I would honestly love to hear why you think there is anybody at all working for or invested in Anthropic who is capable of orchestrating a coup against one of the most well capitalized companies on the planet AND a supplier partner that poses an existential risk to Anthropic itself (per their S-1) in the way of spending commitments and ability to procure compute capacity.
I don't need to hear why you think they did it. I'm already familiar with the "why" behind this crazy theory. I would like to know how you think it's possible they did this.
This is the only market where people think this is a reasonable thing to do, mostly because it's the only market in which most people are sellers not buyers.
Imagine if the government passed a law saying that when you got to a restaurant (or car wash, or bookstore, etc) you had to declare to the business what you're willing to pay before we see the menu so that the business could adjust their prices accordingly.
Because it's not about who's the seller or who's the buyer, it's about who has a whole institutional structure set up for doing the thing and who is doing it casually. When you walk into a pawn shop or one of those tacky-looking high street gold buyer places, you may sell them your valuables or buy someone else's valuables, but either way the shop is the shop and you are the customer. You don't become the shop when the shop is buying from you.
I was at Microsoft when they user-tested Windows 7 to death. Including installing eye tracking software during sessions so you could monitor where the user was looking at on the screen.
That data was in part used to justify the changes made in 8: "Users only click on the start menu 0.1% of the time and when they do they focus on a narrow portion of the screen."
A couple of the UX people who worked on Win8 told me that some of the UX study results were faked to give the desired results that justified the new Win8 design changes.
Def the sort of stories you hear at at a bar after work.
> therefore we need to make this menu as large as possible
shouldn't the conclusion have been "the start menu is so great the user doesn't even need to focus on using it". it feels like desperately trying to make the data fit the predefined conclusion, despite the data screaming and kicking.
That's really the big problem with the telemetry-driven designs today - the same raw data can be interpreted in whatever way you want so designers still end up doing what they feel like but can shut down any counter arguments (including complaints from real users) with "the data says otherwise".
Yes, I'm glad somebody's pointed that out. I read that post, "Users only click on the start menu 0.1% of the time...", and thought, well yes, what do you think the start menu's for? The user will use it to load something from time to time, then spend all their time in that application.
I've seen similar before - people not understanding that certain things are at their most successful if they're unobtrusive.
I recently experimented with some image generation models and I was surprised at how much context/prompting is needed to get something that doesn't look Lovecraftian. There was seemingly no downside to adding more and more detail up to the limit. Prose sits between code and images in terms of fuzziness and I wonder if promptmaxxing is the winning strategy for technical documentation.
I've been thinking that fine-tuning a model to write differently shouldn't be that hard (tedious, perhaps) and if I had the resources I would consider doing it myself.
If you haven't seen the trend on X of people hacking their TVs to remove the bloat and spyware then add in better UI it's been pretty awesome to watch.
I ended up finding some projects for samsung tvs. Not sure if they can help clean them up, but they do seem to allow installing your own apps and controlling the tv.
When I read the top-level comment, I was almost certain it was going to be AI-assisted. Before, your only chance at something like that was either having the skills to RE and mod everything yourself, or finding someone equally motivated and skilled to do it. Now you can pay AI to do your bidding, and while vibecoding is definitely not perfect, it has definitely lowered the bar. What an interesting era we live in.
It's not that 4 isn't a valid seed (though insanely unlikely) it's that once I know your values weren't picked randomly because I guessed one of them I can easily figure out the others.
In this case one cert gets defeated and you go the odds of that number being that low are 2^-384. Essentially impossible if it were randomly selected. It's much more likely you didn't choose that at random so I start guessing a bunch of other low numbers for your other certificates and find them quickly.
There are several very practical reasons (and not emotional) reasons to expect that "4" is the result of bad implementation.
I can think any number of programming mistakes that will result in a supposed (pseudo-)random number generator to be be biased to very small numbers. I have made such mistakes myself.
By any reasonable estimate, the probability of randomly picking a number below 256 is MUCH smaller than the probability of someone having made a mistake.
So you say that programming mistakes make small numbers more likely to be picked, and therefore seeing a small number is less likely than seeing a bigger number?
The claim is “if I generate a random number with equal distribution in a gigantic range and get a very tiny number, it’s more likely that there’s a mistake somewhere than that I just got this number randomly”.
In a working RNG, every number has equal probability.
But there's a probability <1 that the RNG is working as it should and a probability >0 that the RNG erroneously generates numbers that aren't equally distributed.
And because the range of possible numbers is so large and the probability of a mistake is not that tiny, a small result is more likely to be an error than a random result that happens to be small.
If I write a program to flip a coin and it generates heads 30 times in a row, is it more likely that this is just a random result or that I made a mistake in my programming?
One probability is 2^30, and the probability that someone makes an error resulting in this pattern is probably higher than 1/2^30.
If you cherry-pick that statement in isolation you're technically correct. But in practice the whole comment you're replying to is actually correct, in this case if you saw the number four, one would be correct in concluding that's unlikely and searching other low digit numbers, the proof of that is that it happened here.
I think they are right and you are wrong. If I understand properly you say:
"Playing roulette, in isolation after RED - RED - RED - RED - RED, there is not a higher chance of getting BLACK than RED in the next run. But in practice, one would be correct in concluding that it will probably be BLACK. The proof is that in our test, the next run gave BLACK".
Said differently, it was not particularly unlikely that the number would be 4. But maybe it got the attention of the person who decided to try assuming that it was a mistake, and they were indeed lucky.
My understanding is more that it's like seeing RED - RED - RED - RED - RED and wondering if the roulette is broken, testing it, and realising it is.
The hypothesis is "I see a small number, I can test if it is broken or not... oh it is". Could totally have been "no it's all good, just happened to be a small number".
I don't see why you would get downvoted for that. You're right.
It just happened that that particular number got the attention of a human, which was luck. But in terms of statistics, this number was not less likely than another one.
reply