Because an average request would have to travel half the globe, so setting up a simple HTTPS connection would already take a second. That's completely unacceptable for static content.

Not only would transfer be slow, they would also be much more pressing on the network as the request would occupy huge stretches and many interconnects and switches.

Furthermore, it hardens the website against DDoS and adds robustness for regional failures.

Minority of sites have global traffic. This is especially true for non-English sites.

So, while average request may have to travel quarter (not half) of the Internet globe, median request from the set of requests that matter has much lower latency.

Barcelona to Stockholm is about 65ms, ~35ms to Amsterdam, ~43ms to Frankfurt.

HTTP/3 is ubiquitous, you don’t get TCP handshake penalty anymore in major browsers.

Do they also have their own infra or are they just an entity between the customer and AWS?

They have their own infra but its not the size of CF. Chicken and Egg

I guess CSAM detection will remain hard as the training data raises so many ethical concerns.

I wonder if there's a way in which we can forward-hash sensitive material like CSAM and train a classifier that only consumes the 'hashed' version of the material for training and detection.

Very naive approach but wouldn't training a model on a few specific buckets like "someone is naked", "image is explicit" and "there is a child" in the picture, would do the job without having to train explicitely on CSAM? If the model returns both a high probability of "there is a child" and a high probability of any other bucket than this image is classified as CSAM. It'd be high recall,low precision but one would be on a rather safe side.

This is already done, there's hashed indexes of known CSAM and you can apply vector steering to the latent space to shift detections to trigger on CSAM material. You can also use two detectors for age estimation and a separate one for nudity and use them in combination.

I believe this is how many CSAM classifiers work. There isn't a simple legal avenue to host CSAM for classifier training, for good reason.

> Subdomains of existing domains are totally fine

Damn, another push to centralizing and walling off user content in 'fandoms', 'reddits' and other closed communities?

Have the internet cards been handed out now, and are we in the end-game?

Will old domains resell for premiums like low-background-radiation steel?

> Will old domains resell for premiums like low-background-radiation steel?

This has been a thing for a long time in domain name resale for speculation or link farming or similar, domains that are aged with "backlinks" sell for more money.

https://www.google.com/search?&q=domain+resale+backlinks

Private blog networks is another term in that side of the world. They're explicitly designed to look legit even to a human but give link juice to the spam world (which also look legit to a human). And the HN crowd say "I can spot an AI website a mile away!!!". Well cool, and good for you, but many people cannot. In fact, if you can, they are glad you are so smart and are leaving. Why? Because you wont convert.

The targets in the spam world (email and seo) that want to make money want the low hanging fruit. They want uneducated people to visit the site. I remember when google was still young, these same people wanted visitors from AOL. Why? lower traffic and much higher conversion. What kind of person uses AOL when google is available? An uneducated one and a new one to the internet. Or an older person who is likely to fall for a scheme.

I suspect more and more visitors coming from google will be seen this way in the age of AI and social media searching. Easy targets. Whether it's true or not is debatable (maybe people just want to read a real webpage and have AI burnout). Anyway thanks for coming to my ted talk.

I own a single English language dictionary word .net domain that is a common word related to lawsuits/legal activity and have been offered money to sell it. If I sell it, it will almost certainly become the front page of somebody's CRM lead generator for personal injury lawyers (or a blogspam site driving traffic to same through affiliate links).

People who work in actual infrastructure building tech may not know or have any reason to pay attentino to how many shady "affiliate" networks there are in existence for click traffic.

I dont know if I would discredit those folks as not knowing "actual infrastructure". The lengths I saw some folks go to get unique class C ip ranges and maintain automation is pretty impressive.

IP addresses too. Everyone who trades IP blocks pays a premium for ones that haven't been used for spamming in a long time, so they can be the first in a long time to use them for spamming.

It wouldn't surprise me. It seems like a mirror of the social media move from social graphs to recommendation engines. You are either producer or consumer and never shall consumer be promoted organically.

Search used to feel more like infrastructure: put something useful on the web, link to it, and eventually people could find it

The answer is simple. Limit your time on social media the way you'd be limiting a child and focus on IRL interactions.

What if I want to spend my time discussing things that are not represented in my geographical region?

Start a club in your geographical region.

This response sounds like "let them eat cake". You realize a lot of people on the internet (even here on HN) are in the developing world and, whether due to lack of means or weak passports, can't necessarily travel to where the things that interest them can be found locally?

I also find IRL social interactions also want pacing to some degree. But I don't know, I find my brain is sort of wired for these sorts of novel, pseudonymous interactions from a young age.

