Google warns its own employees: Do not use code generated by Bard

I have no experience working with a monorepo but I've read Google has "an army" of people maintaining the monorepo. I would have imagined with enough support like this, you wouldn't need to copy paste code everywhere in a monorepo?

I'd appreciate any insight anyone can share on monorepo. I am fascinated by the idea but I've never had a chance to work with it and people who I've met who have either don't want to or can't talk about their experience with it.

The other one I'm thinking of; I know this isn't the case with Google because their repo is too big, but, if a piece of copyrighted code would end up in an average repository, it would quickly be distributed to all users that have a copy of that repo, so they can't take it back again.

What that looks like, who knows.

Find someone at the same level of competency as you, set a goal, that person gets access to any LLM they want, you have to do it with no assistance other than reading stack or googling. And if you think your ego can handle it, find someone who you know is below you in competency (but not junior) and set the same goal and see if they surpass your own productivity or progress.

We’d have to set up some controls but you see where I’m getting at. It would be an interesting experiment either way.

It’s all in how you use the tools. If you’re zero shotting and hoping for the best then yeah, not great. But that’s where your domain expertise is supposed to come in.

It gets the rough outline of a solution OK but very easily gets lost if the problem requires 2 or 3 levels of what I guess I'd call nested analysis. Or rather, it's fine for creating a surface solution, but it gets lost in the weeds quite easily for things like type system or borrow checker problems.

Because it doesn't reason -- it (re)produces patterns.

At its core, an advertisement is made of two components: the subject - a store, a brand, a product, etc. - with which the victim is supposed to develop positive associations, ideally strong enough to motivate a purchase and/or advertising the subject to their acquaintances, and the message, which is meant to create those associations. Only the first part, the subject, has to be factual. The message does not, and in fact it usually isn't - manipulative bullshit performs much, much better, and generally the optimum for advertising seems to be asymptotically close to the line past which it would be legally fraud.

The only factual part, the subject, needs accurate handling, and is best suited for classical database systems - which is exactly how Google, and everyone else, is handling it. The message part - that's a good match to LLMs, which excel at producing convincingly sounding bullshit. For advertising, it seems what you need is to crank up the hallucinations a bit, but have some plausible deniability built into the whole system, so that when the LLM hallucinates in too obvious a way, no one can actually be held responsible.

It's always good to be aware of how things affect human hormones and absolutely worth study and regulation, but it's a huge jump from frogs are affected to humans are affected. I know right wing people are world class long jumpers so this kind of "it could be!" really isn't helping towards a reasonable debate.

Is that last bit from the Copyright Office, or is it the author's interpretation? Because I could just as easily imagine the battle being over who or what was the actual creator of the content (i.e. is it a derivative work), rather than whether the thing the machine created is eligible for copyright.

To remove the AI from the equation for a second, imagine that I took four images of living artists' work, placed them in a 2x2 grid, and called that a new artwork. There are two seperate questions to consider: (1) have I infringed upon the original authors' copyrights, and (2) is the new thing I have created eligible for copyright.

The stance that there is "no copyright protection for works created by non-humans" only addresses the second question, not the first question of how it interacts with existing copyrights.

Then where is the line? If I use a painting program that runs a Math.random() on a brush and I use this brush to draw something then is that also created by a non-human?

If I program a robot arm to draw a picture then is that picture also created by a non-human? What if I use a printer?

https://copyright.gov/docs/zarya-of-the-dawn.pdf is the letter that lays out why they don't think Midjourney is a paintbrush rather than a monkey.

I can't tell that this is a long-standing interpretation, rather than a recent ruling, conceivably a blip.

Just out curiosity, what makes the code you write, more original than an LLMs who’s training set is way bigger than yours and likely will more variance in how to achieve same goal.

I’m not trying to be facetious but rather stoke the philosophical idea of the reality that humans aren’t as special and unique as we think we are.

While I also think there is a good chance that what you are claiming (that any output is a derived work of the training set), this is definitely not settled, and will require someone going to court over it. I expect the proceedings over something like this to take a long time, and very likely reach to the supreme court. So, we won't know if you're right or not for quite a while.

A good argument against the idea that any output is a derived work of (all) data in the training set is exactly that verbatim copies of snippets from the Linux kernel can't be a derived work of snippets from GCC, even if GCC was also in the training set. So it is arguable that someone intending to claim copyright infringement for a piece of code generated by an LLM has to identify which particular work of theirs from the training set it infringes on, it can't be a blanket finding.

Leave other LLMs to the side on this issue. There are a variety of different issue that arise when Google employees might use a non-Google LLM, so that prohibition is in a category with different reasons.