Most residential proxy providers that I've seen don't support UDP. UDP is what WebRTC uses and has been part of the browser's "core" since 2021. If someone says they're on Chrom(), FireFox, Safari - they have it.

So the setup goes like this:

1) Create a subdomain to get the client's TCP ip address. You can force it to ipv4 by not giving it an AAAA record. This will make future-comparison easier.

2) Create a subdomain to get the client's UDP ip address. Ditto the quad-a bit. You'll want to run your own STUN server (are you there, Claude? Its me, Justin...). To make it tamper resistant, I have the STUN server encrypt the ipv4 address and some other stuff and return it as an ipv6 ice candidate. Its abuse, but it works and I don't have to worry about back-end-server communication bottle necking anything. Just catch, encrypt, return.

And the algorithm looks like this:

** YOU HAVE A TCP AND UDP IP ADDRESS ***

1) If both IP Addresses are the same - probably not a residential proxy

2) If both IP Addresses are different but on the same ASN and in the same .../24(?) - they're on a cell network

3) If the IP Addresses are on different ASNs - that's a MAJOR red flag. Most people who use residential proxies have it drilled into their head - TURN OFF WebRTC. This leaked bad in the mid 2010's and people are still terrified of it.

4) If one of the Addresses is on Cisco Web Umbrella or zScaler - and other is residential - they're behind a corporate protection firewall. Not really a proxy per se, but it'll trip this algorithm. Based on your threat model, they're probably ok to whitelist.

*** YOU __ONLY__ HAVE TCP IP ***

So this is the "one weird trick" / secret sauce / whatever that makes this work. TLS-duration / TCP-rtt > 3.0

* "Many residential proxies split the connection at the exit: the origin’s TCP socket ends there, so Linux TCP_INFO estimates only origin <=> exit RTT. The exit forwards TLS handshake bytes to the true client, making ClientHello handshake-complete include the hidden proxy chain. An unusually high TLS-duration/TCP-RTT ratio exposes that split." *

This is really all we need. From here, we just do the following:

TCP == UDP === not proxy

TCP ~= UDP === if on same ASN, not proxy

TCP != UDP === proxy or zScaler-thing. Whitelist to suit.

No UDP and TLS/TCP RTT Ratio < 3 === probably not a proxy (lower is better here)

No UDP and TLS/TCP RTT > 3 === probably a proxy (most I've seen float between 6-12x)

I've worked with this thing for the last 6 months and have worked mainly with Soax, BrightData, and OxyLabs proxies. I also built my own ProxyDevice (android) and C2 - put it on my own network and it caught it.

This is extremely cool because this means you don't have to blacklist IP addresses (not like it works) since a legitimate user and a proxy will come from the same IP - the proxy client will look _very_ different. They get blocked by NetAcuity/MaxMind/StaticList - the entire IP is killed. This is surgical.

This probably isn't perfect, but based on my limited experience its spooky accurate. I won't claim its 100% accurate for all cases everywhere but if its better than nothing - by definition its better than nothing - and that helps. Which, at the end of the day is all I want to do with this - help.

p.s. more than happy to share what I've built. Test, evidence, code, whatever. Just let me know.

Comments URL: https://news.ycombinator.com/item?id=49668557

Points: 1

# Comments: 0

Because I don't see it. I'll take all the diseases and crt monitors for whatever advances AI is promising at this point.

Comments URL: https://news.ycombinator.com/item?id=49668216

Points: 3

# Comments: 3

The Gemini family had a distinct niche in document comprehension, with thousand page input documents taking only 300k tokens. Nothing quite like that in OpenAI or Anthropic world, even at more than 10x the token adjusted price. Should we just give up on Google at this point and engineer around the competitors' limits and eat the costs? Totally unnecessary own goal by team Google.

Comments URL: https://news.ycombinator.com/item?id=49668196

Points: 5

# Comments: 0

Posting this both because y'all might find it interesting and because I'm doing QA after having GLM-5.3-flash rebase memory safe glibc from 2.40 to 2.44.

Comments URL: https://news.ycombinator.com/item?id=49667945

Points: 1

# Comments: 0

Back to my last post, I miss tactile buttons. Feeling this way about majority of the technology these days.

Comments URL: https://news.ycombinator.com/item?id=49667449

Points: 4

# Comments: 8

https://mrsirg97-rgb.github.io/rig/ https://github.com/mrsirg97-rgb/rig

Comments URL: https://news.ycombinator.com/item?id=49667385

Points: 2

# Comments: 2

I just stumbled across a post[0] on Reddit, in which a user asked an AI model to promote his vibecoding project.

The agent submitted links to a bunch of sites, and the top site was HN with 1.5k views.

Color me surprised, as I've seen an explosion of LLM-content on other more popular platforms, but HN has usually been better than most.

How do we deal with this in the future?

Is flagging enough?[1]

[0]https://news.ycombinator.com/item?id=49665756

[1]I believe I found the post, which was already flagged when I discovered it, but I'll refrain from sharing here.

Comments URL: https://news.ycombinator.com/item?id=49667245

Points: 3

# Comments: 5

Lots of sign ups but people not completing onboarding + not responding to emails.

First time building SaaS.

Just wondering if you have tips how to at least get people complete the onboarding?

