I like the thought of this but how does it work in practice?
The company treats me well and preserves my job while I'm planning not to leave. Until they don't. Because once the transaction is more trouble than its worth - either financially, or politically, or interpersonally - I'm gone. But if I am planning to leave, the company doesn't know that and treats me the exact same way.
"It looks the same" is completely different than having been directly extracted from digital source.
If the digital source is the only thing that can be copyrighted, then you have to prove the digital source is what was used inappropriately. If you can't prove that, either because it didn't happen or because there's no technological way to prove it, then you can't prove copyright infringement.
Ages ago, I consulted on a case where a converted font was included in a product --- it was even simpler than https://luc.devroye.org/kinch.html since no transformations were applied --- just had to figure out which version of which font editor was used to open up the font file and then which settings were used to re-generate the font in the new format used for the infringing product.
If a digital font file is used as a source, it's probably too much work for the thief to erase all traces of that --- whether or no it should be acceptable to re-create a typeface design w/o crediting the original designer and adhering to their intent in terms of licensing and distribution is a different discussion.
You want to know if a new model is actually better, which you won't know if they just added the specific example to the training set. It's like handing a dev on your team some failing test cases, and they keep just adding special cases to make the tests pass.
How many examples does OpenAI train on now that are just variants of counting the Rs in strawberry?
I guess they have a bunch of different wine glasses in their image set now, since that was a meme, but they still completely fail to draw an open book with the cover side up.
> How many examples does OpenAI train on now that are just variants of counting the Rs in strawberry?
Well, that's easy: zero.
Because even a single training example would 'solved' it by memorizing the simple easy answer within weeks of 'strawberry' first going viral , which was like a year and a half ago at this point - and dozens of minor and major model upgrades since. And yet, the strawberry example kept working for most (all?) of that time.
So you can tell that if anything, OA probably put in extra work to filter all those variants out of the training data...
Nope! The knowledge cutoff does not show lack of leakage. Even if you get a non-confabulated cutoff which was before anyone ever asked the strawberry question or any question like it (tokenization 'gotchas' go back to at least davinci in June 2020), there is still leakage from the RLHF and tuning process which collectively constitute post-training, and which would teach the LLMs how to solve the strawberry problem. People are pretty sure about this: the LLMs are way too good at guessing things like who won Oscars or Presidential elections. This leakage is strongest for the most popular questions... which of course the strawberry question would be, as it keeps going viral and has become the deboooonkers' favorite LLM gotcha.
(This is, by the way, why you can't believe any LLM paper about 'forecasting' where they are just doing backtesting, and didn't actually hold out future events. Because there are way too many forms of leakage at this point. This logic may have worked for davinci-001 and davinci-002, or a model whose checkpoints you downloaded yourself, but not for any of the big APIs like GPT or Claude or Gemini...)
To gauge how well the models "think" and what amount of slop they generate.
Keeping it secret because I don't want my answers trained into a model.
Think of it this way, FizzBuzz used to be a good test to weed out bad actors. It's simple enough that any first year programmer can do it and do it quickly. But now everybody knows to prep for FizzBuzz so you can't be sure if your candidate knows basic programming or just memorized a solution without understanding what it does.
The only model I've seen so far that doesn't end up going crazy with long contexts with Gemini 2.5 pro, but tbf I haven't gone past 700-750k total tokens so maybe as it starts to approach the limit (1.05M) things get hairy?
Maybe "this font is offensively priced to the point where I immediately think the person selling it is a criminal using this for money laundering, or clinically insane" and "it's hard to sell enough fonts to live off of" are related?
Maybe if charlatans didn't say with a straight face that a font should be sold on a subscription model they'd sell more? Maybe if it didn't cost as much as a car they'd sell more?
Well the analogy falls apart because (among many, many other reasons) the people eating at Michelin rated restaurants, especially 3-star, are completely insensitive to the price. It will cost whatever it costs and there will still be a long wait to get a table, if you even can.
So rather than pretending we're talking about truffles, let's just talk about fonts directly without strained analogies. Fonts, which the majority of people don't even recognize. 90% of people don't even know what a foundry is. Your average person can't tell the difference between any two fonts if they're both sans-serif or serif.
It doesn't fall apart, you have examples that actually match it. Marketing boutiques of website creators match the 3-star Michelin analogy. High budgets from their customers (think LVMH) are the norm. And they will love and understand paying X for a font. In fact they will almost expect this type of thing in the design process.
At the end of the day if people don't see the difference and the value between a free and a priced one, then they don't need to steal and can just use the free ones. There are plenty of amazing free fonts anyway some being the actual roots of many paying ones, and the gold standards.
The mental overhead from having slightly different drop shadows and button sizes is minimal, and I think pretty overblown by people who prefer identical UIs for everything regardless of form factor (which is of course a valid take).
But someone could just as easily respond to today's UIs with "Yes, it's wonderful that every single app looks identical, as if it was all designed by one pretty boring artist with no creativity whatsoever" and that would also be a perfectly valid take.
Predictability is a feature. That does include button sizes. To me that's where the asymmetry is. The "boring artist with no creativity" complaint is aesthetic, whereas predictability is a functional concern.
They definitely did not say that and what is this constant need for people on the internet to respond to someone saying "maybe this isn't the right way to do something" with "Oh well then you're saying that something can't be done at all and it's pointless and why even try!!!11"
The company treats me well and preserves my job while I'm planning not to leave. Until they don't. Because once the transaction is more trouble than its worth - either financially, or politically, or interpersonally - I'm gone. But if I am planning to leave, the company doesn't know that and treats me the exact same way.
reply