Tiberium

(Not a lawyer, not an expert, just my possibly ignorant/wrong comment)

I checked around a bit more, and this seems to directly contradict the HN title:

https://agewarden.ai/customer-agreement

> AGEWARDEN is an age estimation tool. It does not verify identity and does not guarantee 100% accuracy. Customer is responsible for determining whether AGEWARDEN satisfies the legal requirements applicable to its business and jurisdiction.

I'll be honest: I don't have experience with audio stuff/age verification, but wouldn't it be far easier to bypass this than even face (live video) verification, let alone full KYC with ID? At that point it's not that much better than a "I'm over 18" button, or am I missing something?

EDIT: The website itself partially answers my question:

> Self-declaration, the "I am over 18" checkbox, is explicitly prohibited by every major regulator in the UK, EU, and Australia.

But then:

> Facial scanning works, but it builds infrastructure that outlasts the check. A system that estimates your age today can identify you tomorrow. Platforms that rolled it out met immediate backlash and reversals. Users do not trust platforms with their face.

How would we trust your platform to not store voice fingerprints, then?

(On a side note, all descriptions all LLM-written, but that's expected in 2026)

You're right, HN stripped the deep link. There's also a funnier part below when the agent tells what harness it's running on, and other info: https://github.com/home-assistant/core/pull/173465#issuecomm...

It might not help in your case, but usually LLMs are far more likely to help if you have source code available, or if the service is running on localhost. Ideal if you have a source code where the LLM can make a lab setup and test vulnerabilities by itself. Normal GPT-5.5 with KYC is enough for this.

Probably silently rerouting?

I think what they're saying is that this prompt/jailbreak only lets Mythos discover some really easy vulnerabilities that it probably fixes from a simple "Find and fix bugs in this code" and that this can be easily done by other models like GPT-5.5. Which is very different from targeted security research.

I still also have access, so either they silently reroute Fable 5 to Opus 4.8 or hasn't actually pulled the switch yet.

Last update over a year ago, so I hope (2025) gets added to the title:

> [2025/05/26] (Step 1 completed!) We release Mixture-of-Thoughts--a curated reasoning dataset of 350k verified traces distilled from R1. The dataset spans tasks in mathematics, coding, and science, and is designed to teach language models to reason step-by-step. We also provide a recipe to train OpenR1-Distill-7B, which replicates the reasoning capabilities of deepseek-ai/DeepSeek-R1-Distill-Qwen-7B and marks the completion of step 1 in the Open R1 project.

Doesn't look like they managed to actually reproduce R1, and only stopped on Step 1 out of their 3-step plan.

I tested it myself, seems to reproduce on: 3.1.0-a1 to 3.3.16, 4.0.0-a1 / 4.0.0-a2. Fixed in 3.3.17 and in master.

Gives you auth + access to Moderation Control Panel (if the user is a moderator/admin). Does not give access to the Admin Control Panel though.

I hope GitHub changes their vibecoded badges, what does RETIRED even signify in this context? Why does the preview have to be in ominous red?

I was just pointing out how the article is clearly LLM written, probably including the interactive widgets. It's especially obvious because someone writing such an article in 2026 would at least find what the newest tokenizers are, instead of mentioning LLaMA 2/3 (!), and GPT's old tokenizer that they dropped since GPT-4o (or something close).

And, more obviously, the fact that GPT-4 is being directly named even though that model is over 3 years old by now: "Ask GPT-4, Claude, or Gemini today and they will usually answer three.".

Sorry, I just think that the article wasn't produced by a human at all.

> This isn’t because the model can’t count. It’s because it never sees the letters at all.

> The chunks aren’t characters and they aren’t words. They’re something more specific, and the specificity matters more than most people realize.

> Those numbers are real, but they hide what a token actually is.

> GPT-4’s vocabulary isn’t Claude’s. Claude’s isn’t Llama’s.

> The model never sees text. It sees a sequence of integer indices into its own private alphabet.

> So tokens aren’t “roughly like words” or “kind of like characters”. They’re the atoms of perception for one specific model, and they’re the only language that model speaks.

> The same sentence is nine tokens to GPT-4 and seven tokens to Llama 3. Not because Llama is smarter or the sentence changed, but because the two models have different vocabularies.

> That’s it. No clever scoring, no neural network.

Could people who use LLM to write articles at least prompt them to have a better style? I'm really tired of the default Claude style (a lot of Chinese models also reuse the same style)

The article comes from the "personal" experience of an LLM so it's a very trusted source!

/s

The whole "book" seems to be AI-generated, or at least very heavily AI-edited. Would at this point it not be easier to just tell developers to use their LLM of choice to achieve the same (or, likely, better) result?