Comments URL: https://news.ycombinator.com/item?id=49666457

Points: 1

# Comments: 1

Points: 2

# Comments: 1

https://web.archive.org/web/20241003101005/https://www.angelfire.com/pa3/VAMPMAN/krypt_dcinfo.html

https://web.archive.org/web/20031205061600/http://dreamcast.planetweb.com/community/homepages/dcdesign.html

https://web.archive.org/web/20181029115530/https://dreamcastarena.com/how-to-create-dreamcast-website-friendly

https://www.dreamcast-talk.com/forum/viewtopic.php?t=16378

http://kamui.cool/zz_lab/infinitedream_HTTP_USER_AGENT.html

http://kamui.cool/zz_lab/jeffma_useragent.html

http://kamui.cool/zz_lab/browsermatrix.html

https://dreamcast.wiki/Development

https://dcemulation.org/phpBB/viewtopic.php?t=47902

https://web.archive.org/web/20250713015510/http://pc.psilocybindreams.com/

https://web.archive.org/web/20000521015744/http://tools.planetweb.com/wmhelp.html

Comments URL: https://news.ycombinator.com/item?id=49666410

Points: 4

# Comments: 0

The van depicted an image of the World Trade Center with a plane crashing into it. Eye witnesses describe the van in a video posted here: https://www.reddit.com/r/911archive/comments/14hippk/best_911_street_footage_ive_ever_seen_starts/

Go to 33:02 of the video to hear the eye witness account of the van.

The woman who called was also interviewed: https://www.youtube.com/watch?v=fOiCMMMeXE8

The van was sought by the FBI and the 5 men were eventually arrested, but they released and immediately returned to Israel, never to be heard of again.

So, what happened on that day? Why were they dancing in reaction to the attacks, and why did they have a van with an image of the attack on it?

What did they mean when they said they were there to "document the event"?

Comments URL: https://news.ycombinator.com/item?id=49665258

Points: 5

# Comments: 0

I once arrogantly thought software engineering would be one of the last roles to be impacted by AI, but now it's looking like it's one of the first.

We can debate about to what degree AI can actually take over from software engineers today, but I think there's no denying the trajectory. Companies don't want tenured and skilled developers anymore, they want dime a dozen agent supervisors to pump out features. Quality is taking even more of a backseat than before. Even if developers can't be replaced, AI significantly lowers the leverage we have.

Needless to say I didn't sign up for this. Supervising agents takes away pretty much everything I enjoyed about my job. I am somewhat mourning this loss, but I can at least be grateful I'm still relatively early in my career and young enough to learn a new skill.

So, now I'm thinking about what's next for me so I can protect both my income and my sanity. Since I left school at 16 I have no real qualifications, just my experience, which makes it harder to pivot.

If you feel the same way as me, what's your next move?

Comments URL: https://news.ycombinator.com/item?id=49664807

Points: 8

# Comments: 8

Points: 2

# Comments: 4

Points: 2

# Comments: 0

GitHub: https://github.com/Adarshent/Spnda Python SDK: pip install spnda

Why we built this: Semantic Entropy (Nature 2024) is great at detecting hallucinations, but the quadratic NLI cross-encoder bottleneck (90ms GPU overhead with DeBERTa) makes it unusable for real-time production serving.

We found that exact-match normalized entropy (R_sc) achieves the same discriminative AUROC on structured reasoning (GSM8K) while running in under 1 microsecond on CPU (760 nanoseconds).

Benchmarking against Python proxies: - Kernel latency: 760 ns (Rust) vs 15 ms (Python) - Memory RSS: ~3 MB vs 230 MB - Cold start: 3.6 ms vs 1.17 s

Happy to answer any questions about the Rust implementation, the zero-GPU exact-match math, or the benchmark results!

Comments URL: https://news.ycombinator.com/item?id=49663329

Points: 17

# Comments: 2

The news is full of AI or AI - adjacent articles to the extent it's now a tedious topic.

Where are the exciting innovations happening in the world? What great products are being built we're not hearing about?

Comments URL: https://news.ycombinator.com/item?id=49663205

Points: 3

# Comments: 2

For example, how do you go from Sonnet to Opus? Is Opus trained from scratch, built on Sonnet, or mostly the same model with more compute and training?

And how do models like Astra suddenly make a big jump in some capabilities? What is stopping Anthropic, Mistral, or others from doing the same thing? Is the main difference just more compute and money, or are there training methods, data, architecture, and research breakthroughs that competitors may not know about?

I can't think of a better place to ask this. I am guessing there are people here who actually work on these models and know what goes on behind the scenes.

Comments URL: https://news.ycombinator.com/item?id=49662978

Points: 6

# Comments: 2

Compare to https://web.archive.org/web/20260908235305/https://openai.com/index/navier-stokes-solution/

Also, see GitHub - git history: https://github.com/openai/NavierStokesAndEuler

Comments URL: https://news.ycombinator.com/item?id=49661928

Points: 5

# Comments: 1

For the novel projects I'm working on, Astra did unblock me where I was stuck with GPT 5.6 Sol, which in itself gives it value for myself, but outside of that it's not a significant improvement if at all. I have noticed Astra is more efficient so for my use case it may not be 2.5x cost.

Comments URL: https://news.ycombinator.com/item?id=49661691

Points: 6

# Comments: 3

Comments URL: https://news.ycombinator.com/item?id=49660811

Points: 2

# Comments: 3