had agents review functions one by one for differences, clean up decompiler artifacts and possible semantic bugs, and mark functions as reviewed. So the raw matching % is really low even though the game already works.

It does help that the game's native part is quite small, only 1.5k pure C functions in 700KB of code. Although with LLMs as long as you have enough usage, it's only a matter of time even for huge codebases.

As the LLM I used GPT 5.5, then 5.6 Sol, I trust GPT models the most for reverse engineering, they're very thorough.

It's nice that these people started using LLMs (mentioned in https://lwss.github.io/Kisak-Black/), although IDA MCPs are worse than CLI-based options, and I wouldn't trust Claude models that much for this work. It seems like the work started in 2025 when those models weren't good enough for that, but they absolutely are now.

Decomps are one of those repetitive, mostly non creative tasks that I think humans shouldn't spend their valuable time on, maybe only to guide or clean up. If you have an older favorite game, chances are, you can fully decompile and reimplement it with enough tokens :)

Perhaps you haven't used LLMs enough, this style is extremely common with Claude. And can you link me a Pangram 4 example where it has lots of false positives? The main reason why Pangram is supposed to be reliable in the first place is that they trade false positives for false negatives.

Come on, the LLM language is all over the article:

> The workflow I had in mind was deliberately cautious

> The code was not the difficult part.

> These requirements are not a one-time form to complete and forget. They create an ongoing responsibility around the account, the documented process, community feedback, failures, and potential reversions.

Just a few examples. And yes, Pangram 4 also flags it as 100% LLM written. I don't mind being downvoted or flagged, but I think more people should be aware of the LLM style, even if they're ok with it being used without any disclaimer. It's honestly sad that nowadays people on HN cannot recognize this style.

Fully LLM-written article, maybe that's another reason :)

Can you provide a single example of such paper that would show as AI generated on Pangram?

One of mathematicians working at OpenAI refuted those claims directly on X - https://x.com/AcerFur/status/2083656978719719601

This immediately struck me as Claud-y:

> The gap has two independent consequences, each sufficient to invalidate the claimed disproof. The first is structural.

Then the classic coding agents negation of earlier evidence, instaed of just updating to use new references, they mention that they changed old to new:

> The publicly released monolithic file ConnesRigidity.lean (37,000+ lines) does not use the names CocycleExtension, ZeroCocycle, or TwistedCocycle that appeared in the earlier modular source files (CocycleExtension.lean, ICC.lean, CrossedClosure.lean). However, the identical mathematical construction is present under different names. The following table gives the correspondence, with line numbers in the published file.

> This is the same zero-cocycle / twisted-cocycle structure identified in the earlier modular source files, confirming that the structural analysis of this note applies to the published code

And some other Claud-y stuff:

> Why both paths are closed. A successful defence would have to close both paths simultaneously

> This case illustrates a failure mode that is becoming increasingly well documented in the literature on AI-assisted formal mathematics: the gap between what a formal proof verifies and what it means. The Lean kernel certifies that a proof term inhabits a given type; it does not certify that the type faithfully encodes the intended mathematical claim. As Tao has emphasised

And afterwards I cross-verified with Pangram 4 which I trust, it marked the preamble/starting stuff as 100% AI-generated.

This "paper" itself is 100% AI-generated...

People should just try base models, or models where you have access to prefills. I guess so many who use LLMs nowadays have never actually experimented with a text completion endpoint locally/over API.

A whole Solar System for just $5000? Is Pluto included?

This comment was also auto flagged and deleted, I vouched it to specifically tell you: please write comments on HN yourself. Even this comment is LLM written.

Just a tip: AI generated comments get auto flagged on HN, so if you want to write something in comments for people to reply to, write it yourself.

I think the editor themselves misunderstood the conjecture.

UPD: The edit got reverted and there's this on the talk page now: https://en.wikipedia.org/wiki/Talk:Jacobian_conjecture#c-DaR...

UPD2: There are edit wars happening now: https://en.wikipedia.org/w/index.php?title=Jacobian_conjectu... https://en.wikipedia.org/wiki/Talk:Jacobian_conjecture#c-Sea...

> Don't post generated text or AI-edited text. HN is for conversation between humans.

Sorry, but gaslighting others that you haven't used an LLM to write your comments when you obviously have is incredibly dishonest.

Not even talking about the actual content on the website, OP - why are all of your comments here LLM-generated?

It's extremely clear you're not writing those yourself, which makes the whole thing very disingenuous if you can't engage with others in your own language.

> Some caveats since this crowd will rightly push on them

Maybe you should start by not writing your HN posts with an LLM..

To be fair, when they reset usage for everyone (instead of a banked reset), they move your reset back a week later, so:

1. Your window ends (usage reset) in 3 days

2. They reset usage for everyone

3. Your usage goes back to 100% and your current window ends in 7 days now.

API pricing itself might have extreme margins compared to the real cost. Anyway, in my testing the Pro 20x $200 plan gives you about $2200 API-equivalent weekly usage if you're only using GPT 5.6 Sol, so quite close to $9k-$10k/month API-equivalent, it's a bit inconsistent with cache costs.

It's very interesting that for Anthropic the $100 and $200 plans only differ 2x in weekly limits, the 5 hour limit differences are more severe. But for OpenAI, Pro 20x is, well, 4x of Pro 5x for only 2x cost. So, for example, 100% of weekly usage for Codex on a Plus ($20) account is just 5% of weekly usage for Codex on Pro 20x.

And you can calculate how much extra usage you can get from resets, and especially banked resets by purposefully using the whole quota and using your banked reset - they expire 30 days after they're given out, so if you don't use one, it just disappears.

I don't think it's just the docs that are LLM written.

Was there a reason to add an AI-generated image to the top of the article? :(

More details:

- https://platform.kimi.ai/docs/guide/kimi-k3-quickstart

- https://platform.kimi.ai/docs/pricing/chat-k3

1M context, pricing is $3/$15 for 1M tokens (cache $0.3), which is extremely high for a Chinese open-weight model, but if it's truly competitive with most of the current frontier and is only behind Fable/Sol, the pricing is justified.

This is 1:1 pricing of Anthropic's Sonnet series (except Sonnet 5 which is currently on discount), and very close to 5.6 Terra pricing (Terra's input is $2.5).

One thing to consider, though: reasoning efficiency matters directly for how expensive a model actually is in real use. GPT's models are extremely reasoning efficient, and some Claude models like Fable at lower effort are as well. So if Sol spends 10K reasoning tokens to do something (at $30/1M) vs Kimi K3 that spends 50K reasoning tokens, Sol would win on cost effectiveness.