What happened to TheNumbers.com(stephenfollows.com) |
What happened to TheNumbers.com(stephenfollows.com) |
Despite the "download all data" link being prominently placed on the front page, the AI scrapers decided it would be more efficient to download terabytes upon terabytes of raw HTML by paginating through every possible facet on the search endpoint. Even with CloudFront caching results and a fairly efficient backend setup, the monthly bill ended up with around $1k just going toward network ingress/egress, so I shut down the site the following month.
Where they more exhaustive or more frequent?
Otherwise, I'm very curious to know more about their old and new architecture and what sorts of mitigation/scaling strategies they've started using to keep the site online.
I think what's really going on is that bots expose how underpowered web servers has gotten in recent years. In the 2000s, even poorly-architected PHP sites tended to serve about 200 requests per second, with 1000+ being common for static sites. I remember when Node.js came out and claimed that it could serve more like 100,000 RPS due to its cooperative threading model. But today sites have a remarkable slowness to them, running many hundreds or thousands of database queries due to ORMs and N+1 problems, so that response times can be 500 ms or more and even 1000 simultaneous users stresses servers.
What went wrong is that nobody solved stuff like Russian doll caching in a general way separate from the programming language and database. We should have had ways to make dependency graphs using Etag headers as keys with real cache invalidation of dependent data. I went down that rabbit hole 10 years ago using touch events in Laravel with callbacks to handle cache invalidation when class model data was saved to the database. Also a query cache using Redis which I think might have been handled better at the database level anyway. After that experience, I can honestly say that cache invalidation is so difficult to get right that it's effectively an open problem. Meaning that programmings should use a package instead of rolling it by hand, and it should be a major concern from the start (along with sharding by user id or using something like Firebase).
Don't get me started on how the web should have been a P2P content-addressable memory anyway. Nearly everything should be available from a nearby edge peer, similarly to BitTorrent. But nobody bothered to solve how to make that work with HTTPS/SSL. I suspect that has to do with early flaws in the browser security model where the whole page has to be behind HTTPS or warnings appear. So it was never clear what was personally identifiable information (PII) or merely public data being served over HTTPS. To really solve that, we probably need real trust networks and maybe even zero-knowledge proofs.
Since these problems are so challenging to fix, and big companies can't be bothered to do it since they pulled the ladder up behind them, we're probably stuck with banal "are you human" challenge screens for the foreseeable future.
It was... if you are paying datacenter rates for the bandwidth
If you're paying cloud provider per GB pricing, nah, even text will add up if you happen to be targeted by a bunch of bots
> What went wrong is that nobody solved stuff like Russian doll caching in a general way separate from the programming language and database. We should have had ways to make dependency graphs using Etag headers as keys with real cache invalidation of dependent data.
we did that with nested ESI includes in Varnish so every "box" of content on the page was cached separately + some piping for invalidation, so if a given piece in the database was changed it sent invalidation to all nodes. There was also some grace so if the thing you wanted got updated RIGHT NOW you might get stale version while the new one is updated in background, and don't pay the latency cost
It probably seems daunting but to be honest this feels like a weekend's work at this point with LLM assistance. Not to be glib!
Sure, you could technically redesign to handle the bot traffic, but if the bot traffic is just taking the data and reducing any need for humans to visit the site, why is he putting in the effort to maintain the site?
The article says (speculates?) that malicious users are trying to get privileged access for an edge in prediction market betting. From the article:
> If you could see The Numbers data before everyone else, every single week, you would have a significant edge over all the other traders - learning the answers slightly ahead of publication would allow you to front-run the trades.
Basically bots need to be (somehow) paying for the traffic they create, or prevented from creating it, or told to go away and then fined if they violate the request. No idea how to do these or even at what level in the stack they should happen, but they need to happen eventually somehow.
Disclosure: I am a shareholder and would love for them to solve the AI bot problem and the ad problem like this.
I know this isn't really the point of the article, but I've been thinking about this a lot. I wonder if we're going to see more resources go this way.
I used to publish little doo-dads as opensource software. Not because it was something that was legitimately ground breaking or anything (it absolutely wasn't) but because I had a problem, and I thought "heh wouldn't it be cool if someone else had a similar problem and could use my resource for it."
But now I'm really reluctant to give more stuff to the free web. Because the fact that it gets scraped and added to a pile of training data to later be monetized really rubs me the wrong way.
And I can see why maintainers of sites like this, or other free but incredibly useful resources might start to get irate at that.
Exactly, so use the AI to secure your servers. Ask the AI to audit your site for any security holes. If unable to rewrite, at least harden the existing code. AI’s are really really cheap (and fast) security consultants now.
1. bot traffic causing infrastructure cost
2. scraping circumventing paying for licenses
3. the risk of hacking
Webscraping is among the more benign things that gambling-on-anything can drive. And even that has a negative impact, as seen here.
I have no idea why those sites are legal.
From the article.
