That said, random or exhaustive search is a more scientifically useful method than you might think.
The first commercial antibiotics (Sulfa drugs) were found by systemically testing thousands of random chemicals on infected mice. This was a major drug discovery method up until the 1970s or so, when they had covered most of the search space of biologically-active small molecules.
A few month ago I went to a similar talk. They got a carboxylic acid from a plant (I forgot the name) that has some activity to kill caterpillar that eat corn, and made like 10 or 15 compounds with organic alcohols to get an ester. They tried different doses on the caterpillars and then make a computer model to predict the activity of similar compounds (QSAR). The idea is to use it in a long list of other organic alcohols and try to find a better compound.
But they choose chemical reactions that are usual in the lab, so they guess they will be able to make it work in the lab, and they keep most of the structure without changes. So it's closer to what they classify here as look nearby the known good points instead of a true random search.
Related, I was talking to a computational chemist at a conference a few years ago. Their work was mostly at the intersection of ML and material science.
An interesting concept they mentioned was this idea of "injected serendipity" when they were screening for novel materials with a certain target performance. They proceed as normal, but 10% or so of the screened materials are randomly sampled from the chemical space.
They claimed this had led them to several interesting candidates across several problems.
Jeff Dean literally featured it in a tweet announcing the model. Personally it feels absurd to believe they've put absolutely no thought into optimizing this type of SVG output given the disproportionate amount of attention devoted to a specific test for 1 yr+.
I wouldn't really even call it "cheating" since it has improved models' ability to generate artistic SVG imagery more broadly but the days of this being an effective way to evaluate a model's "interdisciplinary" visual reasoning abilities have long since passed, IMO.
It's become yet another example in the ever growing list of benchmaxxed targets whose original purpose was defeated by teaching to the test.
...only if you deliberately attempt to extract it by repeatedly prompting it to complete fragments of the book. They had to do quite a bit of work to make this happen.
so? It demonstrates that LLM models retain the copyrighted material in their weights. This is an important thing to consider about LLMs and shows that there need to be better protections for the creative industry.
Really? I retain plenty of copyrighted material in my head. What matters is the contexts in which I reproduce it (if any).
A search index might also contain copyrighted material. As long as it's used for search queries as opposed to regurgitation there's no problem. Search indexes and LLMs are both clearly very beneficial tools to have access to.
What does this (thought) experiment accomplish? That is, what point are you trying to make here?
Since we're talking about an electronic system the search index example is the more directly relevant one. Anyone who wants to object to LLMs is going to need to take care to ensure consistency with his views on Google's search index.
You could do the equivalent if they would let you. They don't. That's the point I was getting at. How the thing is used is what actually matters, not that it has "absorbed" copyrighted material.
I never claimed any change in copyright law. Only that one analogy was more direct than the other for the purpose of the current discussion.
You didn't answer my question. What point were you trying to make with your earlier reply?
Professional performers could certainly be viewed as such in this analogy. They memorize and then reproduce copyrighted material as a matter of course.
And when they do is when copyright protections might come into play. But not the basic learning of being a human being.
My playing copyrighted music on my synths at home, or singing lyrics along are different than if I am a professional musician benefiting financially from playing someone else's music in public.
Producing a product = market rules apply
Just living as a human = totally different thing
Yes, I agree. That was my entire point when I said: What matters is the contexts in which I reproduce it (if any).
The issue is not (or at least should not be) that LLMs are trained on material subject to copyright or can be very intentionally coaxed into regurgitating copyrighted material. The issue should be people building or using systems with the explicit intent of reproducing copyrighted material in an unauthorized manner.
> If an LLM is a product, and it contains the work (in this case can spit out Harry Potter) it is derivative. Doesn't matter what it's used for.
That's not the definition of a derivative work in copyright law; further, whether what legally qualifies as a derivative work is within the scope of the exclusive rights of the copyright holder is, in the US, subject to whether it is within one of the exceptions to exclusive rights in the law, notably the fair use exception, which very much does depend on, among other things, what it is used for.
That's dogma on your part. Rather than practical outcome you're opting for human exceptionalism. I can't accept that.
Merely containing a work doesn't make something derivative. A photograph could inadvertently capture a copyrighted image in the background but so long as it isn't the primary focus I think your line of reasoning there fails.
It is your view that's dogmatic. The law in this area has yet to be fully tested in court, let alone any prospective changes that might be made to it in the near future.
Regardless, I thought this was a discussion about what the law ought to say.
The defense is that the model is not designed to output Harry Potter verbatim, and in fact will not unless you jump through lots of hoops. Image generation would probably provide you with a stronger position here since those setups can easily output likenesses without needing to carefully engineer the prompt to cause them to do so. But even then it is clearly not the intention of the people training or deploying them that they be used that way.
> Hourglass. A subject would fall into this shape category when there is a very small difference in the comparison of the circumferences of her bust and hips AND if the ratios of her bust-to-waist and hips-to-waist are about equal and significant (Simmons, 2002)
> Rectangle. A rectangular subject would have her bust and hip measure fairly equal AND her bust-to-waist and hip-to-waist ratios low. She would not possess a clearly discernible waistline (Simmons, 2002)
Over here (E.U) I'd say most women definitely would be "hourglass shaped" in some way more than any other shape - maybe some would be a tie with "rectangle" but I'm breaking the tie by saying it's fair to say hourglass does not mean wasp-waist either - so I couldn't reconcile my anecdotal observation from the stated facts until it dawned on me that this was U.S stats.
> One 2007 study found that half of women (49%) in the U.S. were considered rectangle-shaped. Only 12% of women had a true hourglass figure.
> Results from the 2007–2008 National Health and Nutrition Examination Survey (NHANES), using measured heights and weights, indicate that an estimated 34.2% of U.S. adults aged 20 years and over are overweight, 33.8% are obese, and 5.7% are extremely obese.
And apparently it's worse for women (35.5% obese) than men (32% obese).
Anyway I'm not sure what "true hourglass" is supposed to even mean (wasp-waist?); according to the definition you got some waistline + balanced hip and shoulders => you're hourglass. If you start using "rectangle" as a fallback when in doubt then of course it's going to rate higher.
Funnily enough the very study linked is a comparison with another country (Korea):
This is just the price of being on the bleeding edge.
Unfortunately, prompt injection does strongly limit what you can safely use LLMs for. But people are willing to accept the limitations because they do a lot of really awesome things that can't be done any other way.
They will figure out a solution to prompt injection eventually, probably by training LLMs in a way that separates instructions and data.
I've said it before and I'll say it again, all AI discussion feels like a waste of effort.
“yes it will”, “no it won’t” - nobody really knows, it's just a bunch of extremely opinionated people rehashing the same tired arguments across 800 comments per thread.
There’s no point in talking about it anymore, just wait to see how it all turns out.
Now it's basically here, people really are getting attached to AI bf/gfs. It's only a matter of time until romance scams start using LLMs.
reply