Yeah I get that. But treat them like candy - it's okay to eat a small dessert after lunch, it's okay to eat entire bucket of ice cream once in a while, but you know, your main diet needs to be quality stuff.

Words of wisdom to heed. Thank you, user anal_reactor :‑D

Stay open for more nuggets of wisdom!

I couldn't agree more with this comment!

The really perverse bit is that this rewards exactly the kind of consolidation Google's anti-spam changes are presumably meant to fight

> Have the internet cards been handed out now, and are we in the end-game?

The big companies are trying to corral and kill everybody else, yes.

That's not some "end-game", though. It's just powerful anti-social entities doing their thing while society refuses to punish them.

> Will old domains resell for premiums like low-background-radiation steel?

See also: any number of user-contributed sites that were bought by big players and basically parked. Does anyone even go to mp3.com for any real reason other than to see if anything has been done with it since the original was forced offline? I suspect they'll eventually unload it to someone who forks over enough money.

That or Google is getting so unusable even Discord is a better search engine.

I realize that this is a tangent but one of the more sad developments of recent developments as an advocate for open knowledge is how much of it is trapped on servers like discord that will never be indexed by any search engine.

There isn’t even an ability to turn on a feature to have discord chats indexed or anything. I realize it’s not only discord but it’s illustrative of what I perceive as an issue

It's no worse than how it used to be trapped in IRC greybeards and the solution, whatever it is, is presumably the same.

Not the end, just google bleeding itself out as a search engine while LLMs take over. Chatgpt often links me random, actually decent sites that I wouldn't have gotten from a simple google search on first page, because it would've been exclusively seo slop

ChapGPT (and all other hosted L To me) still use a search engine behind the scenes whenever you ask a question. Often, it uses those seo slop sites as primary sources when formulating an answer. Which is why, even though hallucinations are now somewhat rare these days, you _still_ should independently verify their output for anything important.

Maybe they do that opaque degradation trick that whenever it's asked something questionable, it'll route to a worse model instead.

It's going to be 50% less buggy, but we're going to write 10x as much code too.

It does change the output, they never said it did not. They said it would not _noticeably_ affect performance.

It doesn't necessarily change the output distribution; it depends exactly how it's implemented, and Anthropic haven't told us that. Google's original SynthID paper describes how you can do this.

Toy proof-of-concept: Anthropic owns a secret key which is a coin-flip Bernoulli random variable K with p=1/2. You are paying Anthropic to give you X, a Bernoulli random variable with p=1/2. Anthropic changes from their old strategy, "draw from K, then throw it away and flip a coin, each time you ask for a sample", to their new strategy, "draw from K and send it to you". You cannot observe the difference, but Anthropic knows K and so they know when you are repeating its outputs. (Obviously this is a toy example; in reality the distribution is vastly more complicated than Bernoulli, and Anthropic isn't just storing some model outputs to use as K but instead is computing a correlation with a known pseudorandomness source.)

But at low, or zero temperature, the distribution effectively collapses to top-token. This will not be compatible with watermarking.

Would explain why Antrophic removed T from the API.

You have a misunderstanding. Watermarking does not bias the responses in any way. How is this possible?

Before: "He leaped at the chance" - 33%. "Jumped at the opportunity" - 66%.

After: "He leaped at the chance" - 33%. "Jumped at the opportunity" - 66%.

But if you refresh your response from Anthropic 100 times:

Before: "Jumped at the opportunity" He leaped at the chance" "Jumped at the opportunity"

After: "He leaped at the chance" "He leaped at the chance" "He leaped at the chance"

The second one is detectable as being watermarked.

davmre has a good explanation that's more in-depth.

I understand, but this only applies at non-zero temperatures.

How is it a lie?

The performance of the output is poor and I can detect the watermark but only in a domain i am in the middle of, if it writes a summary of an email i can not tell, but in claude code in my codebase i definitely can tell. I am complaining about it!

Or when it tries a different language to explain a word: "The four items the ledger had标 standing open"!

You have no clue. Low-density suburbs create much more traffic than dense areas, just because you _have_ to travel >5km just to the next supermarket. In dense areas, most people walk.

Now that you can't white knight for the people with medical emergencies, you're white knighting for the poor?

Cars are a money sink, everyone would be richer if you don't need them for personal transport.

I never said anything about medical emergencies, and nobody is forcing you to own one today.

They artificially lock it down, because the app will give them more data to collect than the web app plus a larger surface to annoy the user with notifications, widgets, and other advertisement.