Random chapter so you can judge the quality for yourself: https://microsoft.github.io/RustTraining/python-book/ch09-er...

And the non-stop bullet list slop just looks horrible: https://microsoft.github.io/RustTraining/python-book/ch01-in...

Seems like this isn't limited to the Python book though, and others have the same issues: https://github.com/microsoft/RustTraining/issues/14

The core claim of the paper is related to PRL's prices. The paper says:

> at current PRL prices ($0.21)

But this is the price for Perle - https://coinmarketcap.com/currencies/perle/ - an entirely different project.

The actual Pearl (PRL) price is $0.76 as of now - https://www.coingecko.com/en/coins/pearl-2 - which is more than 3x difference. Based on this, I'm quite skeptical of the usefulness of the rest of the paper.

Plus the authors for at least one cited paper seem to be hallucinated/wrong (I'm leaning into hallucinated because the style looks more similar to Claude). For example, in the paper:

> D. Leinweber, C. Badertscher, and A. Kiayias. Challenges of proof-of-useful-work (PoUW). arXiv:2209.03865, 2022.

The reality if you open that arXiv:

> Challenges of Proof-of-Useful-Work (PoUW). Felix Hoffmann

The article itself is completely AI written, and writes so many words for essentially a very simple answer "no".

It's plausible for a person to prompt an LLM agent to behave that way, and then the rest would be done by the LLM. So the "seed" would still be human intent, but the subsequent actions would be by the LLM.

Active discussions from when it happened (February):

https://news.ycombinator.com/item?id=46990729

https://news.ycombinator.com/item?id=46987559

Also see https://news.ycombinator.com/item?id=48352939

Maybe I was a bit unclear in my post, sorry, I didn't mean that local LLMs were any less/more ethical, I meant that the people who prefer local LLMs over proprietary cloud ones sometimes cite ethics/etc as their reason.

A funny thing about this is that the current top-tier LLMs like GPT 5.5 in Codex and Opus 4.8 in Claude Code are extremely unlikely to act on those instructions. But smaller/cheaper models, especially small local ones, are more likely.

So, in a way, those instructions will realistically only harm whose who try to be more ethical with their LLM usage, rather than the ones who use the frontier ones from the "evil" AI companies.

I tried myself with GPT-5.5 in Codex, it simply ignored that instruction.

Yeah, I suppose that's one of the reasons why they changed it to a much more harmless instruction.

Laptops will also have to contain a much tighter configuration, display, keyboard, camera, etc ;)

DGX Spark is $4700, so I kind of doubt that RTX Spark's top configs will be cheaper than that.

For anyone curious to know how this will fare against Macbooks, at least in CPU perf: DGX Spark has the exact same GPU and CPU as the top RTX Spark laptops will, so you can just directly compare from that.

Of course, DGX Spark is a miniPC, so laptops will likely be slower due to power limits/throttling.

It won't, the top tier RTX Spark has the same exact CPU and GPU as DGX Spark, so you can check DGX Spark CPU benchmarks to see how it fares. Spoiler: it's about M3 Max level. And they're only coming this fall.

Reposting some of the (likely) bad commits with issues open against them, just to show that this isn't a one-off:

- https://github.com/RsyncProject/rsync/commit/4fa7156ccdb2ad3...: https://github.com/RsyncProject/rsync/issues/905 https://github.com/RsyncProject/rsync/issues/900

- https://github.com/RsyncProject/rsync/commit/1d5b5ab83af84db...: https://github.com/RsyncProject/rsync/issues/924

- https://github.com/RsyncProject/rsync/commit/859d44fa4f14207...: https://github.com/RsyncProject/rsync/issues/897

- https://github.com/RsyncProject/rsync/commit/30656c5e358b1c6...: https://github.com/RsyncProject/rsync/issues/896 https://github.com/RsyncProject/rsync/issues/915

- https://github.com/RsyncProject/rsync/commit/8112445318a35e4...: https://github.com/RsyncProject/rsync/issues/910 https://github.com/RsyncProject/rsync/issues/927

EDIT: posted on top, see https://news.ycombinator.com/item?id=48334270

Currently there are 130 Claude-coauthored commits, and the maintainer seems to not be engaging with any of the recent issues and just pushing more of the "security in depth" fixes that are breaking real features for people.

Amazing benchmarks!

> Dimensions: Height 490mm, Width 206mm, Depth 445mm

> Flow Rate: Min < 3 litres per hr, Max 55 litres per min.

> Water Pressure: Min 1 bar, Max 8 bar*

> Regeneration Time: 12 mins approx

> Water 17 litres