Not the same site, but an example of the same issue.
Deplorable behavior indeed
I'm not defending AI bots overwhelming websites or hackers motivated by Polymarket, etc., but I don't really believe a 30 year old website with "approximately 160,000 source files serving around 2 million pages" is a good litmus test for the state of the online world. What's worse is the absurdity that this basic site offering niche data would be targeted because Polymarket depends on it for some of their bets, something that most websites don't have to deal with. Frankly, that seems like a much more interesting angle to explore than "this old website that should have been re-written multiple times over the last three decades now doesn't have a choice but to be re-written."
Oddly, it seems like the solution, at least in this specific case, is relatively simple (and ironic): pay for an LLM to re-write this site in a modern, scalable, secure way and have that LLM monitor the site to ensure it is behaving correctly. Put this thing in a modern platform behind a proper cache (i.e., throw the whole thing up in Cloudflare) and many of these problems just don't exist anymore. If this site is as basic as it seems, this could be a weekend project.
It's certainly an interesting story, and I don't blame the owner of The Numbers for handling his site the way he has, but framing this as "AI is destroying our beloved internet" just seems obtuse, at least through this lens.
The intelligence community came to the conclusion (after research and experiments) that things like "prediction markets" were a bad idea in 1996.
Far worse now than then, with the net and AI of 2026 and the insane number of people now online that want to make an easy buck.
Don't get me wrong; I am guessing some of us on here would do well on those places. But they should not exist.
Or at least somewhere between that and full on protection racket.
LLMs come on the scene, hammer the site until it goes offline or racks up bills that threaten bankruptcy, and the solution then is to pay for an LLM to fix it and monitor it.
That's a real nice site ya got there, it would be a shame if massive datacenters going up around the world were to start hammering it from tens of thousands of IP addresses....
The thing is, people were scraping and monetizing other people's websites long before the current LLM fad. The difference is the magnitude of the problem.
> And I can see why maintainers of sites like this, or other free but incredibly useful resources might start to get irate at that.
Those who object to the scraping fall into several camps, but the biggest complaint I am hearing is that it increases both maintenance costs and time. In other words: it sucks when people are using your work in a manner that you find offensive, but it goes beyond that by doing actual harm.
People doing it with a couple of machines and residential proxies versus Anthropic doing it with 2 data centers worth of machines.
Scale matters.
I used to rank pretty well trashing crappy credit cards and encouraging people to switch to better options, then Google decided that 10 results for the card issuers website was better.
Same shit when I manually wrote proto-gethuman posts on calling telecoms/banks/etc (and also pushed visitors to try an Indy ISP or credit unions), then Google felt it was better to drive users to the telecom’s website that wants you to do anything but call them.
Please do train on my pre-LLM gold!
I feel the same way too. But guess what, all the code you did not publish gets into the training corpus anyway (when you gave Codex or whatever full read access to your filesystem).
AI solves exactly the meta-problem you describe: "I had a problem and needed to write a one-off doo-dad utility program to solve it." Now you can do something with your time besides writing pointless one-off doo-dads.
As for monetizing the training data, (a) it cost hundreds of millions of dollars to generate the weights, so why begrudge the companies that made the investment and did the research necessary to make it happen?; and (b) rest assured, whatever your doo-dad does, an open-weight model like GLM 5.2 can generate it for free using your own hardware.
So you don't have to pay anyone in that case. Well, except nVidia, I guess. Point granted there.
It doesn't have to make sense to you - I just believe that I'm not exactly alone in this thought.
Now apply this to Art, free stories, writing etc. It feels bad to have your free contributions hoovered up and monetized. It doesn't feel particularly fair or ethical to me. And it'd make me double think before making something free and publicly available.
Writing software one-offs to scratch an itch was historically one of my most enjoyable past-times. AI trivializing that has been a very real theft of joy in my life.
Solving the actual problem was never the point, it was just motivation to do geek-out and craft some code.
AI is rapidly diminishing many interesting hobbies (coding, art, music, writing).
Having more free time when there's nothing fun nor exciting to do with it isn't really a benefit.
...thus eliminating all the tedious chatting, relaxing and making friends that people were previously forced to do while waiting for elevators.
That also makes no sense, but I don't know what else I should have expected.
(I never publish my recommendations because they seem to complicated for people. Like using 30 GB plan/$10/monthly from T-Mobile for data and then using Tello for $10 plan for voice/text. This would require a 2 esim/sim card phone. I am currently using a Moto G Power 2024 phone from eBay $90/new, which was better than the $200 slightly used Pixel 6a from Swappa. Most people would just save the hassle, get a Galaxy phone with a phone contract.)
I've spent a lot of time in your shoes, wasting time on busy-work needed to accomplish a larger goal (and absolutely sharing the results freely, over multiple decades)... and I don't miss that part of it one bit.
LLMs are great, but they aren't producing new information. You still need people for that. But if you cut down any incentive for the people to do that, the LLMs will starve.